<?xml version="1.0" encoding="UTF-8"?><feed xmlns="http://www.w3.org/2005/Atom">
  <title>GitHub Python Daily Trending Repositories</title>
  <id>https://mshibanami.github.io/GitHubTrendingRSS</id>
  <updated>2026-11-16T07:43:55Z</updated>
  <subtitle>Daily Trending Repositories of Python on GitHub</subtitle>
  <link href="https://mshibanami.github.io/GitHubTrendingRSS"></link>
  <entry>
    <title>bojieli/ai-infra-book</title>
    <updated>2026-11-16T07:43:55Z</updated>
    <id>tag:github.com,2026-11-16:/bojieli/ai-infra-book</id>
    <link href="https://github.com/bojieli/ai-infra-book" rel="alternate"></link>
    <summary type="html">&lt;p&gt;《深入理解 AI Infra：量化分析与系统设计》（李博杰 著）开源书稿：从硬件约束和模型架构出发，量化推导 LLM 推理与训练系统设计。含全书正文、PDF、配套计算工具与实验&lt;/p&gt;&lt;p&gt;&lt;img src=&#34;https://mshibanami.github.io/GitHubTrendingRSS/assets/icons/link.png&#34; width=&#34;20&#34; height=&#34;20&#34; alt=&#34;link&#34; style=&#34;margin: 0 8px 0 0; padding: 0; display: inline-block; vertical-align: middle;&#34; /&gt;&lt;a href=&#34;https://bojieli.github.io/ai-infra-book/&#34;&gt;https://bojieli.github.io/ai-infra-book/&lt;/a&gt;&lt;/p&gt;&lt;hr&gt;&lt;h1&gt;深入理解 AI Infra：量化分析与系统设计&lt;/h1&gt; &#xA;&lt;p&gt;&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/actions/workflows/book-site.yml&#34;&gt;&lt;img src=&#34;https://github.com/bojieli/ai-infra-book/actions/workflows/book-site.yml/badge.svg?sanitize=true&#34; alt=&#34;Build&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/LICENSE&#34;&gt;&lt;img src=&#34;https://img.shields.io/badge/license-Apache--2.0-blue.svg?sanitize=true&#34; alt=&#34;License&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://github.com/bojieli/ai-infra-book&#34;&gt;&lt;img src=&#34;https://img.shields.io/github/stars/bojieli/ai-infra-book?style=social&#34; alt=&#34;Stars&#34; /&gt;&lt;/a&gt;&lt;/p&gt; &#xA;&lt;p&gt;&lt;strong&gt;简体中文&lt;/strong&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-en/&#34;&gt;English&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-zh-tw/&#34;&gt;繁體中文&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-ru/&#34;&gt;Русский — community edition&lt;/a&gt;&lt;/p&gt; &#xA;&lt;div class=&#34;markdown-alert markdown-alert-tip&#34;&gt;&#xA; &lt;p class=&#34;markdown-alert-title&#34;&gt;&#xA;  &lt;svg class=&#34;octicon octicon-light-bulb mr-2&#34; viewbox=&#34;0 0 16 16&#34; version=&#34;1.1&#34; width=&#34;16&#34; height=&#34;16&#34; aria-hidden=&#34;true&#34;&gt;&#xA;   &lt;path d=&#34;M8 1.5c-2.363 0-4 1.69-4 3.75 0 .984.424 1.625.984 2.304l.214.253c.223.264.47.556.673.848.284.411.537.896.621 1.49a.75.75 0 0 1-1.484.211c-.04-.282-.163-.547-.37-.847a8.456 8.456 0 0 0-.542-.68c-.084-.1-.173-.205-.268-.32C3.201 7.75 2.5 6.766 2.5 5.25 2.5 2.31 4.863 0 8 0s5.5 2.31 5.5 5.25c0 1.516-.701 2.5-1.328 3.259-.095.115-.184.22-.268.319-.207.245-.383.453-.541.681-.208.3-.33.565-.37.847a.751.751 0 0 1-1.485-.212c.084-.593.337-1.078.621-1.489.203-.292.45-.584.673-.848.075-.088.147-.173.213-.253.561-.679.985-1.32.985-2.304 0-2.06-1.637-3.75-4-3.75ZM5.75 12h4.5a.75.75 0 0 1 0 1.5h-4.5a.75.75 0 0 1 0-1.5ZM6 15.25a.75.75 0 0 1 .75-.75h2.5a.75.75 0 0 1 0 1.5h-2.5a.75.75 0 0 1-.75-.75Z&#34;&gt;&lt;/path&gt;&#xA;  &lt;/svg&gt;Tip&lt;/p&gt;&#xA; &lt;p&gt;&lt;strong&gt;下载与阅读&lt;/strong&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/#%E7%89%88%E6%9C%AC%E4%B8%8E%E6%A0%BC%E5%BC%8F%E8%AF%B4%E6%98%8E&#34;&gt;格式与版本说明&lt;/a&gt;&lt;/p&gt; &#xA; &lt;ul&gt; &#xA;  &lt;li&gt;&lt;strong&gt;简体中文（原版，作者 &lt;a href=&#34;https://github.com/bojieli&#34;&gt;@bojieli&lt;/a&gt;）&lt;/strong&gt;：&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases/latest/download/AI-Infra-Book.pdf&#34;&gt;PDF&lt;/a&gt; · &lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases/latest/download/AI-Infra-Book.epub&#34;&gt;EPUB&lt;/a&gt; · &lt;a href=&#34;https://bojieli.github.io/ai-infra-book/&#34;&gt;在线阅读&lt;/a&gt;&lt;/li&gt; &#xA;  &lt;li&gt;&lt;strong&gt;English（社区翻译，&lt;a href=&#34;https://github.com/tg1482&#34;&gt;@tg1482&lt;/a&gt;）&lt;/strong&gt;: &lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases/latest/download/AI-Infra-Book-EN.pdf&#34;&gt;PDF&lt;/a&gt; · &lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases/latest/download/AI-Infra-Book-EN.epub&#34;&gt;EPUB&lt;/a&gt; · &lt;a href=&#34;https://bojieli.github.io/ai-infra-book/en/&#34;&gt;Read online&lt;/a&gt;&lt;/li&gt; &#xA;  &lt;li&gt;&lt;strong&gt;繁體中文（社区翻译，&lt;a href=&#34;https://github.com/edward821220&#34;&gt;@edward821220&lt;/a&gt;）&lt;/strong&gt;：&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases/latest/download/AI-Infra-Book-ZH-TW.pdf&#34;&gt;PDF&lt;/a&gt; · &lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases/latest/download/AI-Infra-Book-ZH-TW.epub&#34;&gt;EPUB&lt;/a&gt; · &lt;a href=&#34;https://bojieli.github.io/ai-infra-book/zh-tw/&#34;&gt;線上閱讀&lt;/a&gt;&lt;/li&gt; &#xA;  &lt;li&gt;&lt;strong&gt;Русский — community edition, &lt;a href=&#34;https://github.com/ilkruglov&#34;&gt;@ilkruglov&lt;/a&gt;&lt;/strong&gt;: &lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases/latest/download/AI-Infra-in-Depth-RU.pdf&#34;&gt;PDF&lt;/a&gt; · &lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases/latest/download/AI-Infra-in-Depth-RU.epub&#34;&gt;EPUB&lt;/a&gt; · &lt;a href=&#34;https://bojieli.github.io/ai-infra-book/ru/&#34;&gt;Онлайн-чтение&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-ru/&#34;&gt;О переводе&lt;/a&gt;&lt;/li&gt; &#xA; &lt;/ul&gt; &#xA;&lt;/div&gt; &#xA;&lt;p&gt;《深入理解 AI Infra》是 GitHub 上获得 &lt;strong&gt;50k+ Star&lt;/strong&gt; 的&lt;a href=&#34;https://github.com/bojieli/ai-agent-book&#34;&gt;《深入理解 AI Agent：设计原理与工程实践》&lt;/a&gt;的姊妹篇。&lt;/p&gt; &#xA;&lt;p&gt;写完&lt;a href=&#34;https://github.com/bojieli/ai-agent-book&#34;&gt;《深入理解 AI Agent》&lt;/a&gt;后，在与读者交流的过程中，我越来越感到：要开发好基于模型的应用，还需要理解它赖以运行的基础设施。大多数软件工程师不必亲自开发操作系统、编译器和芯片，却仍要学习操作系统、编译原理和计算机体系结构，因为申请内存、读取文件、调用函数，背后都有资源与时间代价。基于模型开发应用也是如此。延迟相差几倍，产品体验就可能完全不同；成本相差一个数量级，能够支撑的商业模式也随之改变。&lt;/p&gt; &#xA;&lt;p&gt;更深层的变化是&lt;strong&gt;编程抽象的上移：从操作系统到模型上下文&lt;/strong&gt;。传统的操作系统、编译器和硬件要为事先未知的各种程序提供通用能力，系统优化总要在可编程性与性能之间取舍。如今 LLM 成了最重要的应用，从算子执行到分布式调度，都可以针对特定的模型和加速器架构优化；模型设计也开始反过来适应硬件，DeepSeek V4/V4.1 重新设计长上下文的表示方式，就是一例。从某种意义上说，&lt;strong&gt;模型成了 LLM 时代的操作系统，AI Infra 成了 LLM 时代的计算机体系结构&lt;/strong&gt;。《计算机体系结构：量化研究方法》是我在体系结构领域的入门书，而 AI Infra 领域还缺少一本从硬件约束和模型架构出发、量化推导系统设计的书，这是我写作本书的动机。&lt;/p&gt; &#xA;&lt;p&gt;贯穿全书的方法是&lt;strong&gt;从约束推导设计&lt;/strong&gt;：先明确任务与质量要求，列出计算、存储、通信和依赖关系，对照硬件的容量、带宽和算力做数量级估算。这类估算人容易出错，AI 也一样：只算权重读取而忘了 KV 缓存，按峰值算力推算速度而不查带宽能否供给，把工作平分给多张卡却遗漏卡间通信，漏掉任何一项，结论都可能偏离几倍甚至几个数量级。估算还有另一层用意：读博时导师张霖涛博士反复叮嘱，优化一定要做到物理所允许的极限。本书贯彻这一习惯，先按第一性原理算出硬件允许的上限，再看实测离上限还有多远；差距不是模型漏了项，就是系统有可以去掉的开销。从 FPGA 加速 Bing 搜索排序、昇腾 AKG 算子生成到 UB 万卡互联，我反复遇到的是同一条线索：&lt;strong&gt;数据搬移&lt;/strong&gt;。本书因此反复追问五个问题：&lt;strong&gt;搬什么、搬多少、搬几次、经过哪里、谁必须等它。&lt;/strong&gt; 更多写作背景见&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/00-%E5%89%8D%E8%A8%80.md&#34;&gt;前言&lt;/a&gt;。&lt;/p&gt; &#xA;&lt;p&gt;目前书稿仍是初稿，正在持续修订。&lt;/p&gt; &#xA;&lt;h2&gt;内容目录&lt;/h2&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th style=&#34;text-align:center&#34;&gt;章&lt;/th&gt; &#xA;   &lt;th&gt;主题&lt;/th&gt; &#xA;   &lt;th&gt;主要问题&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;1&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/01-%E5%88%9D%E8%AF%86%20AI%20Infra.md&#34;&gt;初识 AI Infra&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;一次生成需要多少显存、计算和数据读写？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;2&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/02-%E6%A8%A1%E5%9E%8B%E6%9E%B6%E6%9E%84.md&#34;&gt;模型架构&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;注意力、历史状态与专家结构如何改变系统需求？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;3&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/03-%E6%8E%A8%E7%90%86%E4%B8%8E%E8%AE%AD%E7%BB%83%E8%B4%9F%E8%BD%BD.md&#34;&gt;推理与训练负载&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;任务阶段、到达模式和状态寿命如何影响资源需求？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;4&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/04-%E5%8A%A0%E9%80%9F%E5%99%A8%E6%9E%B6%E6%9E%84.md&#34;&gt;加速器架构&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;如何在计算、存储、带宽、功耗与成本之间取舍？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;5&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/05-%E7%AE%97%E5%AD%90%E4%B8%8E%E8%BF%90%E8%A1%8C%E6%97%B6.md&#34;&gt;算子与运行时&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;融合、复用、并发和调度如何减少执行开销？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;6&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/06-%E8%B6%85%E8%8A%82%E7%82%B9.md&#34;&gt;超节点&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;多设备协作如何平衡容量、吞吐和同步代价？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;7&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/07-%E6%95%B0%E6%8D%AE%E4%B8%AD%E5%BF%83%E7%BD%91%E7%BB%9C.md&#34;&gt;数据中心网络&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;网络带宽、通信方式和拥塞如何影响计算效率？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;8&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/08-%E6%8E%A8%E7%90%86%E4%BC%98%E5%8C%96.md&#34;&gt;推理优化&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;批处理、KV 管理、卸载与推测解码何时有效？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;9&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/09-%E5%88%86%E5%B8%83%E5%BC%8F%E6%8E%A8%E7%90%86.md&#34;&gt;分布式推理&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;如何放置计算和状态，并处理扩缩容与恢复？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;10&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/10-%E8%AE%AD%E7%BB%83%E7%B3%BB%E7%BB%9F.md&#34;&gt;训练系统&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;如何安排显存、通信和重算，让训练更高效？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;11&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/11-%E8%B5%84%E6%BA%90%E8%B0%83%E5%BA%A6%E4%B8%8E%E8%BF%90%E8%A1%8C%E7%8E%AF%E5%A2%83.md&#34;&gt;资源调度与运行环境&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;模型服务和工具环境如何共享资源，减少等待？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td style=&#34;text-align:center&#34;&gt;12&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/12-%E7%AB%AF%E8%BE%B9%E4%BA%91%E5%8D%8F%E5%90%8C.md&#34;&gt;端边云协同&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;任务放在本地、边缘还是云端，如何兼顾效果、延迟和成本？&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h2&gt;适合谁读&lt;/h2&gt; &#xA;&lt;p&gt;如果你已经调用过模型 API，或在自己的机器上运行过模型，想进一步弄清模型为什么慢、如何降低成本，这本书可以作为起点。从事系统、网络和芯片工作的工程师，以及相关方向的研究人员和学生，也能从书中看到各层设计与实际模型任务的联系。&lt;/p&gt; &#xA;&lt;p&gt;阅读时需要一些 Python、线性代数和计算机系统基础，具体要求见&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/00-%E5%89%8D%E8%A8%80.md&#34;&gt;前言&lt;/a&gt;。建议先读第 1—3 章，了解模型与负载，再根据自己的兴趣选择重点：&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;做模型应用和推理服务，可以重点读第 8、9 章，再看第 11、12 章的任务环境与部署；遇到容量、算子或通信问题时，回到第 4—7 章追踪原因。&lt;/li&gt; &#xA; &lt;li&gt;做系统或网络，可以重点读第 5—7 章，再看第 9、10 章，检查这些机制如何影响分布式推理与训练。&lt;/li&gt; &#xA; &lt;li&gt;做芯片与体系结构，可以重点读第 4—7 章，并结合后续章节的任务算例，检查硬件指标如何转化为实际的服务能力。&lt;/li&gt; &#xA; &lt;li&gt;希望系统学习，可以按目录顺序阅读，配合计算工具和实验，逐步核对自己的理解。&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;p&gt;遇到书中的算例，不妨先自己估一下，再看推导和实验结果。也可以换成你正在使用的模型和硬件，看看结论是否改变。&lt;/p&gt; &#xA;&lt;h2&gt;配套计算与实验&lt;/h2&gt; &#xA;&lt;p&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/calculations/README.md&#34;&gt;配套计算工具&lt;/a&gt;可以用来复算书中的数字，也可以换一组模型和输入，估算资源需求。工具附有模型配置和&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/calculations/results/README.md&#34;&gt;结果索引&lt;/a&gt;，静态计算只需 Python 3.10+ 标准库，无需 GPU 或模型权重。&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;git clone https://github.com/bojieli/ai-infra-book.git&#xA;cd ai-infra-book&#xA;&#xA;# 查看模型支持情况&#xA;python3 calculations/calc.py models&#xA;&#xA;# 估算 Qwen3-8B 在 8192-token prefill 下的逐算子资源需求&#xA;python3 calculations/calc.py forward --model qwen3-8b --tokens 8192 --format md&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;仓库使用 &lt;a href=&#34;https://git-lfs.com/&#34;&gt;Git LFS&lt;/a&gt; 保存论文和较大的输入与测量记录，合计约 20 GB。为避免克隆时全部下载，&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/.lfsconfig&#34;&gt;.lfsconfig&lt;/a&gt; 默认跳过所有 LFS 文件，工作区中只留下指针；正文、配图和静态计算都不需要它们。复现某个实验时，只下载对应目录：&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;git lfs install&#xA;git lfs pull --include=&#34;experiments/ch05/05-01/**&#34; --exclude=&#34;&#34;&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/experiments/README.md&#34;&gt;配套实验&lt;/a&gt;按章节存放在 &lt;code&gt;experiments/chXX/XX-YY/&lt;/code&gt; 中，每个实验都附有运行方法、输入条件和结果说明。需要 GPU 的实验会注明硬件与依赖要求；没有相应设备，也可以先阅读已有记录。复现或引用结果时，请留意所用的书稿版本、模型、硬件和输入参数。&lt;/p&gt; &#xA;&lt;h2&gt;本地构建&lt;/h2&gt; &#xA;&lt;p&gt;&lt;strong&gt;阅读网站&lt;/strong&gt;（Python 3.10+）：&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;python3 -m venv .venv-site&#xA;source .venv-site/bin/activate&#xA;python -m pip install -r website/requirements.txt&#xA;python scripts/build_site.py&#xA;python scripts/check_site.py&#xA;python scripts/build_site.py --serve&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;预览地址为 &lt;a href=&#34;http://127.0.0.1:8000&#34;&gt;http://127.0.0.1:8000&lt;/a&gt;，简体中文位于根目录，英文、繁體中文和俄语分别位于 &lt;code&gt;en/&lt;/code&gt;、&lt;code&gt;zh-tw/&lt;/code&gt; 和 &lt;code&gt;ru/&lt;/code&gt;。生成文件位于 &lt;code&gt;build/&lt;/code&gt;，详细说明见&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/website/README.md&#34;&gt;网站构建与发布&lt;/a&gt;。&lt;/p&gt; &#xA;&lt;p&gt;&lt;strong&gt;全书 PDF 与 EPUB&lt;/strong&gt;（另需 Pandoc、XeLaTeX 和字体）：&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;bash book/build_pdf.sh                        # 简体中文 → book/AI-Infra-Book.pdf&#xA;bash book-en/build_pdf.sh                     # English → book-en/AI-Infra-Book-EN.pdf&#xA;bash book-zh-tw/build_pdf.sh                  # 繁體中文 → book-zh-tw/AI-Infra-Book-ZH-TW.pdf&#xA;python3 book/build_epub.py --edition zh       # EPUB 只需 Pandoc 与 Poppler；--edition 可选 zh、en、zh-tw、ru&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;译本配图存于 Git LFS，构建前先下载：&lt;code&gt;git lfs pull --include=&#34;book-en/images/**,book-zh-tw/images/**&#34; --exclude=&#34;&#34;&lt;/code&gt;。依赖、字体及单章编译方法见 &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book/README.md&#34;&gt;PDF 编译说明&lt;/a&gt;。GitHub Actions 会检查 Pull Request 的网站与各语言的 PDF、EPUB 构建；推送到 &lt;code&gt;main&lt;/code&gt; 后自动发布 Release（四种语言各含 PDF 与 EPUB）并部署四种语言的在线阅读网站。&lt;/p&gt; &#xA;&lt;h2&gt;仓库结构&lt;/h2&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;目录&lt;/th&gt; &#xA;   &lt;th&gt;内容&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/README.md&#34;&gt;manuscripts/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;前言、十二章正文、配图与绘图脚本&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/experiments/README.md&#34;&gt;experiments/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;按章节组织的实验与运行记录&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/calculations/README.md&#34;&gt;calculations/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;资源计算工具、固定输入与复算结果&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/case-studies/README.md&#34;&gt;case-studies/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;模型、硬件和系统案例分析，按章节索引&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/references/README.md&#34;&gt;references/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;引用资料、来源清单与版本快照&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/research/README.md&#34;&gt;research/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;支撑正文的专题调研，以及各章修订记录&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book/README.md&#34;&gt;book/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;PDF 模板、构建与校验工具&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-en/README.md&#34;&gt;book-en/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;英文版（社区翻译）：前言与十二章正文、英文配图、翻译工具与 PDF 构建脚本&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-zh-tw/README.md&#34;&gt;book-zh-tw/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;繁體中文版（社区贡献）：前言与十二章正文、繁體配图、翻译工具与 PDF 构建脚本&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/website/README.md&#34;&gt;website/&lt;/a&gt;、&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/scripts/README.md&#34;&gt;scripts/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;网站资源、构建与检查脚本&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/archive/README.md&#34;&gt;archive/&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;历史大纲、审阅和写作协调记录&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h2&gt;参与贡献&lt;/h2&gt; &#xA;&lt;p&gt;书稿仍是初稿，欢迎读者参与改进。以下几个方向都有价值，欢迎认领：&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;指出并修正错误&lt;/strong&gt;：数字算错、单位或量级不对、引用与原文不符、公式或图表有误。发现一处就值得提一处，&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/issues/new?template=erratum.yml&#34;&gt;勘误 Issue&lt;/a&gt; 或直接提 PR 都可以。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;改写讲得不清楚的地方&lt;/strong&gt;：推导跳步、概念没有在首次出现时交代、例子不好懂。读不顺的地方多半是书写得不好，欢迎&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/issues/new?template=question.yml&#34;&gt;提出来&lt;/a&gt;，也欢迎直接给出更好的写法。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;补充遗漏的重要内容&lt;/strong&gt;：某个该讲的机制、架构或权衡没有写进来。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;修复配套代码的 bug&lt;/strong&gt;：&lt;code&gt;experiments/&lt;/code&gt; 的实验代码与 &lt;code&gt;calculations/&lt;/code&gt; 的计算 CLI，欢迎修正错误、补充测试或改进可用性。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;改进网页版&lt;/strong&gt;：&lt;a href=&#34;https://bojieli.github.io/ai-infra-book/&#34;&gt;在线阅读版&lt;/a&gt;的排版、导航、搜索和移动端体验都还有提升空间。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;翻译&lt;/strong&gt;：英文版与繁體中文版已由社区贡献，欢迎翻译为更多语言，可参考 &lt;code&gt;book-en/&lt;/code&gt; 和 &lt;code&gt;book-zh-tw/&lt;/code&gt; 的组织方式；也欢迎跟进英文版，使其与中文原版同步。翻译前请先开 Issue 说明计划，便于协调进度、避免重复劳动。&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;p&gt;正文的唯一来源是 &lt;code&gt;manuscripts/&lt;/code&gt; 下的 Markdown，网页版和 PDF 都由它构建生成，改正文请直接改这里。&lt;/p&gt; &#xA;&lt;h2&gt;作者与致谢&lt;/h2&gt; &#xA;&lt;p&gt;作者：&lt;a href=&#34;https://01.me/&#34;&gt;李博杰&lt;/a&gt;（&lt;a href=&#34;https://github.com/bojieli&#34;&gt;@bojieli&lt;/a&gt;）。&lt;/p&gt; &#xA;&lt;p&gt;感谢这些年一起做研究和工程的合作者，也感谢相关论文、开源项目与技术文档的作者，以及参与勘误和实验复现的读者。详细致谢见&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/00-%E5%89%8D%E8%A8%80.md&#34;&gt;前言&lt;/a&gt;，引用来源见正文脚注和&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/references/README.md&#34;&gt;参考资料库&lt;/a&gt;。本书 PDF 沿用《深入理解 AI Agent》的 ElegantBook / XeLaTeX 模板。&lt;/p&gt; &#xA;&lt;p&gt;英文版由 &lt;a href=&#34;https://github.com/tg1482&#34;&gt;@tg1482&lt;/a&gt; 翻译并贡献（&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/pull/4&#34;&gt;#4&lt;/a&gt;）。&lt;/p&gt; &#xA;&lt;p&gt;繁體中文版由 &lt;a href=&#34;https://github.com/edward821220&#34;&gt;@edward821220&lt;/a&gt; 翻译并贡献（&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/pull/10&#34;&gt;#10&lt;/a&gt;）。&lt;/p&gt; &#xA;&lt;p&gt;俄语 community edition 由 &lt;a href=&#34;https://github.com/ilkruglov&#34;&gt;@ilkruglov&lt;/a&gt; 组织并完成翻译与校验（&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/pull/20&#34;&gt;#20&lt;/a&gt;）。&lt;/p&gt; &#xA;&lt;h2&gt;贡献者&lt;/h2&gt; &#xA;&lt;p&gt;感谢所有为本书贡献内容、代码、翻译和修正的贡献者！点击头像即可访问他们的 GitHub 主页，也欢迎你加入。&lt;/p&gt; &#xA;&lt;!-- CONTRIBUTORS: GitHub contributors for bojieli/ai-infra-book, plus translators credited above. --&gt; &#xA;&lt;p&gt; &lt;a href=&#34;https://github.com/bojieli&#34;&gt;&lt;img src=&#34;https://avatars.githubusercontent.com/u/1421793?v=4&amp;amp;s=128&#34; width=&#34;64&#34; height=&#34;64&#34; alt=&#34;@bojieli&#34; title=&#34;@bojieli&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://github.com/tg1482&#34;&gt;&lt;img src=&#34;https://avatars.githubusercontent.com/u/31661606?v=4&amp;amp;s=128&#34; width=&#34;64&#34; height=&#34;64&#34; alt=&#34;@tg1482&#34; title=&#34;@tg1482&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://github.com/edward821220&#34;&gt;&lt;img src=&#34;https://github.com/edward821220.png?s=128&#34; width=&#34;64&#34; height=&#34;64&#34; alt=&#34;@edward821220&#34; title=&#34;@edward821220&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://github.com/determine123&#34;&gt;&lt;img src=&#34;https://avatars.githubusercontent.com/u/182319045?v=4&amp;amp;s=128&#34; width=&#34;64&#34; height=&#34;64&#34; alt=&#34;@determine123&#34; title=&#34;@determine123&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://github.com/sleeep-early&#34;&gt;&lt;img src=&#34;https://avatars.githubusercontent.com/u/186607517?v=4&amp;amp;s=128&#34; width=&#34;64&#34; height=&#34;64&#34; alt=&#34;@sleeep-early&#34; title=&#34;@sleeep-early&#34; /&gt;&lt;/a&gt; &lt;/p&gt; &#xA;&lt;!-- /CONTRIBUTORS --&gt; &#xA;&lt;p&gt;&lt;a href=&#34;https://github.com/bojieli/ai-infra-book/graphs/contributors&#34;&gt;查看 GitHub 上的全部贡献记录 →&lt;/a&gt;&lt;/p&gt; &#xA;&lt;h2&gt;许可&lt;/h2&gt; &#xA;&lt;p&gt;本书原创正文、配图及配套代码采用 &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/LICENSE&#34;&gt;Apache License 2.0&lt;/a&gt; 许可。Copyright © 2026 Bojie Li（李博杰）。&lt;/p&gt; &#xA;&lt;p&gt;仓库中的第三方代码、字体、模板与参考资料保留各自的版权和许可声明，不因收录于本仓库而改用 Apache-2.0；具体来源和使用条件见相应目录。&lt;/p&gt; &#xA;&lt;h2&gt;版本与格式说明&lt;/h2&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;PDF&lt;/strong&gt;：由 XeLaTeX 排版，公式、表格、脚注和交叉引用完整，推荐通读时使用。GitHub 直接显示 Markdown 时，LaTeX 公式和部分排版常常渲染不全或错位；上方目录链接到各章 Markdown 源文件，便于查找原文和提交勘误。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;EPUB&lt;/strong&gt;：字号可调，版面随屏幕自动重排，适合手机和电子书阅读器。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;在线阅读&lt;/strong&gt;：&lt;a href=&#34;https://bojieli.github.io/ai-infra-book/&#34;&gt;阅读网站&lt;/a&gt;提供章节导航、全文搜索和深色模式，右上角可切换语言。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;自动发布&lt;/strong&gt;：每次更新 &lt;code&gt;main&lt;/code&gt; 后自动构建全部格式并发布到 &lt;a href=&#34;https://github.com/bojieli/ai-infra-book/releases&#34;&gt;Releases&lt;/a&gt;，上方链接始终指向最新版。&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;翻译版本与译者&lt;/strong&gt;：简体中文为原版，作者为 &lt;a href=&#34;https://github.com/bojieli&#34;&gt;Bojie Li（李博杰）&lt;/a&gt;，正文位于 &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/manuscripts/README.md&#34;&gt;&lt;code&gt;manuscripts/&lt;/code&gt;&lt;/a&gt;。英文版由社区译者 &lt;a href=&#34;https://github.com/tg1482&#34;&gt;@tg1482&lt;/a&gt; 翻译，位于 &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-en/&#34;&gt;&lt;code&gt;book-en/&lt;/code&gt;&lt;/a&gt;，包含前言与十二章正文、重绘为英文标注的配图和独立的构建脚本。繁體中文版由社区贡献者 &lt;a href=&#34;https://github.com/edward821220&#34;&gt;@edward821220&lt;/a&gt; 翻译并整理，位于 &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-zh-tw/&#34;&gt;&lt;code&gt;book-zh-tw/&lt;/code&gt;&lt;/a&gt;，包含前言与十二章正文、繁體配图和独立的构建脚本；两项已知配图例外见&lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-zh-tw/README.md&#34;&gt;繁體版说明&lt;/a&gt;。俄语 community edition 由 &lt;a href=&#34;https://github.com/ilkruglov&#34;&gt;@ilkruglov&lt;/a&gt; 组织并完成翻译与校验，位于 &lt;a href=&#34;https://raw.githubusercontent.com/bojieli/ai-infra-book/main/book-ru/&#34;&gt;&lt;code&gt;book-ru/&lt;/code&gt;&lt;/a&gt;，依据固定版本的中文原稿制作，原作者仍为 &lt;a href=&#34;https://github.com/bojieli&#34;&gt;Bojie Li&lt;/a&gt;。各译本的译者、来源版本和许可证见对应目录说明；译本数字、公式和引用以原版为准。&lt;/li&gt; &#xA;&lt;/ul&gt;</summary>
  </entry>
  <entry>
    <title>kajisho5/ffmpeg-skill</title>
    <updated>2026-11-16T07:43:55Z</updated>
    <id>tag:github.com,2026-11-16:/kajisho5/ffmpeg-skill</id>
    <link href="https://github.com/kajisho5/ffmpeg-skill" rel="alternate"></link>
    <summary type="html">&lt;p&gt;Local video editor for coding agents: 42 FFmpeg tools with a machine-readable contract and an MCP server, for Claude Code, Cursor and Codex. Cut, captions, reframe, loudness, platform exports. No cloud, no API keys.&lt;/p&gt;&lt;p&gt;&lt;img src=&#34;https://mshibanami.github.io/GitHubTrendingRSS/assets/icons/link.png&#34; width=&#34;20&#34; height=&#34;20&#34; alt=&#34;link&#34; style=&#34;margin: 0 8px 0 0; padding: 0; display: inline-block; vertical-align: middle;&#34; /&gt;&lt;a href=&#34;https://www.npmjs.com/package/ffmpeg-skill&#34;&gt;https://www.npmjs.com/package/ffmpeg-skill&lt;/a&gt;&lt;/p&gt;&lt;hr&gt;&lt;p align=&#34;center&#34;&gt; &#xA; &lt;picture&gt; &#xA;  &lt;source media=&#34;(prefers-color-scheme: dark)&#34; srcset=&#34;brand/logo/ffmpeg-skill-horizontal-on-dark.svg&#34; /&gt; &#xA;  &lt;img src=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/brand/logo/ffmpeg-skill-horizontal.svg?sanitize=true&#34; height=&#34;72&#34; alt=&#34;ffmpeg-skill&#34; /&gt; &#xA; &lt;/picture&gt; &lt;/p&gt; &#xA;&lt;p align=&#34;center&#34;&gt;&lt;b&gt;Give your coding agent a video editor.&lt;/b&gt;&lt;br /&gt; Claude Code, Cursor, Codex or any MCP client edits your footage with local FFmpeg, then checks its own output.&lt;br /&gt; No cloud · No API keys · Python standard library&lt;/p&gt; &#xA;&lt;p align=&#34;center&#34;&gt; &lt;a href=&#34;https://kajisho5.github.io/ffmpeg-skill/&#34;&gt;&lt;b&gt;Website&lt;/b&gt;&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#quick-start&#34;&gt;&lt;b&gt;Quick start&lt;/b&gt;&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/demos.md&#34;&gt;Demos&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#tools&#34;&gt;Tools&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#mcp&#34;&gt;MCP&lt;/a&gt; · &lt;a href=&#34;https://www.npmjs.com/package/ffmpeg-skill&#34;&gt;npm&lt;/a&gt; &lt;/p&gt; &#xA;&lt;p align=&#34;center&#34;&gt; &lt;a href=&#34;https://github.com/kajisho5/ffmpeg-skill/actions/workflows/ci.yml&#34;&gt;&lt;img src=&#34;https://github.com/kajisho5/ffmpeg-skill/actions/workflows/ci.yml/badge.svg?sanitize=true&#34; alt=&#34;tests&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://scorecard.dev/viewer/?uri=github.com/kajisho5/ffmpeg-skill&#34;&gt;&lt;img src=&#34;https://api.scorecard.dev/projects/github.com/kajisho5/ffmpeg-skill/badge&#34; alt=&#34;OpenSSF Scorecard&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://www.npmjs.com/package/ffmpeg-skill&#34;&gt;&lt;img src=&#34;https://img.shields.io/npm/v/ffmpeg-skill&#34; alt=&#34;npm&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#ffmpeg-compatibility&#34;&gt;&lt;img src=&#34;https://img.shields.io/badge/ffmpeg-6.1%20%7C%207.1%20%7C%208%20%7C%209%20tested-orange&#34; alt=&#34;FFmpeg 6.1, 7.1, 8 and 9 tested in CI&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/LICENSE&#34;&gt;&lt;img src=&#34;https://img.shields.io/badge/license-MIT-green&#34; alt=&#34;MIT&#34; /&gt;&lt;/a&gt; &lt;a href=&#34;https://github.com/sponsors/kajisho5&#34;&gt;&lt;img src=&#34;https://img.shields.io/badge/sponsor-%E2%9D%A4-ea4aaa?logo=githubsponsors&#34; alt=&#34;Sponsor&#34; /&gt;&lt;/a&gt; &lt;/p&gt; &#xA;&lt;p align=&#34;center&#34;&gt; &lt;a href=&#34;https://kajisho5.github.io/ffmpeg-skill/&#34;&gt;&lt;img src=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/brand/og.png&#34; width=&#34;720&#34; alt=&#34;ffmpeg-skill v3: give your coding agent a video editor. An agent runs probe, cut, fit, export and check, and the final file is measured: 59.98 s, 1080x1920, 30 fps.&#34; /&gt;&lt;/a&gt; &lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;npx ffmpeg-skill&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;table&gt; &#xA; &lt;tbody&gt;&#xA;  &lt;tr&gt; &#xA;   &lt;td width=&#34;50%&#34;&gt;&lt;img src=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/demos/captions_pop_karaoke.gif&#34; alt=&#34;animated captions with a karaoke highlight&#34; /&gt;&lt;br /&gt;&lt;sub&gt;&lt;code&gt;caption.py --animate pop --karaoke&lt;/code&gt;&lt;/sub&gt;&lt;/td&gt; &#xA;   &lt;td width=&#34;50%&#34;&gt;&lt;img src=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/demos/reframe_crop.gif&#34; alt=&#34;16:9 reframed to 9:16&#34; /&gt;&lt;br /&gt;&lt;sub&gt;&lt;code&gt;fit.py --aspect 9:16 --fit crop&lt;/code&gt;&lt;/sub&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td width=&#34;50%&#34;&gt;&lt;img src=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/demos/silence_removal.gif&#34; alt=&#34;silence removed, timeline shorter&#34; /&gt;&lt;br /&gt;&lt;sub&gt;&lt;code&gt;silence.py --threshold -35&lt;/code&gt;&lt;/sub&gt;&lt;/td&gt; &#xA;   &lt;td width=&#34;50%&#34;&gt;&lt;img src=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/demos/loudness.gif&#34; alt=&#34;waveform before and after loudness normalisation&#34; /&gt;&lt;br /&gt;&lt;sub&gt;&lt;code&gt;loudness.py -I -14 --tp -1&lt;/code&gt;&lt;/sub&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt;&#xA;&lt;/table&gt; &#xA;&lt;p&gt;Left half is the input, right half is what the command produced. &lt;strong&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/demos.md&#34;&gt;Every before/after demo, with the exact command under each one →&lt;/a&gt;&lt;/strong&gt; — all of it built by &lt;code&gt;python3 demos/build.py&lt;/code&gt;, so you can rebuild every frame of it yourself. The footage is synthetic (FFmpeg&#39;s own test sources) except in three demos, &lt;code&gt;reframe_talk&lt;/code&gt;, &lt;code&gt;filler_en&lt;/code&gt; and &lt;code&gt;pauses_en&lt;/code&gt;, which use two AI-generated clips of people who do not exist (&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/demos/footage/README.md&#34;&gt;demos/footage/&lt;/a&gt;).&lt;/p&gt; &#xA;&lt;p&gt;&lt;code&gt;ffmpeg-skill&lt;/code&gt; is an &lt;a href=&#34;https://docs.anthropic.com/en/docs/agents-and-tools/agent-skills&#34;&gt;Agent Skill&lt;/a&gt; for Claude Code, Cursor, Codex and any agent that reads &lt;code&gt;SKILL.md&lt;/code&gt;. It teaches the agent a fixed workflow (probe → edit losslessly where possible → check → verify) and ships &lt;strong&gt;42 tools&lt;/strong&gt; that do the actual work with &lt;code&gt;ffmpeg&lt;/code&gt; / &lt;code&gt;ffprobe&lt;/code&gt;: cut, join, silence removal, fit to duration and aspect, captions and karaoke, overlays and motion graphics, HDR → SDR and LUTs, audio clean-up and typed dynamics, sync with drift correction, multicam, loudness, delivery checks, whole-edit project rendering, batch folders. Every tool is also callable as an MCP tool (&lt;code&gt;tools/list&lt;/code&gt; advertises a core 12 by default to keep client context small, with &lt;code&gt;FFMPEG_SKILL_MCP_FULL=1&lt;/code&gt; listing all 42; every tool is reachable by name through &lt;code&gt;tools/call&lt;/code&gt; either way), and the whole set is described by a machine-readable contract.&lt;/p&gt; &#xA;&lt;p&gt;If &lt;code&gt;ffmpeg&lt;/code&gt; and &lt;code&gt;python3&lt;/code&gt; are on your PATH, it works: offline, on footage you would rather not upload.&lt;/p&gt; &#xA;&lt;blockquote&gt; &#xA; &lt;p&gt;&lt;strong&gt;SPEC&lt;/strong&gt; (Self-Producing Execution Contract): each tool&#39;s &lt;code&gt;input_schema&lt;/code&gt; — the part of its contract and MCP tool definition that has to track the CLI flag-for-flag — is never hand-authored beside the code. It is derived, at run time, from the same &lt;code&gt;argparse&lt;/code&gt; parser that already defines the CLI, and CI fails the build if any of it drifts. → &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#what-is-spec&#34;&gt;full explanation&lt;/a&gt;&lt;/p&gt; &#xA;&lt;/blockquote&gt; &#xA;&lt;hr /&gt; &#xA;&lt;h2&gt;Standalone, and in an ecosystem&lt;/h2&gt; &#xA;&lt;p&gt;&lt;strong&gt;Standalone&lt;/strong&gt;, this is a local FFmpeg engine: probe → edit → verify, &lt;code&gt;npx ffmpeg-skill&lt;/code&gt; and nothing else. No API key, no account, no other repo required. Everything above and below this section describes that standalone tool, and none of it changes if you never read the rest of this one.&lt;/p&gt; &#xA;&lt;p&gt;&lt;strong&gt;In &lt;a href=&#34;https://github.com/kajisho5&#34;&gt;kajisho5&lt;/a&gt;&#39;s wider video-production ecosystem&lt;/strong&gt;, this repo is the &lt;em&gt;hands&lt;/em&gt;: it cuts, measures and exports files, and reports back in structured JSON. It does not decide what to cut, whether a deliverable is approvable, what makes a highlight interesting, or what a caption should say (the user&#39;s cue text is burned as written, never rewritten to fit) — those are a &lt;em&gt;brain&lt;/em&gt;&#39;s job, sitting in front of this engine, not inside it.&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;You want to...&lt;/th&gt; &#xA;   &lt;th&gt;Use&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Cut / join / measure / export a file right now&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;this repo&lt;/strong&gt; (&lt;code&gt;ffmpeg-skill&lt;/code&gt;), standalone&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Decide cut points, approve a deliverable, plan a whole edit&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://github.com/kajisho5/video-production-agent&#34;&gt;&lt;code&gt;video-production-agent&lt;/code&gt;&lt;/a&gt; / &lt;a href=&#34;https://github.com/kajisho5/AI-video-production-OS&#34;&gt;&lt;code&gt;AI-video-production-OS&lt;/code&gt;&lt;/a&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Build a typed editing graph across a workspace, without writing raw &lt;code&gt;ffmpeg&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://github.com/kajisho5/video-editing-skill&#34;&gt;&lt;code&gt;video-editing-skill&lt;/code&gt;&lt;/a&gt; / &lt;a href=&#34;https://github.com/kajisho5/audio-production-skill&#34;&gt;&lt;code&gt;audio-production-skill&lt;/code&gt;&lt;/a&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Put Remotion motion graphics (animated captions, silence jump cuts, audio-peak motion) on a real recording&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://github.com/kajisho5/remotion-ffmpeg-skill&#34;&gt;&lt;code&gt;remotion-ffmpeg-skill&lt;/code&gt;&lt;/a&gt;: its compositions take their length, fps, frame, silences, scene cuts, audio peaks and beats from this repo&#39;s &lt;code&gt;probe.py&lt;/code&gt;, &lt;code&gt;silence.py&lt;/code&gt; and &lt;code&gt;scenes.py&lt;/code&gt;. It verifies its render with &lt;code&gt;check.py&lt;/code&gt; and &lt;code&gt;look.py&lt;/code&gt;, composites a graphics layer with &lt;code&gt;overlay.py&lt;/code&gt; (&lt;code&gt;--place&lt;/code&gt; for its &lt;code&gt;--sparse&lt;/code&gt; mode), and delivers through &lt;code&gt;render.py --template&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;Other repos in the ecosystem — &lt;a href=&#34;https://github.com/kajisho5/media-analysis-skill&#34;&gt;&lt;code&gt;media-analysis-skill&lt;/code&gt;&lt;/a&gt;, &lt;a href=&#34;https://github.com/kajisho5/transcription-skill&#34;&gt;&lt;code&gt;transcription-skill&lt;/code&gt;&lt;/a&gt;, &lt;a href=&#34;https://github.com/kajisho5/subtitle-skill&#34;&gt;&lt;code&gt;subtitle-skill&lt;/code&gt;&lt;/a&gt;, &lt;a href=&#34;https://github.com/kajisho5/thumbnail-skill&#34;&gt;&lt;code&gt;thumbnail-skill&lt;/code&gt;&lt;/a&gt;, &lt;a href=&#34;https://github.com/kajisho5/color-grading-skill&#34;&gt;&lt;code&gt;color-grading-skill&lt;/code&gt;&lt;/a&gt;, &lt;a href=&#34;https://github.com/kajisho5/motion-graphics-skill&#34;&gt;&lt;code&gt;motion-graphics-skill&lt;/code&gt;&lt;/a&gt;, &lt;a href=&#34;https://github.com/kajisho5/qc-skill&#34;&gt;&lt;code&gt;qc-skill&lt;/code&gt;&lt;/a&gt; — read this repo&#39;s &lt;code&gt;contract --json&lt;/code&gt;, its tools&#39; &lt;code&gt;--json&lt;/code&gt; output and &lt;code&gt;doctor&lt;/code&gt;, the same way any agent framework would; this repo does not call into any of them. The dependency runs one way.&lt;/p&gt; &#xA;&lt;hr /&gt; &#xA;&lt;p&gt;&lt;strong&gt;Contents&lt;/strong&gt; &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#standalone-and-in-an-ecosystem&#34;&gt;Standalone, and in an ecosystem&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#why&#34;&gt;Why&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#quick-start&#34;&gt;Quick start&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#how-it-works&#34;&gt;How it works&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#design-principles&#34;&gt;Design principles&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#tools&#34;&gt;Tools&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#audio-is-a-first-class-input&#34;&gt;Audio&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#built-for-agents&#34;&gt;Built for agents&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#faq&#34;&gt;FAQ&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#ffmpeg-compatibility&#34;&gt;FFmpeg compatibility&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#tested-on-real-footage&#34;&gt;Tested on real footage&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#install&#34;&gt;Install&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#requirements&#34;&gt;Requirements&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#development&#34;&gt;Development&lt;/a&gt; · &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#docs&#34;&gt;Docs&lt;/a&gt;&lt;/p&gt; &#xA;&lt;hr /&gt; &#xA;&lt;h2&gt;Why&lt;/h2&gt; &#xA;&lt;p&gt;An agent that &#34;knows FFmpeg&#34; still guesses: it assumes a frame rate, picks a codec the container cannot hold, re-encodes a file that only needed a stream copy, and reports &#34;done&#34; without opening the result. This skill takes the guessing out:&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;Real files first.&lt;/strong&gt; Every job starts with &lt;code&gt;probe.py&lt;/code&gt;; the agent decides from the measured duration, fps, resolution, colour and audio layout, not from the file name.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Structured tools, not shell strings.&lt;/strong&gt; Each operation is a script with typed arguments. Nothing runs through a shell; no filter graph is accepted from the caller.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;A contract the agent can read.&lt;/strong&gt; &lt;code&gt;contract --json&lt;/code&gt; states, for every tool, what it takes, what it writes, which FFmpeg components it needs and how the result is verified. The MCP surface is derived from it.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Verification after execution.&lt;/strong&gt; The result is probed, checked against the destination&#39;s spec and, when the picture changed, looked at as a contact sheet.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Local first.&lt;/strong&gt; No cloud, no API keys, no Python dependencies. Optional local transcription is used when a speech engine (Parakeet or a whisper) is installed, never required — by &lt;code&gt;caption.py --transcribe&lt;/code&gt; and, since 1.17, by &lt;code&gt;silence.py --transcribe&lt;/code&gt; (with &lt;code&gt;--filler&lt;/code&gt;, or since 3.6 with &lt;code&gt;--sentence-min-silence&lt;/code&gt; / &lt;code&gt;--sentence-margin-tail&lt;/code&gt;); both take a transcript you already have instead, and both refuse with the install lines rather than guessing. Both take &lt;code&gt;--engine&lt;/code&gt; and &lt;code&gt;--model&lt;/code&gt;; &lt;code&gt;FFMPEG_SKILL_ASR_ENGINE&lt;/code&gt; and &lt;code&gt;FFMPEG_SKILL_ASR_MODEL&lt;/code&gt; set this machine&#39;s defaults (e.g. &lt;code&gt;large-v3-turbo&lt;/code&gt; where that whisper.cpp model is installed; without either, auto and &lt;code&gt;base&lt;/code&gt;). The engine runs on your machine and the audio is never uploaded; parakeet-mlx, faster-whisper and openai-whisper download their model weights the first time they run (parakeet-mlx&#39;s default from Hugging Face).&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;h2&gt;Paper&lt;/h2&gt; &#xA;&lt;p&gt;The design discipline behind this skill — per-operation verification for AI agents — is written up as a position paper. The paper scores ffmpeg-skill itself as only partially conformant with the discipline it defines.&lt;/p&gt; &#xA;&lt;p&gt;Kajiyama, S. (2026). &lt;em&gt;What Is Verification-Driven Coding? — A Discipline for Not Trusting Generation.&lt;/em&gt; Zenodo. &lt;a href=&#34;https://doi.org/10.5281/zenodo.23720216&#34;&gt;https://doi.org/10.5281/zenodo.23720216&lt;/a&gt;&lt;/p&gt; &#xA;&lt;h2&gt;Quick start&lt;/h2&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;# 0. FFmpeg itself, once (Ubuntu/Debian shown; macOS and Windows are in the Install table below)&#xA;sudo apt install ffmpeg&#xA;&#xA;# 1. install the skill for Claude Code (Cursor: --cursor, Codex: --codex, all three: --all)&#xA;npx ffmpeg-skill&#xA;&#xA;# 2. check the machine: ffmpeg, ffprobe and every FFmpeg component the tools need&#xA;npx ffmpeg-skill doctor&#xA;&#xA;# 3. (for agent frameworks) read the machine-readable contract&#xA;npx ffmpeg-skill contract --json | head -40&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;Captions from speech (&lt;code&gt;caption.py --transcribe&lt;/code&gt;) need one local speech engine, installed separately. &lt;code&gt;pip install faster-whisper&lt;/code&gt; is the simplest; the first run downloads its model, and the audio never leaves the machine. Without an engine, everything else works, &lt;code&gt;doctor&lt;/code&gt; lists the engine as missing optional, and &lt;code&gt;--transcribe&lt;/code&gt; refuses with the install line for each engine it supports.&lt;/p&gt; &#xA;&lt;p&gt;Already installed? re-run &lt;code&gt;npx ffmpeg-skill&lt;/code&gt; to refresh &lt;code&gt;~/.claude/skills/ffmpeg-skill&lt;/code&gt;. Copies are not updated automatically.&lt;/p&gt; &#xA;&lt;p&gt;&lt;code&gt;doctor&lt;/code&gt;&#39;s overall &lt;code&gt;ok&lt;/code&gt; and a single tool&#39;s &lt;code&gt;usable&lt;/code&gt; are different signals: &lt;code&gt;ok&lt;/code&gt; is false when anything &lt;em&gt;some&lt;/em&gt; tool requires is missing, so one missing filter (a plain Homebrew &lt;code&gt;ffmpeg&lt;/code&gt; without &lt;code&gt;subtitles&lt;/code&gt;, or a build without libvidstab) turns it false even though most tools still run. Check &lt;code&gt;doctor --json&lt;/code&gt;&#39;s &lt;code&gt;tools&lt;/code&gt; field for the per-tool answer, not just &lt;code&gt;ok&lt;/code&gt;.&lt;/p&gt; &#xA;&lt;p&gt;Then talk to your agent:&lt;/p&gt; &#xA;&lt;blockquote&gt; &#xA; &lt;p&gt;&#34;Take &lt;code&gt;interview.mp4&lt;/code&gt;, keep 0:45–3:10 and 5:00–6:30, and make it exactly 60 seconds for Reels.&#34;&lt;/p&gt; &#xA;&lt;/blockquote&gt; &#xA;&lt;p&gt;The agent runs &lt;code&gt;probe.py&lt;/code&gt;, &lt;code&gt;cut.py --segments 0:45-3:10,5:00-6:30&lt;/code&gt;, &lt;code&gt;fit.py --duration 60 --aspect 9:16 --fit crop&lt;/code&gt;, &lt;code&gt;export.py --preset reels&lt;/code&gt;, &lt;code&gt;check.py --platform reels&lt;/code&gt; and &lt;code&gt;look.py&lt;/code&gt;, then reports &#34;final.mp4: 59.98 s, 1080×1920, 30 fps, AAC stereo&#34; with the contact sheet it inspected.&lt;/p&gt; &#xA;&lt;h3&gt;Deliver to a platform&lt;/h3&gt; &#xA;&lt;p&gt;One command per destination, with the app&#39;s own UI taken into account:&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;python3 $S/render.py talk.mp4 --template tiktok --cues cues.txt&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;That fills the shipped &lt;code&gt;templates/tiktok.json&lt;/code&gt;: 9:16 crop, captions popped word by word &lt;em&gt;above&lt;/em&gt; TikTok&#39;s description bar and clear of its like column, −14 LUFS, the &lt;code&gt;tiktok&lt;/code&gt; export preset, and a &lt;code&gt;check.py --platform tiktok&lt;/code&gt; on the file it wrote. The caption size comes from the delivery table, so since 1.17.1 the filled project also states &lt;code&gt;&#34;fit_size&#34;: &#34;on&#34;&lt;/code&gt;: a size nobody asked for shrinks to fit the cue instead of splitting the sentence across two cues. A template that states its own &lt;code&gt;fit_size&lt;/code&gt;, and a &lt;code&gt;--brand&lt;/code&gt; that states a caption size, still win, and a project with &lt;code&gt;&#34;fit_size&#34;: &#34;off&#34;&lt;/code&gt; renders the captions 1.17.0 rendered. Templates ship for &lt;code&gt;tiktok&lt;/code&gt;, &lt;code&gt;reels&lt;/code&gt;, &lt;code&gt;shorts&lt;/code&gt;, &lt;code&gt;youtube-shorts&lt;/code&gt;, &lt;code&gt;youtube&lt;/code&gt;, &lt;code&gt;x&lt;/code&gt;, &lt;code&gt;linkedin&lt;/code&gt;, &lt;code&gt;facebook&lt;/code&gt;, &lt;code&gt;podcast&lt;/code&gt; and &lt;code&gt;audiogram&lt;/code&gt;; &lt;code&gt;--template all&lt;/code&gt; (or a comma-separated list) renders the seven video destinations (&lt;code&gt;tiktok&lt;/code&gt;, &lt;code&gt;reels&lt;/code&gt;, &lt;code&gt;shorts&lt;/code&gt;, &lt;code&gt;youtube&lt;/code&gt;, &lt;code&gt;x&lt;/code&gt;, &lt;code&gt;linkedin&lt;/code&gt;, &lt;code&gt;facebook&lt;/code&gt;) from the same edit and writes a &lt;code&gt;&amp;lt;name&amp;gt;_pack.md&lt;/code&gt; table of what each one produced. Files land next to the input unless &lt;code&gt;-o&lt;/code&gt; says otherwise, and &lt;code&gt;--dry-run&lt;/code&gt; shows every planned command rather than a result. &lt;code&gt;render.py --list-templates&lt;/code&gt; prints them with their frames, limits and safe zones. Alias spellings work everywhere a platform is named (&lt;code&gt;youtube-shorts&lt;/code&gt; = &lt;code&gt;shorts&lt;/code&gt;, &lt;code&gt;ig&lt;/code&gt; = &lt;code&gt;reels&lt;/code&gt;, &lt;code&gt;twitter&lt;/code&gt; = &lt;code&gt;x&lt;/code&gt;, &lt;code&gt;fb&lt;/code&gt; = &lt;code&gt;facebook&lt;/code&gt;).&lt;/p&gt; &#xA;&lt;p&gt;The tools also work on their own, from any shell:&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;S=~/.claude/skills/ffmpeg-skill/scripts&#xA;python3 $S/probe.py input.mp4 --compact&#xA;python3 $S/fit.py input.mp4 --duration 60 --aspect 9:16 --dry-run    # print the plan, run nothing&#xA;python3 $S/export.py input.mp4 --preset reels --json                 # structured result with a probe of the output&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;On Windows in Git Bash, &lt;code&gt;python3&lt;/code&gt; is only on PATH if Python was installed from the Microsoft Store; a &lt;a href=&#34;http://python.org&#34;&gt;python.org&lt;/a&gt; install exposes &lt;code&gt;python&lt;/code&gt; (or the &lt;code&gt;py&lt;/code&gt; launcher) instead — replace &lt;code&gt;python3&lt;/code&gt; with &lt;code&gt;python&lt;/code&gt; above if you see a &#34;command not found&#34;. &lt;code&gt;bin/install.js&lt;/code&gt; and &lt;code&gt;doctor&lt;/code&gt;/&lt;code&gt;contract&lt;/code&gt; already handle this for you; only the raw script examples above need it spelled out manually.&lt;/p&gt; &#xA;&lt;p&gt;More requests and the commands behind them: &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/examples/README.md&#34;&gt;examples/README.md&lt;/a&gt;. To see everything run end-to-end on generated footage: &lt;code&gt;npm run demo&lt;/code&gt; (the gallery in &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/demos.md&#34;&gt;docs/demos.md&lt;/a&gt;).&lt;/p&gt; &#xA;&lt;h2&gt;How it works&lt;/h2&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-mermaid&#34;&gt;flowchart TD&#xA;    U[User request] --&amp;gt; A[AI agent&amp;lt;br/&amp;gt;Claude Code · Cursor · Codex]&#xA;    A --&amp;gt;|reads| S[SKILL.md&amp;lt;br/&amp;gt;workflow, request → tool map, report format]&#xA;    A --&amp;gt;|runs| T[Structured tool&amp;lt;br/&amp;gt;scripts/&amp;amp;lt;name&amp;amp;gt;.py, typed argparse flags]&#xA;    T --&amp;gt; C[Contract&amp;lt;br/&amp;gt;input schema · role · capabilities · verification policy]&#xA;    C --&amp;gt; D[Capability detection&amp;lt;br/&amp;gt;doctor: available / missing / unknown]&#xA;    D --&amp;gt; F[FFmpeg execution&amp;lt;br/&amp;gt;no shell, stream copy when possible]&#xA;    F --&amp;gt; V[Verification&amp;lt;br/&amp;gt;probe · check · look.py contact sheet]&#xA;    V --&amp;gt; R[Structured result&amp;lt;br/&amp;gt;--json: status, output, commands, probe]&#xA;    R --&amp;gt; A&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;Over MCP the same tools are reached through a transport that holds no tool table of its own:&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-mermaid&#34;&gt;flowchart LR&#xA;    M[MCP client&amp;lt;br/&amp;gt;Claude Desktop · Cursor · any client] --&amp;gt; P[mcp/server.py&amp;lt;br/&amp;gt;stdio JSON-RPC]&#xA;    P --&amp;gt;|tools/list| C[Contract-derived ToolSpecs&amp;lt;br/&amp;gt;names · order · inputSchema]&#xA;    P --&amp;gt;|tools/call| T[scripts/&amp;amp;lt;name&amp;amp;gt;.py]&#xA;    C -.derived from.-&amp;gt; K[scripts/_contract.py]&#xA;    T -.described by.-&amp;gt; K&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;Names, order and &lt;code&gt;inputSchema&lt;/code&gt; in &lt;code&gt;tools/list&lt;/code&gt; are translated from each tool&#39;s argparse parser at start-up, so a new flag or a new script appears in MCP with no edit to &lt;code&gt;mcp/&lt;/code&gt;. A test copies the skill, adds, removes and edits a script, and reads &lt;code&gt;tools/list&lt;/code&gt; again to prove it.&lt;/p&gt; &#xA;&lt;h2&gt;Design principles&lt;/h2&gt; &#xA;&lt;p&gt;These are the rules the skill file gives the agent and the code enforces.&lt;/p&gt; &#xA;&lt;ol&gt; &#xA; &lt;li&gt;&lt;strong&gt;Probe first.&lt;/strong&gt; No tool decides from the file name. &lt;code&gt;probe.py&lt;/code&gt; measures duration, fps (with variable-frame-rate detection), resolution, rotation, bit depth, HDR format including Dolby Vision, colour tags and every audio stream before anything is cut.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Lossless when possible.&lt;/strong&gt; &lt;code&gt;cut.py&lt;/code&gt; and &lt;code&gt;loudness.py&lt;/code&gt; stream-copy what they do not need to touch (&lt;code&gt;join.py&lt;/code&gt; re-encodes so it can crossfade and conform mismatched clips; &lt;code&gt;--copy --transition none&lt;/code&gt; joins clips that already match by stream copy). Re-encoding happens only when it must: frame-accurate cuts, filters, format changes, or a keyframe farther than the tolerance.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Plan before render.&lt;/strong&gt; Every tool takes &lt;code&gt;--dry-run&lt;/code&gt; (print the ffmpeg command lines, write nothing; &lt;code&gt;verify&lt;/code&gt; ignores it), &lt;code&gt;--plan FILE&lt;/code&gt; (the dry run saved as a plan with fingerprinted inputs that &lt;code&gt;render.py FILE&lt;/code&gt; executes later, refusing if an input changed; &lt;code&gt;render&lt;/code&gt; and &lt;code&gt;verify&lt;/code&gt; refuse it), &lt;code&gt;--json&lt;/code&gt; (structured result with a probe of the output), &lt;code&gt;--fast&lt;/code&gt; (preview quality), &lt;code&gt;--progress&lt;/code&gt; (percent and ETA), &lt;code&gt;--timeout&lt;/code&gt; (a hung ffmpeg is killed and reported, never waited on forever; Ctrl-C or SIGTERM likewise stops the running ffmpeg, removes its partial output and reports &lt;code&gt;kind: interrupted&lt;/code&gt;) and &lt;code&gt;--overwrite&lt;/code&gt; (explicit consent before an existing output is replaced). A test runs every tool whose contract says its dry run runs no ffmpeg (&lt;code&gt;ffmpeg_execution: none&lt;/code&gt;, plus &lt;code&gt;cut&lt;/code&gt; and &lt;code&gt;metadata&lt;/code&gt; without their measuring flags) under &lt;code&gt;--dry-run&lt;/code&gt; behind a fake ffmpeg and asserts that no ffmpeg call happened and no file appeared; another runs every tool the contract says supports &lt;code&gt;--dry-run&lt;/code&gt; (all but &lt;code&gt;verify&lt;/code&gt;) behind an ffmpeg that logs each decode, and asserts that no file appeared and that every tool that decoded is one the contract lists as &lt;code&gt;analysis_only&lt;/code&gt;.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Machine-readable contract.&lt;/strong&gt; &lt;code&gt;contract --json&lt;/code&gt; describes all 42 tools: input schema generated from the parser, output schema, role, required and conditional FFmpeg capabilities, dry-run support, the verification tools to run afterwards, whether a visual check is required, &lt;code&gt;mutates_input: false&lt;/code&gt;. &lt;code&gt;provides&lt;/code&gt; lists all 42 by a cross-repository Capability id (&lt;code&gt;ffmpeg-skill.cut&lt;/code&gt;, &lt;code&gt;ffmpeg-skill.loudness&lt;/code&gt;, ...) for &lt;a href=&#34;https://github.com/kajisho5/AI-video-production-OS&#34;&gt;&lt;code&gt;kajisho5/AI-video-production-OS&lt;/code&gt;&lt;/a&gt;&#39;s &lt;code&gt;CapabilityContract.provides&lt;/code&gt; — see &lt;code&gt;docs/contract.md&lt;/code&gt;.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Contract-derived MCP.&lt;/strong&gt; &lt;code&gt;mcp/server.py&lt;/code&gt; builds its &lt;code&gt;tools/list&lt;/code&gt; from the contract. Tool names, order and &lt;code&gt;inputSchema&lt;/code&gt; cannot drift from the scripts; a test keeps the two byte-identical.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Capability detection.&lt;/strong&gt; &lt;code&gt;doctor&lt;/code&gt; reads &lt;code&gt;ffmpeg -encoders / -decoders / -filters / -bsfs&lt;/code&gt; and reports which of the components the tools need are present on this build (libx264, libass, zscale, loudnorm, xfade, …), before a job fails inside ffmpeg.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Unknown is not missing.&lt;/strong&gt; When a listing cannot be read (a layout the parser does not know, ffmpeg exiting non-zero) the affected capabilities are &lt;code&gt;unknown&lt;/code&gt;: never &lt;code&gt;missing&lt;/code&gt;, never silently &lt;code&gt;available&lt;/code&gt;. An installed filter is not reported absent; a failed detection is not a pass.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Verify the result.&lt;/strong&gt; The output is probed, and when the picture changed (captions, overlays, crops, colour, transitions) the agent runs &lt;code&gt;look.py&lt;/code&gt; and inspects the PNG. The report is not finished until its &lt;code&gt;Look:&lt;/code&gt; line names that image; audio-only jobs say &lt;code&gt;Look: not needed&lt;/code&gt;. &lt;strong&gt;&#34;Inspects&#34; means the calling agent&#39;s own vision, not a feature of this skill:&lt;/strong&gt; &lt;code&gt;look.py&lt;/code&gt; only renders a PNG; nothing in this repository detects faces, products, subjects, or &#34;the interesting part&#34; of a frame or a scene. When a crop or reframe needs to keep a specific part of the frame (&lt;code&gt;fit.py --fit crop --crop-x/-y&lt;/code&gt;, see &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#tools&#34;&gt;Tools&lt;/a&gt;), it is the multimodal agent looking at that PNG and choosing the anchor — a non-visual caller (a script, a CLI user without eyes on the sheet) has to supply that decision itself, and the default is a plain centre crop. Likewise &lt;code&gt;scenes.py --highlights&lt;/code&gt; ranks candidate scenes by a measured proxy (&lt;code&gt;--rank-by audio&lt;/code&gt; or &lt;code&gt;--rank-by duration&lt;/code&gt;), never by content; it is the agent that turns a look at the sheet into a judgement.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;One label per report.&lt;/strong&gt; A finished job is &lt;code&gt;Done:&lt;/code&gt;, a failure or a refusal is &lt;code&gt;Failed:&lt;/code&gt;, and a partial result is &lt;code&gt;Done:&lt;/code&gt; with the shortfall named in &lt;code&gt;Notes:&lt;/code&gt; — never a third label such as &lt;code&gt;Done (partially):&lt;/code&gt;.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Keep originals.&lt;/strong&gt; No tool overwrites its input. Outputs are new files named &lt;code&gt;&amp;lt;input&amp;gt;_&amp;lt;operation&amp;gt;.&amp;lt;ext&amp;gt;&lt;/code&gt; unless told otherwise, and a test hashes every input after the run.&lt;/li&gt; &#xA;&lt;/ol&gt; &#xA;&lt;h2&gt;Tools&lt;/h2&gt; &#xA;&lt;p&gt;42 public tools, all Python 3.10+ standard library, all with &lt;code&gt;--help&lt;/code&gt;, &lt;code&gt;--dry-run&lt;/code&gt;, &lt;code&gt;--json&lt;/code&gt;, &lt;code&gt;--plan FILE&lt;/code&gt; (a dry run written as a plan &lt;code&gt;render.py&lt;/code&gt; executes later), non-zero exit and a reason on stderr on failure. &lt;code&gt;--json-brief&lt;/code&gt; (1.11.0) prints the same result document trimmed to what a caller acts on — status, output, &lt;code&gt;verified&lt;/code&gt;, a compact &lt;code&gt;summary&lt;/code&gt; of the output probe, the tool&#39;s own keys and the command count instead of the command lines — for roughly a third of the bytes; &lt;code&gt;--json&lt;/code&gt; itself is unchanged. Every re-encoding tool except &lt;code&gt;export.py&lt;/code&gt;, &lt;code&gt;render.py&lt;/code&gt; and &lt;code&gt;batch.py&lt;/code&gt; (a preset or the project decides there) takes &lt;code&gt;--codec h264|hevc|av1|prores&lt;/code&gt; and &lt;code&gt;--quality N&lt;/code&gt; (1.8) and, with &lt;code&gt;export.py&lt;/code&gt;, &lt;code&gt;--hw&lt;/code&gt;/&lt;code&gt;--no-hw&lt;/code&gt; (Apple VideoToolbox on Apple Silicon, opt-in; see &lt;code&gt;doctor&lt;/code&gt; below), and every time flag takes seconds, &lt;code&gt;mm:ss&lt;/code&gt;, &lt;code&gt;hh:mm:ss.fff&lt;/code&gt; or SMPTE &lt;code&gt;hh:mm:ss:ff&lt;/code&gt; with an optional &lt;code&gt;@fps&lt;/code&gt; suffix (1.9).&lt;/p&gt; &#xA;&lt;p&gt;&lt;strong&gt;Analysis and inspection&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Tool&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;probe.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Duration, fps (+ VFR detection), resolution, codecs, bit depth, HDR format incl. Dolby Vision (&lt;code&gt;hdr&lt;/code&gt; for a real PQ/HLG/DV transfer only, &lt;code&gt;bt2020_or_hdr&lt;/code&gt; also for BT.2020 SDR), colour space, rotation, every audio stream; &lt;code&gt;--analyze&lt;/code&gt; flags Log footage&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;scenes.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Scene changes, audio peaks, highlight proposals (&lt;code&gt;--rank-by audio&lt;/code&gt; loudest, or &lt;code&gt;--rank-by duration&lt;/code&gt; longest — both proxies, not &#34;best&#34;) and a per-scene sheet; cut list for &lt;code&gt;cut.py --segments&lt;/code&gt;; &lt;code&gt;--beats&lt;/code&gt; measures the music&#39;s beat grid (tempo, beat times, confidence); &lt;code&gt;--shots&lt;/code&gt; classifies each shot static/pan/motion by measured optical flow; &lt;code&gt;--audio-peaks&lt;/code&gt; lists per-second dBFS; &lt;code&gt;--speech&lt;/code&gt; reports a speech-vs-music ratio, not a classification&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;look.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Contact sheet, single frames, side-by-side comparison as PNG so the agent can see what it made; &lt;code&gt;--safe NAME&lt;/code&gt; shades the zones a platform&#39;s own UI covers; &lt;code&gt;--guides title,action,4:3&lt;/code&gt; outlines the broadcast safe areas (on the PNG only)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;Editing&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Tool&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;cut.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;In/out or multi-segment cuts, lossless &lt;code&gt;-c copy&lt;/code&gt; first, re-encode fallback, &lt;code&gt;--accurate&lt;/code&gt; for frame-exact video and sample-exact audio, &lt;code&gt;--smart&lt;/code&gt; for a frame-exact H.264 cut that re-encodes only the partial GOPs at the cut points (matched to the source, verified against its packets and frames) and copies the rest, an SDR HEVC source re-encodes as HEVC (&lt;code&gt;--no-keep-hevc&lt;/code&gt; for x264), a single &lt;code&gt;.mp4&lt;/code&gt;/&lt;code&gt;.mov&lt;/code&gt; copy keeps its edit list, so the picture starts at &lt;code&gt;--start&lt;/code&gt; unless a keyframe lands later (&lt;code&gt;start_snapped&lt;/code&gt; then says so; &lt;code&gt;--no-edit-list&lt;/code&gt; for 2.x&#39;s shifted start), &lt;code&gt;--vfr-guard&lt;/code&gt; to choose the VFR check (&lt;code&gt;sampled&lt;/code&gt; by default); reports &lt;code&gt;precision&lt;/code&gt;; &lt;code&gt;--snap beats&lt;/code&gt; moves the in/out points onto a measured beat, or refuses when there is no measurable pulse ; &lt;code&gt;--remove A-B,...&lt;/code&gt; drops ranges and keeps the rest&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;join.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Concatenate clips with xfade transitions, normalising size, fps, sample rate and channel layout (the widest clip&#39;s, or &lt;code&gt;--channels&lt;/code&gt;); audio-only inputs are joined as audio; &lt;code&gt;--list parts.txt&lt;/code&gt; reads generated segments (TTS lines) and checks every one first — missing, empty or unreadable parts are all named in one refusal, or skipped and reported with &lt;code&gt;--on-missing skip&lt;/code&gt; ; &lt;code&gt;--copy --transition none&lt;/code&gt; joins clips with matching streams by stream copy, measured afterwards, and refuses a mismatch by name&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;silence.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Detect and remove dead air (jump cuts) with a margin around speech (&lt;code&gt;--margin&lt;/code&gt;, 0.2 s each side); list or export the cut list, or a source-to-output time map (&lt;code&gt;--timemap&lt;/code&gt;, JSON or CSV); reports the level at every cut and warns where one may land in a breath or a word&#39;s tail; &lt;code&gt;--filler&lt;/code&gt; also removes filler words, but only where a speech engine timed them; &lt;code&gt;--speech-aware&lt;/code&gt; keeps a breath inside a sentence and cuts only at sentence-boundary pauses, composing with &lt;code&gt;--filler&lt;/code&gt; through the same &lt;code&gt;keep_ranges()&lt;/code&gt;; &lt;code&gt;--margin-lead&lt;/code&gt; / &lt;code&gt;--margin-tail&lt;/code&gt; keep different amounts before and after speech, and with word timings &lt;code&gt;--sentence-min-silence&lt;/code&gt; / &lt;code&gt;--sentence-margin-tail&lt;/code&gt; give the pause after a sentence end (the transcript&#39;s punctuation) its own floor and tail&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;fit.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Fit to a duration (pitch-preserving speed change or trim, smooth slow-mo) and/or aspect ratio (pad, crop or &lt;code&gt;--fit blur&lt;/code&gt;&#39;s blurred, darkened fill, with &lt;code&gt;--crop-x&lt;/code&gt;/&lt;code&gt;--crop-y&lt;/code&gt; to keep an off-centre subject) and/or exact &lt;code&gt;--width&lt;/code&gt;/&lt;code&gt;--height&lt;/code&gt;, with the result&#39;s &lt;code&gt;picture&lt;/code&gt; saying where the picture sits in the frame written (measured); rotate 90/180/270, flip h/v; force constant fps&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;crop.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Crop to an exact pixel rectangle (&lt;code&gt;--x --y --width --height&lt;/code&gt;) — distinct from &lt;code&gt;fit.py --fit crop&lt;/code&gt;, which crops to an aspect ratio it computes itself&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;cropdetect.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Measure existing black letterbox/pillarbox bars and report the &lt;code&gt;crop.py&lt;/code&gt;-ready rectangle that removes them — analysis only, writes no file; &lt;code&gt;--motion-centre&lt;/code&gt; reports the per-second motion centroid instead, for a 9:16 reframe — report only, the calling agent picks the crop, no auto-reframing&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;deinterlace.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Deinterlace interlaced source footage (&lt;code&gt;yadif&lt;/code&gt;), &lt;code&gt;--mode frame&lt;/code&gt;/&lt;code&gt;field&lt;/code&gt;, &lt;code&gt;--parity&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;denoise.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Reduce video noise/grain (&lt;code&gt;hqdn3d&lt;/code&gt;), &lt;code&gt;--strength low/medium/high&lt;/code&gt; or individual spatial/temporal overrides&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;redact.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Blur, pixelate, or (&lt;code&gt;--mode delogo&lt;/code&gt;) interpolate away an exact pixel rectangle for the whole clip. The box is given: nothing here finds a face, a plate or a logo, and &lt;code&gt;delogo&lt;/code&gt; is a border blend, not inpainting&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;sphere.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Extract a flat rectilinear viewport from a 360/spherical video (&lt;code&gt;--yaw --pitch --roll --h-fov --v-fov&lt;/code&gt;); no subject tracking, only the aim you give it&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;straighten.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Rotate by an arbitrary angle for horizon correction (&lt;code&gt;--degrees&lt;/code&gt;, &lt;code&gt;--fit crop/pad&lt;/code&gt;) — distinct from &lt;code&gt;fit.py --rotate&lt;/code&gt;&#39;s exact 90-degree turns&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;insert.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Turn a still image into a silent, fixed-duration video clip (title card, end slate) at an exact frame size / fps, with an optional Ken Burns zoom/pan&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;background.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Generate a solid-colour or two-colour gradient clip at an exact size/duration — no input file; &lt;code&gt;--transparent&lt;/code&gt; writes a fully transparent canvas (ProRes 4444, QuickTime Animation, PNG or VP9 with alpha) to draw graphics on&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;reverse.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Reverse playback (video and, unless &lt;code&gt;--no-audio&lt;/code&gt;, audio)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;stabilize.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Two-pass motion stabilisation (&lt;code&gt;vidstabdetect&lt;/code&gt;/&lt;code&gt;vidstabtransform&lt;/code&gt;)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;sequence.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Numbered (&lt;code&gt;frame_%04d.png&lt;/code&gt;) or glob-matched still images into a video&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;waveform.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Render an audio track as a waveform or spectrum visualization video (&lt;code&gt;showwaves&lt;/code&gt;/&lt;code&gt;showspectrum&lt;/code&gt;) — for audio-only inputs with no picture worth showing; &lt;code&gt;--image&lt;/code&gt; puts the visualisation over a still plate (an audiogram), with &lt;code&gt;--platform&lt;/code&gt;, &lt;code&gt;--title&lt;/code&gt; and burnt-in captions&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;freeze.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Hold a frame for N seconds (&lt;code&gt;--at&lt;/code&gt;, &lt;code&gt;--hold&lt;/code&gt;, &lt;code&gt;--mode insert/extend&lt;/code&gt;) — an end-card hold or a comedic beat&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;pad.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Add black/silent padding at the start and/or end of the timeline (&lt;code&gt;--start&lt;/code&gt;, &lt;code&gt;--end&lt;/code&gt;) — distinct from &lt;code&gt;fit.py --fit pad&lt;/code&gt;&#39;s per-frame letterbox bars&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;speedramp.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Step through different constant speeds across a clip via &lt;code&gt;--segment START-END:FACTOR&lt;/code&gt; (repeatable) — distinct from &lt;code&gt;fit.py&lt;/code&gt;&#39;s single whole-clip speed factor&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;loop.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Repeat a clip &lt;code&gt;--times&lt;/code&gt; N or to a target &lt;code&gt;--duration&lt;/code&gt; — for background loops and filling a fixed slot length; &lt;code&gt;--crossfade&lt;/code&gt; smooths the seam, &lt;code&gt;--boomerang&lt;/code&gt; plays it forward then backward&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;broll.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Cut away to a B-roll clip over the A-roll for a window (&lt;code&gt;--insert B --at T --duration D&lt;/code&gt;, repeatable) and come back; A&#39;s length and audio untouched by default&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;metadata.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Write container chapter markers from a &lt;code&gt;TIME TITLE&lt;/code&gt; text file and title/artist/comment tags, every stream copied bit for bit; &lt;code&gt;--auto-chapters&lt;/code&gt; proposes the markers from measured pauses and scene cuts and names them &lt;code&gt;Chapter N&lt;/code&gt; for you to rename&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;grid.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Composite &lt;code&gt;--cols&lt;/code&gt;x&lt;code&gt;--rows&lt;/code&gt; clips into one grid, each cell letterboxed and labelled with its filename by default (&lt;code&gt;--label none&lt;/code&gt; to skip)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;Audio&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Tool&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;audio.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Voice clean-up chain at three strengths (&lt;code&gt;--voice light|medium|strong&lt;/code&gt;), FFT denoise, typed compressor / limiter / gate, music bed with sidechain ducking (&lt;code&gt;--duck-amount/-threshold/-attack/-release&lt;/code&gt;), a never-ducked effects bed (&lt;code&gt;--effects&lt;/code&gt;), &lt;code&gt;--stereo-widen&lt;/code&gt;, fades, 5.1 → stereo, track replacement, extraction (&lt;code&gt;-o out.wav&lt;/code&gt;), &lt;code&gt;--audio-stream N&lt;/code&gt;, a tempo change that keeps the pitch (&lt;code&gt;--tempo&lt;/code&gt;, atempo stages in 0.5–2.0) and an exact length in samples (&lt;code&gt;--length-samples&lt;/code&gt;), resampling (&lt;code&gt;--sample-rate&lt;/code&gt;) and one channel by index (&lt;code&gt;--channel N&lt;/code&gt;), each read back from the output&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;sync.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Offset between two recordings by audio cross-correlation (1 ms, pure Python), clock-drift correction; aligned video or audio out (audio-to-audio only — no lip-sync/face detection); a third or later camera is an additional &lt;code&gt;more_sources&lt;/code&gt; positional (&lt;code&gt;sync.py REF A B ...&lt;/code&gt;), the &lt;code&gt;second&lt;/code&gt; positional itself keeps its name&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;loudness.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Two-pass EBU R128 &lt;code&gt;loudnorm&lt;/code&gt; to −14 LUFS / −1 dBTP or any target (&lt;code&gt;--lra&lt;/code&gt; for the range), video stream-copied; the written file is measured again and re-encoded until it meets &lt;code&gt;--tp&lt;/code&gt; (lossy encoders overshoot); &lt;code&gt;--measure-only&lt;/code&gt;; &lt;code&gt;--versions streaming,ebu-r128,broadcast-jp&lt;/code&gt; (or &lt;code&gt;LUFS/TP&lt;/code&gt; pairs) writes one file per target from one master, each measured after and listed in &lt;code&gt;versions&lt;/code&gt; (&lt;code&gt;output&lt;/code&gt; is null), &lt;code&gt;verified&lt;/code&gt; only when every version met its own target&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;Picture&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Tool&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;caption.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Burn SRT/ASS with font, size, colour, outline, position; build SRT from timed plain text; wraps to the safe area — with &lt;code&gt;--platform&lt;/code&gt;, the destination&#39;s own left/right safe zone, which since 1.17.2 is also what the burnt ASS states as its side margins — with a phrase-aware breaker (&lt;code&gt;--wrap phrase|measured&lt;/code&gt;) and &lt;code&gt;--max-lines&lt;/code&gt;/&lt;code&gt;--min-duration&lt;/code&gt;/&lt;code&gt;--offset&lt;/code&gt;; &lt;code&gt;--mode mux&lt;/code&gt; takes a repeated &lt;code&gt;--srt file:lang&lt;/code&gt; for several language-tagged, toggleable tracks in one file; picks a font by script for non-Latin text (&lt;code&gt;--lang&lt;/code&gt;); animated and word-by-word karaoke timed to the speech energy or real word timings; &lt;code&gt;--fit-size&lt;/code&gt; shrinks the size until a cue fits &lt;code&gt;--max-lines&lt;/code&gt; instead of splitting the sentence (on by default on the delivery-template path since 1.17.1, where the size comes from the platform table); says &lt;code&gt;caption text unchanged&lt;/code&gt; when it burned the cues exactly as given; optional local transcription&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;overlay.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Logos, watermarks and titles with position, time range, opacity, fades; &lt;code&gt;--platform NAME&lt;/code&gt; keeps them clear of that destination&#39;s UI; &lt;code&gt;--video&lt;/code&gt; for picture-in-picture, &lt;code&gt;--chromakey&lt;/code&gt; for green-screen compositing, &lt;code&gt;--place&lt;/code&gt; to lay ranges of a layer at chosen times; &lt;code&gt;--keep-alpha&lt;/code&gt; composites onto a transparent canvas or alpha clip and writes the result with its alpha; a VP9 layer with alpha keeps its transparency&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;graphics.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Lower-thirds, title cards, chapter chips, progress bars, countdowns, corner bugs, social stickers, opening hook cards and meme captions drawn by FFmpeg from a brand kit; &lt;code&gt;--platform NAME&lt;/code&gt; keeps them inside that destination&#39;s safe zone; &lt;code&gt;--text-render&lt;/code&gt; routes shaping scripts through libass and &lt;code&gt;--emoji-assets&lt;/code&gt; composites colour emoji; &lt;code&gt;--keep-alpha&lt;/code&gt; renders a template on a transparent canvas as a per-element alpha layer&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;color.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;HDR10 / HLG / Dolby Vision → SDR BT.709 tone mapping, DV layer stripping, 3D LUT (.cube), colour-tag rewriting, typed primary correction (exposure/contrast/saturation/gamma/white balance/lift-gain/levels/curves); &lt;code&gt;--keep-alpha&lt;/code&gt; grades an alpha clip and keeps its alpha, checked frame by frame; &lt;code&gt;--legalize bt709|ebu-r103&lt;/code&gt;: a legal-range clamp after the 4:2:0 conversion, measured on the written file&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;Delivery&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Tool&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;export.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Presets &lt;code&gt;youtube&lt;/code&gt;, &lt;code&gt;youtube4k&lt;/code&gt;, &lt;code&gt;reels&lt;/code&gt;, &lt;code&gt;tiktok&lt;/code&gt;, &lt;code&gt;shorts&lt;/code&gt;, &lt;code&gt;linkedin&lt;/code&gt;, &lt;code&gt;facebook&lt;/code&gt;, &lt;code&gt;x&lt;/code&gt;, &lt;code&gt;youtube-hdr&lt;/code&gt; (HEVC Main10, source HDR tags kept), &lt;code&gt;youtube-av1&lt;/code&gt;, &lt;code&gt;prores&lt;/code&gt;, &lt;code&gt;dnxhr-lb&lt;/code&gt;/&lt;code&gt;sq&lt;/code&gt;/&lt;code&gt;hq&lt;/code&gt;/&lt;code&gt;hqx&lt;/code&gt;/&lt;code&gt;444&lt;/code&gt;, the alpha masters &lt;code&gt;prores4444&lt;/code&gt;/&lt;code&gt;prores4444xq&lt;/code&gt;/&lt;code&gt;qtrle&lt;/code&gt;/&lt;code&gt;vp9-alpha&lt;/code&gt;/&lt;code&gt;hap-alpha&lt;/code&gt; (alpha read back), &lt;code&gt;key-fill&lt;/code&gt; (a switcher&#39;s fill + key pair), &lt;code&gt;h265&lt;/code&gt;, &lt;code&gt;gif&lt;/code&gt;, &lt;code&gt;copy&lt;/code&gt;, all tagged BT.709 unless they carry HDR or are an editing master (&lt;code&gt;prores&lt;/code&gt;, &lt;code&gt;dnxhr-*&lt;/code&gt;: the source&#39;s tags kept, codec, profile and pixel format read back) (a BT.601-matrix or RGB source is converted to BT.709, not relabelled); &lt;code&gt;--normalize&lt;/code&gt; meets the platform&#39;s loudness in the same call (&lt;code&gt;render.py&lt;/code&gt; turns it on by default for platform presets) ; &lt;code&gt;--max-mb MB&lt;/code&gt; fits the file under a size limit by a two-pass encode and measures it ; &lt;code&gt;--preset copy --repair&lt;/code&gt; remuxes a damaged file (timestamps regenerated, corrupt packets dropped) and counts the errors before and after&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;proxy.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Small, low-bitrate proxy for downstream AI analysis/preview/editing decisions — resize by &lt;code&gt;--width&lt;/code&gt;/&lt;code&gt;--scale&lt;/code&gt;, proxy-grade &lt;code&gt;--quality&lt;/code&gt; (default 30), &lt;code&gt;--fps&lt;/code&gt;, &lt;code&gt;--no-audio&lt;/code&gt;; not a delivery preset&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;check.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;PASS / WARN / FAIL against YouTube, Shorts, Reels, TikTok, X, LinkedIn, Facebook, broadcast and podcast specs, from the same delivery table the export presets and templates read (podcast also reports chapter markers and channel count), with the fix for each failure and a &lt;code&gt;format&lt;/code&gt; / &lt;code&gt;judgement&lt;/code&gt; kind per row; a length or file size over the table&#39;s upload limit is a WARN marked &lt;code&gt;basis: guidance&lt;/code&gt; (platforms change these, the uploader checks the current one), while &lt;code&gt;--max-duration&lt;/code&gt; / &lt;code&gt;--max-mb&lt;/code&gt; state the caller&#39;s own limit and FAIL when a &lt;code&gt;--platform&lt;/code&gt; is named (with none, an over-length is WARN like every judgement row); &lt;code&gt;--content&lt;/code&gt; adds black-frame, frozen-frame, silence, short audio-dropout, single-frame-flash and flat-first-frame rows ; &lt;code&gt;--compare REF&lt;/code&gt; measures SSIM/PSNR against a reference, and with &lt;code&gt;--expect-changed SPAN|none&lt;/code&gt; fails when the picture changed outside the spans the caller declared; &lt;code&gt;--levels bt709|ebu-r103&lt;/code&gt; counts the luma and chroma samples outside the broadcast legal range, per clip and per window, with min/max Y/U/V; &lt;code&gt;--expect-identity&lt;/code&gt; checks that the file is the one a run reported writing (&lt;code&gt;output_identity&lt;/code&gt;)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;report.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Single-file HTML delivery report: before/after sheets, media facts, loudness, compliance, the commands run; &lt;code&gt;--pack&lt;/code&gt; renders a social pack table&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;Orchestration&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Tool&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;render.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Render a whole edit from a declarative &lt;code&gt;project.json&lt;/code&gt; (clips, transitions, captions, overlays including the social sticker/hook/meme graphics and a picture-in-picture or chroma-keyed &lt;code&gt;video&lt;/code&gt; layer sized for the final frame, music and stem levels, loudness, export, chapter markers, check); &lt;code&gt;--init&lt;/code&gt;, &lt;code&gt;--dry-run&lt;/code&gt;, &lt;code&gt;--stop-after&lt;/code&gt;, &lt;code&gt;--cache DIR&lt;/code&gt;/&lt;code&gt;--from STAGE&lt;/code&gt; (reuse the stages that did not change); the captions block takes the fit-size policy (&lt;code&gt;fit_size&lt;/code&gt;, &lt;code&gt;min_size&lt;/code&gt;, &lt;code&gt;fit_size_scope&lt;/code&gt;) and the run reports the caption stage&#39;s counts as &lt;code&gt;caption&lt;/code&gt;; &lt;code&gt;--template NAME INPUT&lt;/code&gt; renders a shipped delivery template (&lt;code&gt;--template all&lt;/code&gt; writes the whole social pack plus its table); &lt;code&gt;--export-timeline edit.fcpxml|.edl|.otio&lt;/code&gt; (2.1) writes the cut as an editor timeline for Final Cut, DaVinci Resolve or Premiere instead of rendering it — clips, speed, dissolves, the music bed and chapter markers carry over, and &lt;code&gt;not_exported&lt;/code&gt; names what stays behind&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;batch.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Apply a step recipe or a project to a folder with a content-hash cache; &lt;code&gt;--watch&lt;/code&gt;; &lt;code&gt;--jobs N&lt;/code&gt; processes several files at once under one shared &lt;code&gt;--timeout&lt;/code&gt;. A step that writes must name &lt;code&gt;{out}&lt;/code&gt; or its own &lt;code&gt;-o&lt;/code&gt;/&lt;code&gt;--output&lt;/code&gt;: one that writes its default output instead is refused (&lt;code&gt;step_writes_without_out&lt;/code&gt;)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;multicam.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Align any number of cameras and recorders by audio (with drift correction) and cut between them from a switch list; &lt;code&gt;--switch energy&lt;/code&gt; auto-cuts to whichever camera is loudest at that moment instead of a hand-built list (audio level as a stand-in for who is talking: there is no speaker identification, so it needs each speaker on their own camera&#39;s mic), with &lt;code&gt;--min-shot&lt;/code&gt; (minimum shot length) and &lt;code&gt;--edl&lt;/code&gt; (write a plain START-END cut list, not a CMX EDL; for Premiere/Resolve use &lt;code&gt;--write-project&lt;/code&gt; then &lt;code&gt;render.py --export-timeline&lt;/code&gt;)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;verify.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Run the toolchain on real device files and report PASS / FAIL per step&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;Not tools, but part of the surface: &lt;code&gt;mcp/server.py&lt;/code&gt; (the MCP transport) and &lt;code&gt;scripts/_contract.py&lt;/code&gt; (&lt;code&gt;contract --json&lt;/code&gt;, &lt;code&gt;doctor&lt;/code&gt;). Per-flag reference for every tool: &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/references/scripts.md&#34;&gt;references/scripts.md&lt;/a&gt;.&lt;/p&gt; &#xA;&lt;h2&gt;Audio is a first-class input&lt;/h2&gt; &#xA;&lt;p&gt;WAV, FLAC, MP3, M4A/AAC, OGG and Opus go through &lt;code&gt;probe&lt;/code&gt;, &lt;code&gt;cut&lt;/code&gt;, &lt;code&gt;join&lt;/code&gt;, &lt;code&gt;silence&lt;/code&gt;, &lt;code&gt;loudness&lt;/code&gt;, &lt;code&gt;audio&lt;/code&gt;, &lt;code&gt;sync&lt;/code&gt; and &lt;code&gt;check --platform podcast&lt;/code&gt; with the same commands as video. The output extension picks the codec: &lt;code&gt;-o out.wav&lt;/code&gt; writes PCM, &lt;code&gt;-o out.aiff&lt;/code&gt; (&lt;code&gt;.aif&lt;/code&gt;) and &lt;code&gt;-o out.caf&lt;/code&gt; 16-bit PCM in that container, &lt;code&gt;-o out.flac&lt;/code&gt; FLAC, &lt;code&gt;-o out.mp3&lt;/code&gt; MP3, &lt;code&gt;-o out.m4a&lt;/code&gt; AAC; &lt;code&gt;.wma&lt;/code&gt; is written only as a stream copy of WMA sound (&lt;code&gt;cut.py&lt;/code&gt; without &lt;code&gt;--accurate&lt;/code&gt;, &lt;code&gt;metadata.py&lt;/code&gt;), and a run that would encode into it is refused.&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;Extraction.&lt;/strong&gt; An audio extension on a video input drops the picture: &lt;code&gt;audio.py talk.mp4 -o talk.wav&lt;/code&gt;, or &lt;code&gt;--voice -o talk.m4a&lt;/code&gt; to clean it on the way (sound that starts after the picture keeps its place: the lead is written as silence). A tool that encodes a picture (and &lt;code&gt;metadata.py&lt;/code&gt; for a source with one) refuses an audio extension (&lt;code&gt;output_extension_audio_only&lt;/code&gt;). &lt;code&gt;--audio-stream N&lt;/code&gt; picks a track; &lt;code&gt;probe&lt;/code&gt; lists them under &lt;code&gt;audio_streams&lt;/code&gt;.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Join.&lt;/strong&gt; &lt;code&gt;join.py intro.wav episode.m4a outro.wav -o full.flac&lt;/code&gt; resamples every clip to one rate and channel layout and crossfades them (&lt;code&gt;--transition none&lt;/code&gt; for a butt join). Audio and video inputs cannot be mixed in one join.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Sample-accurate trims.&lt;/strong&gt; &lt;code&gt;cut.py talk.wav --start 1.2345 --end 2.3456 --accurate&lt;/code&gt; trims at the sample; the JSON reports &lt;code&gt;precision&lt;/code&gt; (&lt;code&gt;packet&lt;/code&gt; for a stream copy, &lt;code&gt;sample&lt;/code&gt; for PCM / FLAC, &lt;code&gt;codec_frame&lt;/code&gt; when a lossy encoder frames the audio again, &lt;code&gt;frame&lt;/code&gt; for video) and the measured &lt;code&gt;duration_error_ms&lt;/code&gt;. A &lt;code&gt;.wav&lt;/code&gt; never receives compressed packets.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Typed dynamics.&lt;/strong&gt; &lt;code&gt;audio.py --compress --comp-threshold -20 --comp-ratio 4&lt;/code&gt;, &lt;code&gt;--limit --limit-ceiling -1&lt;/code&gt;, &lt;code&gt;--gate --gate-threshold -45&lt;/code&gt;. Each flag is one documented option of FFmpeg&#39;s &lt;code&gt;acompressor&lt;/code&gt;, &lt;code&gt;alimiter&lt;/code&gt; or &lt;code&gt;agate&lt;/code&gt;, range-checked before ffmpeg runs; no filter string is accepted from the caller.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Loudness.&lt;/strong&gt; &lt;code&gt;loudness.py talk.wav -I -16 --tp -1.5 -o talk.m4a&lt;/code&gt; for podcast levels; &lt;code&gt;check.py talk.m4a --platform podcast&lt;/code&gt; measures LUFS and true peak and reports the chapter markers and channel count.&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;p&gt;Picture tools (&lt;code&gt;fit&lt;/code&gt;, &lt;code&gt;caption&lt;/code&gt;, &lt;code&gt;overlay&lt;/code&gt;, &lt;code&gt;graphics&lt;/code&gt;, &lt;code&gt;color&lt;/code&gt;, &lt;code&gt;export&lt;/code&gt;, &lt;code&gt;look&lt;/code&gt;) refuse an audio file with &#34;input has no video stream&#34; instead of inventing a picture. &lt;code&gt;scenes.py&lt;/code&gt; measures an audio file (&lt;code&gt;audio_only: true&lt;/code&gt;, one scene): &lt;code&gt;--beats&lt;/code&gt;, &lt;code&gt;--audio-peaks&lt;/code&gt;, &lt;code&gt;--speech&lt;/code&gt; and &lt;code&gt;--highlights&lt;/code&gt; run on the sound, and &lt;code&gt;--sheet&lt;/code&gt; / &lt;code&gt;--shots&lt;/code&gt; are refused.&lt;/p&gt; &#xA;&lt;h2&gt;Built for agents&lt;/h2&gt; &#xA;&lt;h3&gt;What is SPEC?&lt;/h3&gt; &#xA;&lt;p&gt;&lt;strong&gt;SPEC&lt;/strong&gt; (Self-Producing Execution Contract) is the name this project&#39;s author, &lt;a href=&#34;https://github.com/kajisho5&#34;&gt;kajisho5&lt;/a&gt;, gave the pattern the tool layer is built on: each tool&#39;s &lt;code&gt;input_schema&lt;/code&gt; — the part of its contract that has to track the CLI exactly, flag for flag — is never hand-authored side by side with the code. It is derived, at run time, from the one thing that has to be correct for the CLI to work at all: the script&#39;s own &lt;code&gt;argparse&lt;/code&gt; parser.&lt;/p&gt; &#xA;&lt;p&gt;Concretely, &lt;code&gt;scripts/_contract.py&lt;/code&gt;&#39;s &lt;code&gt;_capture_parser()&lt;/code&gt; imports every tool script and intercepts its &lt;code&gt;parse_args()&lt;/code&gt; call to get the live, fully-built parser object — flags, types, choices, defaults, required/positional, mutually exclusive groups, all of it. &lt;code&gt;input_schema&lt;/code&gt; is built straight from that object. (The rest of a &lt;code&gt;ToolSpec&lt;/code&gt; — &lt;code&gt;role&lt;/code&gt;, &lt;code&gt;capabilities&lt;/code&gt;, &lt;code&gt;inputs&lt;/code&gt;, &lt;code&gt;outputs&lt;/code&gt;, &lt;code&gt;output_schema&lt;/code&gt; — comes from a hand-authored table, &lt;code&gt;TOOL_META&lt;/code&gt;, since those facts aren&#39;t things a parser can express; only &lt;code&gt;input_schema&lt;/code&gt; is parser-derived.)&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;The contract&lt;/strong&gt;&#39;s &lt;code&gt;input_schema&lt;/code&gt; for every tool is generated from the live parser directly.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;&lt;a href=&#34;http://SKILL.md&#34;&gt;SKILL.md&lt;/a&gt; is two-tier&lt;/strong&gt; (1.11.0): the file the agent loads every session keeps the workflow, the request→script table and one line per gotcha; the long-form detail lives in &lt;code&gt;references/gotchas.md&lt;/code&gt; and the other &lt;code&gt;references/&lt;/code&gt; files, read only when a job needs it.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;The MCP server&lt;/strong&gt; (&lt;code&gt;mcp/server.py&lt;/code&gt;) carries no schema of its own; &lt;code&gt;tools/list&lt;/code&gt; is translated straight from the contract, &lt;code&gt;input_schema&lt;/code&gt; included.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;The docs&lt;/strong&gt; (&lt;code&gt;docs/contract.md&lt;/code&gt;&#39;s field reference, this README&#39;s tool table) describe the same shape. &lt;code&gt;tests/test_contract.py&lt;/code&gt; runs on every CI run and fails the build if any of them drift out of sync with what the code actually does — it catches drift, it doesn&#39;t fix it for you.&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;p&gt;So adding a flag to a script&#39;s &lt;code&gt;argparse&lt;/code&gt; block updates &lt;code&gt;input_schema&lt;/code&gt; and the MCP tool definition with no second edit, and a docs page or &lt;code&gt;TOOL_META&lt;/code&gt; entry that falls behind fails CI rather than drifting silently. There is no separate &lt;code&gt;input_schema&lt;/code&gt; file to forget to update.&lt;/p&gt; &#xA;&lt;h3&gt;Machine-readable contract&lt;/h3&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;npx ffmpeg-skill contract --json            # or: python3 scripts/_contract.py --json&#xA;npx ffmpeg-skill contract --json --static   # without environment detection&#xA;npx ffmpeg-skill contract --json --tool cut # one tool&#39;s entry under `tools` (repeatable): ~60 KB, not ~950 KB&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;The contract is generated from the code that runs, not maintained beside it. For each of the 42 tools (&lt;code&gt;ffmpeg-skill/&amp;lt;name&amp;gt;&lt;/code&gt;) it states:&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Field&lt;/th&gt; &#xA;   &lt;th&gt;Meaning&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;input_schema&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;generated from the tool&#39;s argparse parser: properties, types, enums, defaults, required, positional order, mutually exclusive groups&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;output_schema&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;what &lt;code&gt;--json&lt;/code&gt; prints: &lt;code&gt;status&lt;/code&gt;, &lt;code&gt;output&lt;/code&gt;, &lt;code&gt;commands&lt;/code&gt;, &lt;code&gt;probe&lt;/code&gt;, &lt;code&gt;output_identity&lt;/code&gt; (&lt;code&gt;{size, mtime_ns}&lt;/code&gt; of the file written, on every tool but &lt;code&gt;probe&lt;/code&gt;), plus tool-specific fields (&lt;code&gt;precision&lt;/code&gt;, &lt;code&gt;checks&lt;/code&gt;, &lt;code&gt;offset_seconds&lt;/code&gt;, …)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;role&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;analysis&lt;/code&gt;, &lt;code&gt;analysis_and_execution&lt;/code&gt;, &lt;code&gt;execution&lt;/code&gt; or &lt;code&gt;verification&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;capabilities&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;the FFmpeg encoders, filters and bitstream filters the tool always needs, and the ones needed only for a flag or input&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;supports_dry_run&lt;/code&gt;, &lt;code&gt;supports_json&lt;/code&gt;, &lt;code&gt;supports_json_brief&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;measured by the tests, not declared&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;verification&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;which tools to run on the output afterwards (&lt;code&gt;probe&lt;/code&gt;, &lt;code&gt;check&lt;/code&gt;, &lt;code&gt;look&lt;/code&gt;)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;requires_visual_verification&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;the picture changed; inspect the contact sheet&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;audio_only&lt;/code&gt;, &lt;code&gt;video_required&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;whether an audio-only input is accepted or refused&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;mutates_input&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;always &lt;code&gt;false&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;idempotency_hint&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;bit_exact&lt;/code&gt;, &lt;code&gt;content_equivalent&lt;/code&gt;, &lt;code&gt;cached&lt;/code&gt; or &lt;code&gt;environment_dependent&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;Next to the tool list the document carries a top-level &lt;code&gt;deprecated&lt;/code&gt; list (since 1.10): what the next major removes, since when, the replacement and the surface it lives on. It is empty in 3.0.0: Python 3.9 and FFmpeg 5.x, the two 2.x entries, moved to &lt;code&gt;removed&lt;/code&gt; with what 2.0 removed, and &lt;code&gt;docs/contract.md&lt;/code&gt; &#34;What 2.0 changed&#34; and &#34;What 3.0 changed&#34; are written from that.&lt;/p&gt; &#xA;&lt;p&gt;&lt;code&gt;contract_version&lt;/code&gt; (2.0) is separate from the skill version, so a consumer can pin the shape and read the version for provenance. The document also states the invocation mapping (structured arguments → argv), the JSON shapes for success and failure (&lt;code&gt;{&#34;status&#34;: &#34;failed&#34;, &#34;error&#34;: {&#34;kind&#34;: &#34;input | ffmpeg | output | missing_tool | timeout | verification | interrupted&#34;, &#34;message&#34;: …}}&lt;/code&gt;; a refusal adds &lt;code&gt;error.reason&lt;/code&gt;, a stable code such as &lt;code&gt;output_exists&lt;/code&gt; or &lt;code&gt;plan_inputs_changed&lt;/code&gt;, and &lt;code&gt;error.details&lt;/code&gt;, the facts behind it; warnings, fallbacks and skipped steps a run decided on its own are in &lt;code&gt;diagnostics&lt;/code&gt;), and that no tool runs a shell or executes anything other than the named script, &lt;code&gt;ffmpeg&lt;/code&gt; and &lt;code&gt;ffprobe&lt;/code&gt;. Field-by-field reference: &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/contract.md&#34;&gt;docs/contract.md&lt;/a&gt;.&lt;/p&gt; &#xA;&lt;h3&gt;MCP&lt;/h3&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-json&#34;&gt;{&#34;mcpServers&#34;: {&#34;ffmpeg-skill&#34;: {&#34;command&#34;: &#34;python3&#34;, &#34;args&#34;: [&#34;/Users/you/.claude/skills/ffmpeg-skill/mcp/server.py&#34;]}}}&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;On Windows, &lt;code&gt;python3&lt;/code&gt; is only on PATH if Python was installed from the Microsoft Store; a &lt;a href=&#34;http://python.org&#34;&gt;python.org&lt;/a&gt; install exposes &lt;code&gt;python&lt;/code&gt; (or the &lt;code&gt;py&lt;/code&gt; launcher) instead — if your MCP client reports the server failed to start, change &lt;code&gt;&#34;command&#34;&lt;/code&gt; above to &lt;code&gt;&#34;python&#34;&lt;/code&gt; (or the full path from &lt;code&gt;where python&lt;/code&gt;).&lt;/p&gt; &#xA;&lt;p&gt;&lt;code&gt;mcp/server.py&lt;/code&gt; is a stdio JSON-RPC transport with no tool table of its own. &lt;code&gt;tools/list&lt;/code&gt; is derived from the contract at start-up, in contract order, with &lt;code&gt;inputSchema&lt;/code&gt; translated from each tool&#39;s &lt;code&gt;input_schema&lt;/code&gt;. By default it lists only the core 12 (&lt;code&gt;render&lt;/code&gt;, &lt;code&gt;look&lt;/code&gt;, &lt;code&gt;caption&lt;/code&gt;, &lt;code&gt;export&lt;/code&gt;, &lt;code&gt;check&lt;/code&gt;, &lt;code&gt;fit&lt;/code&gt;, &lt;code&gt;cut&lt;/code&gt;, &lt;code&gt;audio&lt;/code&gt;, &lt;code&gt;loudness&lt;/code&gt;, &lt;code&gt;graphics&lt;/code&gt;, &lt;code&gt;silence&lt;/code&gt;, &lt;code&gt;probe&lt;/code&gt;) so a client doesn&#39;t pay context for 30 schemas it rarely calls directly; set &lt;code&gt;FFMPEG_SKILL_MCP_FULL=1&lt;/code&gt; to list all 42. Every tool, listed or not, is callable through &lt;code&gt;tools/call&lt;/code&gt;, which maps structured arguments to argv and runs the named script; a raw &lt;code&gt;argv&lt;/code&gt; form is accepted for compatibility and marked non-canonical. &lt;code&gt;python3 mcp/server.py --list&lt;/code&gt; prints the tools; &lt;code&gt;--call probe &#39;{&#34;inputs&#34;: [&#34;a.mp4&#34;]}&#39;&lt;/code&gt; runs one from the shell.&lt;/p&gt; &#xA;&lt;p&gt;No &lt;code&gt;inputSchema&lt;/code&gt; lists &lt;code&gt;json&lt;/code&gt; or &lt;code&gt;progress&lt;/code&gt;: they are transport flags the server sets itself, not tool arguments (2.0; 1.x listed them unless &lt;code&gt;FFMPEG_SKILL_MCP_LEAN=1&lt;/code&gt;). A client that still sends them is not refused.&lt;/p&gt; &#xA;&lt;p&gt;A structured value is never read as an option: one that starts with &lt;code&gt;-&lt;/code&gt; (overlay text &lt;code&gt;-50%&lt;/code&gt;) is passed as &lt;code&gt;--flag=value&lt;/code&gt;, and positionals that start with &lt;code&gt;-&lt;/code&gt; follow &lt;code&gt;--&lt;/code&gt;. &lt;code&gt;probe&lt;/code&gt;, &lt;code&gt;check&lt;/code&gt; and &lt;code&gt;cropdetect&lt;/code&gt; are annotated &lt;code&gt;readOnlyHint&lt;/code&gt; and write no file over MCP (their &lt;code&gt;plan&lt;/code&gt; is refused, structured or in a raw &lt;code&gt;argv&lt;/code&gt;); every other tool is &lt;code&gt;destructiveHint&lt;/code&gt;, because with &lt;code&gt;overwrite&lt;/code&gt; it replaces an existing file. See &lt;code&gt;docs/contract.md&lt;/code&gt; § MCP relationship.&lt;/p&gt; &#xA;&lt;p&gt;The server also advertises a &lt;code&gt;prompts&lt;/code&gt; capability: five canned workflow recipes an agent can ask for by name instead of composing the individual tool calls itself — &lt;code&gt;reel&lt;/code&gt;, &lt;code&gt;podcast&lt;/code&gt;, &lt;code&gt;multicam&lt;/code&gt;, &lt;code&gt;delivery_check&lt;/code&gt;, &lt;code&gt;hdr&lt;/code&gt;. &lt;code&gt;prompts/list&lt;/code&gt; returns each one&#39;s name, description and arguments; &lt;code&gt;prompts/get {name, arguments}&lt;/code&gt; fills the matching template and returns one text message naming the actual command lines to run and what to check afterward. These are recipes, not new tool calls — every command line a prompt names is one &lt;code&gt;tools/call&lt;/code&gt; (or the CLI) can already run.&lt;/p&gt; &#xA;&lt;h3&gt;Capability detection&lt;/h3&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;npx ffmpeg-skill doctor          # human-readable&#xA;npx ffmpeg-skill doctor --json   # available / missing / missing_optional / unknown / detection / errors / tools / gpu_encoders / ffmpeg_security&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;&lt;code&gt;doctor&lt;/code&gt; reads &lt;code&gt;ffmpeg -encoders&lt;/code&gt;, &lt;code&gt;-decoders&lt;/code&gt;, &lt;code&gt;-filters&lt;/code&gt; and &lt;code&gt;-bsfs&lt;/code&gt; and resolves every capability the contract declares against this machine&#39;s build. Three states per capability: &lt;code&gt;available&lt;/code&gt;, &lt;code&gt;missing&lt;/code&gt;, &lt;code&gt;unknown&lt;/code&gt;. Exit 0 when everything required is available, 1 when something required is missing, 2 when nothing is proven missing but a required capability is unknown. With detection on (the default), &lt;code&gt;contract --json&lt;/code&gt; carries the same lists under &lt;code&gt;capabilities&lt;/code&gt;. &lt;code&gt;doctor&lt;/code&gt; also reports fonts: the default drawtext family, and &lt;code&gt;fonts.scripts&lt;/code&gt; — one &lt;code&gt;available&lt;/code&gt;/&lt;code&gt;missing&lt;/code&gt;/&lt;code&gt;unknown&lt;/code&gt; per writing system (ja, zh, ko, ar, he, hi, bn, ta, th, lo, ru, el, latin) with the file it would use — so &#34;can this machine render Korean captions&#34; is answered before the job, not after. &lt;code&gt;doctor --json&lt;/code&gt;&#39;s &lt;code&gt;tools&lt;/code&gt; field folds that down to one answer per tool — &lt;code&gt;{&#34;caption&#34;: {&#34;usable&#34;: &#34;no&#34;, &#34;missing&#34;: [&#34;filter:subtitles&#34;], &#34;fix&#34;: &#34;...&#34;}, ...}&lt;/code&gt; — so &#34;is &lt;code&gt;doctor&lt;/code&gt; overall &lt;code&gt;ok&lt;/code&gt;&#34; and &#34;can I run &lt;code&gt;caption.py&lt;/code&gt; on this machine&#34; are answered separately: a plain Homebrew &lt;code&gt;ffmpeg&lt;/code&gt; is &lt;code&gt;ok&lt;/code&gt; for tools that don&#39;t need &lt;code&gt;subtitles&lt;/code&gt;/&lt;code&gt;drawtext&lt;/code&gt;/&lt;code&gt;zscale&lt;/code&gt;, while &lt;code&gt;caption&lt;/code&gt;&#39;s own &lt;code&gt;usable&lt;/code&gt; is &lt;code&gt;&#34;no&#34;&lt;/code&gt;.&lt;/p&gt; &#xA;&lt;p&gt;&lt;code&gt;doctor --json&lt;/code&gt;&#39;s &lt;code&gt;gpu_encoders&lt;/code&gt; reports which GPU-backed encoders (&lt;code&gt;nvenc&lt;/code&gt;, &lt;code&gt;videotoolbox&lt;/code&gt;, &lt;code&gt;qsv&lt;/code&gt;, &lt;code&gt;vaapi&lt;/code&gt;, &lt;code&gt;amf&lt;/code&gt;) this ffmpeg &lt;em&gt;build&lt;/em&gt; was compiled with — read from &lt;code&gt;-encoders&lt;/code&gt; alone, so it proves the capability shipped, not that the GPU/driver on this machine will actually accept a job (that needs a real encode, which &lt;code&gt;doctor&lt;/code&gt;&#39;s introspection never runs). It never affects &lt;code&gt;ok&lt;/code&gt; or any tool&#39;s &lt;code&gt;usable&lt;/code&gt;. The one GPU path a tool uses is Apple VideoToolbox, opt-in: &lt;code&gt;--hw&lt;/code&gt; on every re-encoding tool (h264/hevc/prores on the GPU, AV1 stays on SVT-AV1), &lt;code&gt;FFMPEG_SKILL_HW=1&lt;/code&gt; to make it this machine&#39;s default for everything except &lt;code&gt;export.py&lt;/code&gt;&#39;s delivery presets (those take &lt;code&gt;--hw&lt;/code&gt; explicitly, or &lt;code&gt;render.py --hw&lt;/code&gt; for a whole project). It runs only on Apple Silicon and never by default, except on a build with no &lt;code&gt;libx264&lt;/code&gt;/&lt;code&gt;libx265&lt;/code&gt; (an LGPL ffmpeg that has only &lt;code&gt;h264_videotoolbox&lt;/code&gt;/&lt;code&gt;hevc_videotoolbox&lt;/code&gt;): there it is the only encoder, so every re-encoding tool and &lt;code&gt;export.py&lt;/code&gt;&#39;s presets use it without &lt;code&gt;--hw&lt;/code&gt;, the result says &lt;code&gt;hw.source: &#34;auto&#34;&lt;/code&gt;, and &lt;code&gt;doctor&lt;/code&gt; lists the missing x264/x265 under &lt;code&gt;covered_by&lt;/code&gt; instead of calling the tools unusable (&lt;code&gt;--no-hw&lt;/code&gt; still asks for the CPU). It falls back to the CPU encoder — reported in the result&#39;s &lt;code&gt;hw.fallback&lt;/code&gt;/&lt;code&gt;hw.notes&lt;/code&gt; — only when VideoToolbox itself refuses a job (any other ffmpeg failure is reported as it happened, not retried); every result names the &lt;code&gt;encoder&lt;/code&gt; that actually ran (&lt;code&gt;batch.py&lt;/code&gt;, which encodes nothing itself: each &lt;code&gt;results&lt;/code&gt; row), and a GPU encode adds a &lt;code&gt;notes&lt;/code&gt; line that its quality is not CRF-equivalent. Measured on an M4 Max, 2–7× faster than x264/x265 &lt;code&gt;medium&lt;/code&gt;; at matched SSIM (&lt;code&gt;tests/bench_vt.py&lt;/code&gt;) 1.1–2.9× the bytes on SDR and 1.9–3.5× on HDR, which is why it is a flag and not the default (&lt;code&gt;docs/design-decisions.md&lt;/code&gt;). &lt;code&gt;doctor --json&lt;/code&gt;&#39;s &lt;code&gt;hw&lt;/code&gt; says whether this machine can run it (&lt;code&gt;platform_ok&lt;/code&gt;) and whether it is the default here (&lt;code&gt;default_on&lt;/code&gt;).&lt;/p&gt; &#xA;&lt;p&gt;&lt;code&gt;doctor&lt;/code&gt; also compares the FFmpeg version with a small, dated table of high-severity FFmpeg CVEs (&lt;code&gt;doctor --json&lt;/code&gt;&#39;s &lt;code&gt;ffmpeg_security&lt;/code&gt;; source &lt;a href=&#34;https://ffmpeg.org/security.html&#34;&gt;ffmpeg.org/security.html&lt;/a&gt;, checked 2026-10-22). ffmpeg-skill does not ship FFmpeg, so a decoder bug in the build you have stays yours until you upgrade it. When the version is older than the first upstream release on its branch that carries a fix (today: CVE-2026-8461, the MagicYUV decoder heap write, fixed in 8.0.3, 8.1.2 and 9.0), &lt;code&gt;doctor&lt;/code&gt; prints one &lt;code&gt;warning:&lt;/code&gt; line on stderr with the update command and lists it under &lt;code&gt;ffmpeg_security.warnings&lt;/code&gt;. It is advice, never a failure: &lt;code&gt;ok&lt;/code&gt;, each tool&#39;s &lt;code&gt;usable&lt;/code&gt; and the exit code do not change. A distribution package (Ubuntu &lt;code&gt;8.0.1-3ubuntu2&lt;/code&gt;, Debian &lt;code&gt;+deb13u1&lt;/code&gt;) may carry the fix as a backport under the old version number, so the line says the build is older than the fix, not that it is vulnerable. A version it cannot read (a git snapshot such as &lt;code&gt;N-12345-g…&lt;/code&gt;) gets no warning, and nothing is fetched: the table ships with the skill.&lt;/p&gt; &#xA;&lt;h2&gt;Gotchas and best practices&lt;/h2&gt; &#xA;&lt;p&gt;The short list for humans. The agent-facing version, with the reasoning, is the &#34;Things that look right but are wrong&#34; and &#34;Gotchas&#34; sections of &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/SKILL.md&#34;&gt;SKILL.md&lt;/a&gt;.&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;Variable frame rate (phone and screen recordings).&lt;/strong&gt; &lt;code&gt;probe.py&lt;/code&gt; flags it; every re-encoding tool conforms to a constant rate automatically. &lt;code&gt;cut.py&lt;/code&gt; samples the frame timing before a copy and switches to frame-accurate mode when it is variable or too short to judge, because copy-cuts on VFR land on the wrong frame (&lt;code&gt;--vfr-guard average&lt;/code&gt; restores 2.x&#39;s rule, re-encoding whenever the nominal and average rates differ; &lt;code&gt;--vfr-guard off&lt;/code&gt; keeps the copy). Choose the rate yourself with &lt;code&gt;fit.py input.mp4 --fps 30&lt;/code&gt; when the measured average is odd.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Lossless cuts snap to keyframes.&lt;/strong&gt; A stream-copy cut can start up to one GOP earlier than asked. &lt;code&gt;cut.py&lt;/code&gt; re-encodes when the snap exceeds 0.5 s (&lt;code&gt;--tolerance&lt;/code&gt; changes the limit). For a strictly lossless file pass &lt;code&gt;--tolerance -1&lt;/code&gt; (plus &lt;code&gt;--vfr-guard off&lt;/code&gt; on a variable-frame-rate source), and expect the cut to land on the nearest earlier keyframe; the JSON result lists them under &lt;code&gt;nearest_keyframes&lt;/code&gt;. A &lt;code&gt;--segments&lt;/code&gt; join can still be re-cut: open-GOP video, a later segment of B-frame &lt;code&gt;.mp4&lt;/code&gt;/&lt;code&gt;.mov&lt;/code&gt; video that starts between keyframes (start it on a keyframe, or write &lt;code&gt;.mkv&lt;/code&gt;), or a join that fails its check. A single &lt;code&gt;.mp4&lt;/code&gt;/&lt;code&gt;.mov&lt;/code&gt; copy keeps its edit list by default (since 3.0.0), so the picture starts where asked: the keyframe&#39;s pre-roll is stored but hidden by an MP4 edit list, which a player that ignores edit lists shows (&lt;code&gt;--no-edit-list&lt;/code&gt; starts the file at the keyframe instead).&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Frame-exact without re-encoding the whole cut (H.264).&lt;/strong&gt; &lt;code&gt;cut.py --smart&lt;/code&gt; re-encodes only from each cut point to the next keyframe and copies the GOPs between, matched to the source&#39;s profile, level, pixel format, B-frame depth and colour tags. It checks the result against the source (frame grid, copied packets, decoded frames). A source it cannot match (HEVC, 10-bit, interlaced, VFR, open GOPs) is re-encoded in full, and &lt;code&gt;smart.reason&lt;/code&gt; says why. &lt;code&gt;probe.py --keyframes-near T&lt;/code&gt; shows where the keyframes are.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;HDR stays HDR.&lt;/strong&gt; When the probe reports HDR (HDR10, HLG, Dolby Vision, BT.2020), the tools keep it rather than flatten it. Convert deliberately with &lt;code&gt;color.py --to-sdr&lt;/code&gt; before H.264 deliverables or LUT work. &lt;code&gt;export.py&lt;/code&gt; platform presets are SDR and warn on HDR input.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Loudness targets.&lt;/strong&gt; −14 LUFS / −1 dBTP for YouTube and social platforms (the &lt;code&gt;loudness.py&lt;/code&gt; default), &lt;code&gt;-I -16 --tp -1.5&lt;/code&gt; for podcasts, &lt;code&gt;-I -23&lt;/code&gt; for broadcast. A clip measured at −40 LUFS or below is room tone, not content; raising it raises the noise. Check true peak as well as LUFS: &lt;code&gt;check.py file --platform podcast&lt;/code&gt; measures both.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Alpha is kept only when asked, and a lost alpha is not verified.&lt;/strong&gt; The re-encoding tools write opaque H.264 by default. On a source whose alpha is used (ProRes 4444, QuickTime Animation, PNG, VP9 with alpha), &lt;code&gt;color.py&lt;/code&gt;, &lt;code&gt;graphics.py&lt;/code&gt; and &lt;code&gt;overlay.py&lt;/code&gt; take &lt;code&gt;--keep-alpha&lt;/code&gt; (ProRes 4444 by default, &lt;code&gt;--alpha-codec qtrle|png|vp9&lt;/code&gt;, &lt;code&gt;--alpha-mode straight|premultiplied&lt;/code&gt;) and read the alpha back; without it the result says &lt;code&gt;verified: false&lt;/code&gt; with a failing &lt;code&gt;alpha&lt;/code&gt; step, and &lt;code&gt;--drop-alpha&lt;/code&gt; records that an opaque output is wanted. &lt;code&gt;background.py --transparent&lt;/code&gt; makes the canvas to draw graphics on.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Broadcast legal levels are measured, not assumed.&lt;/strong&gt; &lt;code&gt;check.py final.mxf --levels ebu-r103&lt;/code&gt; (or &lt;code&gt;bt709&lt;/code&gt;) counts every luma and chroma sample outside the legal range and FAILs when one frame has more than 1 % out (&lt;code&gt;--levels-tolerance&lt;/code&gt;); the JSON &lt;code&gt;levels&lt;/code&gt; adds min/max Y/U/V per &lt;code&gt;--levels-window&lt;/code&gt;. &lt;code&gt;color.py --correct --legalize ebu-r103&lt;/code&gt; clamps after the 4:2:0 conversion and measures what the encode left: a lossy encoder can push a few samples around a clamped edge back out, so the result says how many.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Frame changes first, text second.&lt;/strong&gt; Captions and overlays burned before a crop or resize end up off-frame. Reframe, then caption.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Cropping 16:9 to 9:16 discards 70 % of the width.&lt;/strong&gt; &lt;code&gt;fit.py --fit crop&lt;/code&gt; centres by default; pass &lt;code&gt;--crop-x&lt;/code&gt;/&lt;code&gt;--crop-y&lt;/code&gt; toward the subject, or pad with &lt;code&gt;--fit pad --pad-fill blur&lt;/code&gt;. Look at the contact sheet before deciding.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Phrase-aware caption breaking (1.16).&lt;/strong&gt; &lt;code&gt;caption.py&lt;/code&gt;/&lt;code&gt;graphics.py --wrap phrase&lt;/code&gt; (the default) never breaks inside a word or on the wrong side of a hyphen, never leaves a lone digit, kana or punctuation pair on a line, prefers Japanese sentence ends and particles over a mid-word break, and never ends a line on an article or preposition. All four are penalties over break positions that already fit, so no line is widened and the line count never changes; &lt;code&gt;--wrap measured&lt;/code&gt; restores 1.15&#39;s width-only wrap. The text itself is never rewritten or shortened. Since 1.16.1 a Thai run and a katakana word are never broken inside (Thai writes no space inside a phrase and there is no dictionary: the break goes where you put a space or &lt;code&gt;|&lt;/code&gt;), and a line wider than the safe width is reported as &lt;code&gt;overlong&lt;/code&gt; with the fix named.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Caption size fitted to the cue (1.17, reachable from the templates since 1.17.1).&lt;/strong&gt; At a platform caption size a line holds about six em, so an ordinary sentence needs four lines and &lt;code&gt;--max-lines 2&lt;/code&gt; used to cut it into consecutive cues — half the sentence arriving late. &lt;code&gt;caption.py --fit-size&lt;/code&gt; (default &lt;code&gt;auto&lt;/code&gt;) walks the size down until every cue fits, &lt;em&gt;before&lt;/em&gt; laying the cues out, with a legibility floor of 4.5 % of the frame height (&lt;code&gt;--min-size&lt;/code&gt;, default 13 ASS units). An explicit &lt;code&gt;--size&lt;/code&gt; or a &lt;code&gt;brand.json&lt;/code&gt; size is a statement about the look and is never overridden — which in 1.17.0 also silenced the fitter on every &lt;code&gt;render.py --template&lt;/code&gt; run, since a template fills the size from the platform table; 1.17.1 marks that size as the default it is (&lt;code&gt;&#34;fit_size&#34;: &#34;on&#34;&lt;/code&gt; in the filled project), so the type shrinks and no cue is split. &lt;code&gt;--fit-size off&lt;/code&gt; restores that earlier behaviour byte for byte, and the caption text is still never rewritten to make it fit.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Beat-synced cuts (1.17).&lt;/strong&gt; &lt;code&gt;scenes.py --beats&lt;/code&gt; reports the measured grid — tempo, beat times, and a confidence built from how far the winning autocorrelation lag stands above the others and how many onsets land on it. &lt;code&gt;cut.py --snap beats&lt;/code&gt; (and a &lt;code&gt;&#34;snap&#34;&lt;/code&gt; block in a &lt;code&gt;render.py&lt;/code&gt; project) moves in/out points to the nearest beat within &lt;code&gt;--snap-tolerance&lt;/code&gt; — and only onto the grid points a measured onset actually marks, never onto the regular grid&#39;s continuation through a passage with no music in it. Below &lt;code&gt;--min-confidence&lt;/code&gt; it &lt;strong&gt;refuses&lt;/strong&gt;: a cut point may move to a measured beat and may not appear from one, so speech and ambience get an honest &#34;no steady pulse here&#34; instead of an invented grid. A music bed is measured as it is delivered, WAV/FLAC/MP3 with no picture; on a synthetic click track at 90 and 120 BPM every beat is within 1 ms (WAV/FLAC) or 9 ms (MP3) of its click, &lt;code&gt;tests/test_analysis.py&lt;/code&gt; holds it to 20 ms.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Filler words (1.17).&lt;/strong&gt; &lt;code&gt;silence.py --filler --words transcript.json&lt;/code&gt; removes &#34;um&#34; and &#34;uh&#34; through the same cut graph the silences use. Never without measured word timings — there is no heuristic that finds an &#34;um&#34; without them that would not also cut real speech — and &lt;code&gt;like&lt;/code&gt;, &lt;code&gt;tipo&lt;/code&gt; and &lt;code&gt;cioè&lt;/code&gt; are deliberately not in the built-in lists, because a discourse marker is a content word. Built-in lists since 2.19: English, Japanese, Korean, Chinese, Spanish, German, French, Portuguese, Italian, Dutch, Polish, Russian and Turkish; any other language takes &lt;code&gt;--filler-words FILE&lt;/code&gt;. Whisper stays optional: &lt;code&gt;--transcribe&lt;/code&gt; with no engine installed refuses and names the five installs.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Throughput (1.17).&lt;/strong&gt; &lt;code&gt;batch.py --jobs N&lt;/code&gt; runs several files at once, capped at &lt;code&gt;min(N, cpu_count, 8)&lt;/code&gt; and sharing one &lt;code&gt;--timeout&lt;/code&gt; budget rather than one per item; the per-item table keeps its order. &lt;code&gt;render.py --cache DIR&lt;/code&gt; reuses stages whose inputs and arguments did not change, so swapping an export preset re-runs export only. The cache is opt-in with no default directory, and the ffmpeg, skill and contract versions are part of every key, so a cache is never reused across them.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Audiogram (1.16).&lt;/strong&gt; &lt;code&gt;waveform.py --image cover.png&lt;/code&gt; (or &lt;code&gt;render.py --template audiogram&lt;/code&gt;) puts the waveform over a still plate for an episode that has no picture, with &lt;code&gt;--platform&lt;/code&gt; for the frame, &lt;code&gt;--title&lt;/code&gt; and burnt-in captions. The image is a local file you give: nothing is fetched and no cover art is ever invented.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Emoji in captions and titles (1.15).&lt;/strong&gt; &lt;code&gt;caption.py&lt;/code&gt;/&lt;code&gt;graphics.py --emoji-assets DIR&lt;/code&gt; composites a PNG per emoji (Twemoji naming, &lt;code&gt;1f389.png&lt;/code&gt;, or Noto&#39;s &lt;code&gt;emoji_u1f389.png&lt;/code&gt;) on top of the text, because drawtext cannot load a colour emoji font at all and an installed one does not prove libass will draw it in colour — &lt;code&gt;doctor --json&lt;/code&gt;&#39;s &lt;code&gt;fonts.emoji&lt;/code&gt; answers that from a render probe. Without assets the result&#39;s &lt;code&gt;emoji.mode&lt;/code&gt; says which of three things happened: &lt;code&gt;color&lt;/code&gt; when this build&#39;s libass draws the colour glyph (also for &lt;code&gt;--emoji mono&lt;/code&gt; on such a build, with a warning), &lt;code&gt;mono&lt;/code&gt; when it draws a monochrome one, and &lt;code&gt;none&lt;/code&gt; when the emoji are left out of the text: with an &lt;code&gt;emoji_not_drawn&lt;/code&gt; warning when that happened because nothing here draws the glyph, and with only an info line when &lt;code&gt;--emoji none&lt;/code&gt; asked for it (and a &lt;code&gt;graphics.py&lt;/code&gt; title that is nothing but emoji is refused). Nothing is ever downloaded.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Indic and Thai text shaped correctly in titles and lower-thirds (1.15).&lt;/strong&gt; &lt;code&gt;graphics.py&lt;/code&gt; renders Devanagari, Bengali, Tamil, Thai and Lao through libass automatically (&lt;code&gt;text_renderer: &#34;ass&#34;&lt;/code&gt;), because drawtext never reorders matras or re-clusters marks; Arabic and Hebrew were already correct on a fribidi build. &lt;code&gt;--text-render drawtext&lt;/code&gt; with such a script is refused, never rendered wrongly.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Non-Latin text picks a font by script (1.12).&lt;/strong&gt; Japanese, Chinese, Korean, Arabic, Hebrew, Devanagari, Thai, Cyrillic and Greek cues, titles and overlays resolve a font file that covers them automatically, and a machine with no such font fails the job (&lt;code&gt;kind: input&lt;/code&gt;) instead of rendering boxes. &lt;code&gt;doctor --json&lt;/code&gt;&#39;s &lt;code&gt;fonts.scripts&lt;/code&gt; says which languages this machine can render; &lt;code&gt;--lang ja|ko&lt;/code&gt; disambiguates Han-only text, and &lt;code&gt;--lang zh-TW&lt;/code&gt; (or &lt;code&gt;zh-HK&lt;/code&gt;, &lt;code&gt;zh-Hant&lt;/code&gt;) picks a Traditional Chinese face (2.19); an explicit &lt;code&gt;--font&lt;/code&gt;/&lt;code&gt;--font-file&lt;/code&gt; is always kept.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Silence detection finds nothing?&lt;/strong&gt; The default threshold is −35 dBFS. The tool prints a hint with the track&#39;s measured level; raise the threshold (&lt;code&gt;silence.py --threshold -25&lt;/code&gt;) or shorten &lt;code&gt;--min-silence&lt;/code&gt;.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Sync results carry a confidence.&lt;/strong&gt; Below 0.3, or an offset near the edge of the analysis window, is probably wrong: enlarge &lt;code&gt;--analyze-seconds&lt;/code&gt; or find a clap. Recordings over ten minutes from separate devices need &lt;code&gt;sync.py --fix-drift&lt;/code&gt;.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Outputs are never overwritten silently.&lt;/strong&gt; An existing output path is refused (&lt;code&gt;kind: input&lt;/code&gt;, &lt;code&gt;error.reason: output_exists&lt;/code&gt;, before anything runs); pass &lt;code&gt;--overwrite&lt;/code&gt; where a replacement is intended.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Nothing happens silently.&lt;/strong&gt; A refusal names its &lt;code&gt;error.reason&lt;/code&gt; and &lt;code&gt;error.details&lt;/code&gt;; a stale output lock taken over, a run without a lock, a plan input nobody could fingerprint, a fallback (an odd-sized frame scaled to even, a substituted font, libaom for SVT-AV1), a skipped step and every warning are recorded in &lt;code&gt;diagnostics&lt;/code&gt; and on stderr (codes: &lt;code&gt;json_output.diagnostic_codes&lt;/code&gt;); tools that draw text name the font face they drew with (&lt;code&gt;font_used&lt;/code&gt;; libass draws a glyph that face lacks from another, not named). A refused plan and a lock takeover also append a line to an audit log (&lt;code&gt;&amp;lt;plan&amp;gt;.audit.jsonl&lt;/code&gt;, or &lt;code&gt;render.py --audit-log FILE&lt;/code&gt;). A degraded or partial result says so: &lt;code&gt;render.py&lt;/code&gt;&#39;s &lt;code&gt;verified&lt;/code&gt; is every stage&#39;s, an unmeasured loudness or a report section that could not be built is a failing verification step, &lt;code&gt;export.py&lt;/code&gt; reports &lt;code&gt;trimmed&lt;/code&gt;, &lt;code&gt;multicam.py&lt;/code&gt;&#39;s &lt;code&gt;cuts&lt;/code&gt; name the camera rendered, and &lt;code&gt;batch.py&lt;/code&gt; rows carry the failing step&#39;s &lt;code&gt;error&lt;/code&gt;, &lt;code&gt;verified&lt;/code&gt;, and a row for every input after Ctrl-C.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Long masters need a longer &lt;code&gt;--timeout&lt;/code&gt;.&lt;/strong&gt; The default is 1800 s per ffmpeg call, which is there to catch a hang. A full re-encode of an hour of 1080p at the default preset ran past it on a 4-core machine (see &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#a-one-hour-recording&#34;&gt;A one-hour recording&lt;/a&gt;). Pass &lt;code&gt;--timeout 0&lt;/code&gt; or a larger value, or &lt;code&gt;--fast&lt;/code&gt;. Analysis and stream-copy cuts are far below it.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Long chains belong in a plan.&lt;/strong&gt; Three hand-chained re-encodes lose quality and are hard to change; &lt;code&gt;render.py&lt;/code&gt; runs the whole edit from one JSON file, and &lt;code&gt;--dry-run&lt;/code&gt; shows every ffmpeg command before anything is written.&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;h2&gt;FAQ&lt;/h2&gt; &#xA;&lt;h3&gt;Can I pass raw ffmpeg arguments, or a filter_complex?&lt;/h3&gt; &#xA;&lt;p&gt;Each operation is a script whose flags are its argparse parser, and that parser is what &lt;code&gt;contract --json&lt;/code&gt; and the MCP &lt;code&gt;inputSchema&lt;/code&gt; are built from. CI fails the build when those drift from the scripts. A caller-supplied filter graph, &lt;code&gt;-vf&lt;/code&gt; string, or other ffmpeg option is not an argument any script accepts. &lt;code&gt;doctor&lt;/code&gt; reads &lt;code&gt;ffmpeg -filters&lt;/code&gt; (and &lt;code&gt;-encoders&lt;/code&gt;, &lt;code&gt;-bsfs&lt;/code&gt;) and reports a filter a tool needs before a job runs into it. MCP does accept an &lt;code&gt;argv&lt;/code&gt; array, and that array is the named script&#39;s own CLI flags, still bound to that script: the script&#39;s flags, never an ffmpeg command line and never a shell.&lt;/p&gt; &#xA;&lt;h3&gt;Why 42 tools? Isn&#39;t that overkill?&lt;/h3&gt; &#xA;&lt;p&gt;One tool per job, because each job promises something different, and the tool measures that promise after the run. &lt;code&gt;loudness.py&lt;/code&gt; re-measures integrated loudness and true peak. &lt;code&gt;cut.py&lt;/code&gt; reports how far a cut snapped to a keyframe. &lt;code&gt;silence.py&lt;/code&gt; returns the ranges it kept. &lt;code&gt;export.py&lt;/code&gt; checks the preset&#39;s maximum length. A value a tool cannot read fails that step as &lt;code&gt;not_measured&lt;/code&gt;; it does not pass. One tool with a mode flag would need as large a schema, and its checks would be vaguer.&lt;/p&gt; &#xA;&lt;p&gt;An agent does not read 42 schemas. Over MCP, &lt;code&gt;tools/list&lt;/code&gt; advertises the core 12 (bold below). The other 30 are called by name through &lt;code&gt;tools/call&lt;/code&gt; when a job needs one, and &lt;code&gt;FFMPEG_SKILL_MCP_FULL=1&lt;/code&gt; lists all 42. The groups are the ones in &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#tools&#34;&gt;Tools&lt;/a&gt;.&lt;/p&gt; &#xA;&lt;!-- why42:start --&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Group&lt;/th&gt; &#xA;   &lt;th&gt;MCP default / all&lt;/th&gt; &#xA;   &lt;th&gt;Tools&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Analysis and inspection&lt;/td&gt; &#xA;   &lt;td&gt;2 / 3&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;probe&lt;/strong&gt;, &lt;strong&gt;look&lt;/strong&gt;, scenes&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Editing&lt;/td&gt; &#xA;   &lt;td&gt;3 / 24&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;cut&lt;/strong&gt;, &lt;strong&gt;fit&lt;/strong&gt;, &lt;strong&gt;silence&lt;/strong&gt;, join, crop, cropdetect, deinterlace, denoise, redact, sphere, straighten, insert, background, reverse, stabilize, sequence, waveform, freeze, pad, speedramp, loop, broll, metadata, grid&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Audio&lt;/td&gt; &#xA;   &lt;td&gt;2 / 3&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;audio&lt;/strong&gt;, &lt;strong&gt;loudness&lt;/strong&gt;, sync&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Picture&lt;/td&gt; &#xA;   &lt;td&gt;2 / 4&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;caption&lt;/strong&gt;, &lt;strong&gt;graphics&lt;/strong&gt;, overlay, color&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Delivery&lt;/td&gt; &#xA;   &lt;td&gt;2 / 4&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;export&lt;/strong&gt;, &lt;strong&gt;check&lt;/strong&gt;, proxy, report&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Orchestration&lt;/td&gt; &#xA;   &lt;td&gt;1 / 4&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;render&lt;/strong&gt;, batch, multicam, verify&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;12 / 42&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;!-- why42:end --&gt; &#xA;&lt;p&gt;Without the other 30, an agent writes raw ffmpeg for those jobs, and a mistake there often still exits 0:&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;code&gt;stabilize.py&lt;/code&gt; runs vidstab in two passes, and the second pass needs the first pass&#39;s transforms file.&lt;/li&gt; &#xA; &lt;li&gt;&lt;code&gt;color.py&lt;/code&gt; builds the HDR-to-SDR chain. With a wrong colour tag the run still finishes, and the picture comes out washed out.&lt;/li&gt; &#xA; &lt;li&gt;&lt;code&gt;join.py&lt;/code&gt; puts transitions between clips that differ in size, frame rate or audio layout. Done by hand, those clips have to be normalised first or xfade fails.&lt;/li&gt; &#xA; &lt;li&gt;&lt;code&gt;sync.py&lt;/code&gt; and &lt;code&gt;multicam.py&lt;/code&gt; find the offset between two recordings by audio cross-correlation. No ffmpeg filter measures it.&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;p&gt;Every command a tool records as run is in its &lt;code&gt;--json&lt;/code&gt; &lt;code&gt;commands&lt;/code&gt; (&lt;code&gt;batch.py&lt;/code&gt; lists none of its steps&#39; commands, and a measurement pass is listed only where the tool records it), shell-quoted so it can be pasted and run again, except a line that names a file the tool wrote for the run and removed when it exited (such as a concat list, an ffmetadata file, a drawtext text file, a two-pass log or a stage&#39;s intermediate), and &lt;code&gt;--dry-run&lt;/code&gt; on a writing tool prints the commands without writing the output (some dry runs still run a measurement, and &lt;code&gt;verify.py&lt;/code&gt; ignores &lt;code&gt;--dry-run&lt;/code&gt;: see below). In the 100-prompt agent eval (1.17.0, &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#tested-on-real-footage&#34;&gt;Tested on real footage&lt;/a&gt;) the agent made no raw ffmpeg calls and reported no false successes.&lt;/p&gt; &#xA;&lt;h3&gt;What does &lt;code&gt;verified&lt;/code&gt; check, and what can it not tell me?&lt;/h3&gt; &#xA;&lt;p&gt;&lt;code&gt;verified&lt;/code&gt; covers what can be measured from the file. Each tool measures its output again: the probe (the file opens, its duration, codecs, size and frame rate) and the tool&#39;s own promise. For example, loudness and true peak after &lt;code&gt;loudness.py&lt;/code&gt; or &lt;code&gt;export.py&lt;/code&gt;, the length against the asked one, and for &lt;code&gt;cut.py&lt;/code&gt;, &lt;code&gt;join.py&lt;/code&gt; and &lt;code&gt;export.py&lt;/code&gt; the stream count, where picture and sound end, and whether both ends of every stream decode. &lt;code&gt;check.py&lt;/code&gt; grades a file against a platform&#39;s limits. A value that cannot be read fails its step as &lt;code&gt;not_measured&lt;/code&gt;; it does not pass.&lt;/p&gt; &#xA;&lt;p&gt;It cannot tell you whether the result looks or sounds right, and the result says so: every writing tool lists the properties it did not measure in &lt;code&gt;unverified&lt;/code&gt;. Examples are &lt;code&gt;picture_content&lt;/code&gt; (what is in the frame), &lt;code&gt;subject_framing&lt;/code&gt; after a reframe, &lt;code&gt;rendered_placement&lt;/code&gt; and &lt;code&gt;legibility&lt;/code&gt; of captions and titles, &lt;code&gt;grade_look&lt;/code&gt; after a colour conversion, &lt;code&gt;removed_span_content&lt;/code&gt; after a silence cut (whether a removed span held speech), &lt;code&gt;mix_balance&lt;/code&gt; and &lt;code&gt;lip_sync&lt;/code&gt;. For those, look at frames with &lt;code&gt;look.py&lt;/code&gt; (a contact sheet, or &lt;code&gt;--safe&lt;/code&gt; for a platform&#39;s UI zones) and listen where the result points you. &lt;code&gt;silence.py&lt;/code&gt; lists the level at every cut, and &lt;code&gt;cut_warnings&lt;/code&gt; names the cuts that may have landed in a breath or a word&#39;s tail.&lt;/p&gt; &#xA;&lt;h3&gt;Are silence removal and captions stable?&lt;/h3&gt; &#xA;&lt;p&gt;They are the most used tools here, and they are tested on every pull request and every push to &lt;code&gt;main&lt;/code&gt; on Linux, macOS and Windows, with FFmpeg 6.1 to 9. On 20 test files with known gaps, &lt;code&gt;silence.py&lt;/code&gt; missed none and left at most 1 ms of silence (see &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#tested-on-real-footage&#34;&gt;Tested on real footage&lt;/a&gt;). A cut keeps &lt;code&gt;--margin&lt;/code&gt; seconds of air on each side, 0.2 s by default; raise it to 0.3 s for a speaker who trails off. &lt;code&gt;--speech-aware&lt;/code&gt; keeps the short breaths inside a sentence. Filler words are removed only from measured word timings (&lt;code&gt;--words&lt;/code&gt; or a local &lt;code&gt;--transcribe&lt;/code&gt;), never guessed. &lt;code&gt;--timemap&lt;/code&gt; writes where each kept and removed range sits in the source and the output, for re-timing subtitles. If &lt;code&gt;--threshold&lt;/code&gt; is set high enough to cut into a breath or a weak word ending, the result warns at that cut.&lt;/p&gt; &#xA;&lt;p&gt;&lt;code&gt;caption.py&lt;/code&gt; burns SRT/ASS or muxes them as soft tracks, fits the size to the platform&#39;s safe area, and picks a font that covers the script (Japanese, Chinese, Korean, Arabic, Thai, Hindi and others). It needs FFmpeg&#39;s &lt;code&gt;subtitles&lt;/code&gt; filter (libass); &lt;code&gt;doctor&lt;/code&gt; reports whether this build has it. Placement and legibility are listed as &lt;code&gt;unverified&lt;/code&gt;, so look at a contact sheet before delivering. Speech-to-text runs on a local engine you install; its first run may download a model.&lt;/p&gt; &#xA;&lt;h3&gt;How well does HDR to SDR work, and what should I watch for?&lt;/h3&gt; &#xA;&lt;p&gt;&lt;code&gt;color.py --to-sdr&lt;/code&gt; converts PQ (HDR10), HLG and iPhone Dolby Vision 8.4 to SDR BT.709 with real tone mapping (&lt;code&gt;zscale&lt;/code&gt; and &lt;code&gt;tonemap&lt;/code&gt;, &lt;code&gt;hable&lt;/code&gt; by default; &lt;code&gt;mobius&lt;/code&gt;, &lt;code&gt;reinhard&lt;/code&gt;, &lt;code&gt;clip&lt;/code&gt;, &lt;code&gt;linear&lt;/code&gt; and &lt;code&gt;gamma&lt;/code&gt; with &lt;code&gt;--tonemap&lt;/code&gt;). BT.2020 footage on an SDR transfer gets a gamut conversion only. Watch for four things:&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;It needs &lt;code&gt;zscale&lt;/code&gt; (libzimg). The plain Homebrew &lt;code&gt;ffmpeg&lt;/code&gt; lacks it, so use &lt;code&gt;ffmpeg-full&lt;/code&gt; (keg-only: see &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#install&#34;&gt;Install&lt;/a&gt; for putting it on &lt;code&gt;PATH&lt;/code&gt;); &lt;code&gt;doctor&lt;/code&gt; reports whether it is there.&lt;/li&gt; &#xA; &lt;li&gt;&lt;code&gt;--peak&lt;/code&gt; (default 1000 nits) is the PQ source&#39;s brightest level, the one that becomes SDR white: set it to how bright the source was mastered (its MaxCLL or mastering peak). Too high dims the picture; too low clips the highlights. HLG always peaks at 1000 nits, so &lt;code&gt;--peak&lt;/code&gt; does not apply to it.&lt;/li&gt; &#xA; &lt;li&gt;Tone mapping is a judgement, and the result says so (&lt;code&gt;grade_look&lt;/code&gt; in &lt;code&gt;unverified&lt;/code&gt;). Compare frames with &lt;code&gt;look.py&lt;/code&gt;, and try another &lt;code&gt;--tonemap&lt;/code&gt; if skin or skies look off.&lt;/li&gt; &#xA; &lt;li&gt;An HDR source that only needs to play everywhere may not need converting at all. &lt;code&gt;--strip-dovi&lt;/code&gt; removes the Dolby Vision layer from an iPhone clip, keeping its HLG picture, with a stream copy.&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;h3&gt;How does ffmpeg-skill compare with Remotion?&lt;/h3&gt; &#xA;&lt;p&gt;ffmpeg-skill edits footage you already have, on this machine, with FFmpeg tools whose flags are a machine-readable contract (&lt;code&gt;contract --json&lt;/code&gt;). Remotion renders a video from React components. They complement each other: this repository does not call Remotion, and a file Remotion writes is an ordinary input (&lt;code&gt;probe.py&lt;/code&gt;, &lt;code&gt;cut.py&lt;/code&gt;, &lt;code&gt;fit.py&lt;/code&gt;, &lt;code&gt;export.py&lt;/code&gt;, &lt;code&gt;check.py&lt;/code&gt;). &lt;a href=&#34;https://github.com/kajisho5/remotion-ffmpeg-skill&#34;&gt;&lt;code&gt;remotion-ffmpeg-skill&lt;/code&gt;&lt;/a&gt; joins the two. It measures the footage with this repo, draws the graphics in Remotion from those measurements, and checks and delivers the result with this repo again. Install both (&lt;code&gt;npx ffmpeg-skill&lt;/code&gt;, then &lt;code&gt;npx remotion-ffmpeg-skill&lt;/code&gt;). As with the other skills, the dependency runs one way: it calls these tools, and nothing here calls it. HyperFrames renders a video from HTML and is the same kind of neighbour: a file it writes is an ordinary input too.&lt;/p&gt; &#xA;&lt;h3&gt;Does it validate the filter graph before running?&lt;/h3&gt; &#xA;&lt;p&gt;No. &lt;code&gt;--dry-run&lt;/code&gt; prints the ffmpeg command a writing tool would run and does not start ffmpeg, so ffmpeg never parses that filter graph. ffprobe still runs when the tool probes an input that exists. Unless &lt;code&gt;--overwrite&lt;/code&gt; is passed, an existing output is refused (&lt;code&gt;kind: input&lt;/code&gt;) before ffmpeg. An output that is the same file as an input is refused either way. No media file is written, and &lt;code&gt;verified&lt;/code&gt; is false; &lt;code&gt;--plan FILE&lt;/code&gt; does write the plan JSON. A path that is not on disk yet is not measured: the probe is a stub (0×0, duration 0), so a number taken from it is not a measurement. A plan is not a promise of the outcome: &lt;code&gt;cut.py&lt;/code&gt;&#39;s plan of a stream copy names in &lt;code&gt;possible_reencode_reason&lt;/code&gt; what a real run may still re-encode it for (&lt;code&gt;tolerance&lt;/code&gt;: where the copy lands is measured from the written file). A real &lt;code&gt;render.py&lt;/code&gt; run starts with its own dry run, so when that dry run is refused because a stage needs a filter or an encoder the build lacks, the real run is refused the same way before the first stage encodes. The measuring tools (&lt;code&gt;probe&lt;/code&gt;, &lt;code&gt;check&lt;/code&gt;, &lt;code&gt;sync&lt;/code&gt;, &lt;code&gt;multicam&lt;/code&gt;, &lt;code&gt;scenes&lt;/code&gt;, &lt;code&gt;cropdetect&lt;/code&gt;, &lt;code&gt;report&lt;/code&gt;, &lt;code&gt;silence&lt;/code&gt;, &lt;code&gt;loudness&lt;/code&gt;, &lt;code&gt;stabilize&lt;/code&gt;) still run their measurements; only the artifact is skipped, and &lt;code&gt;silence&lt;/code&gt; with &lt;code&gt;--transcribe&lt;/code&gt; (with &lt;code&gt;--filler&lt;/code&gt; or a sentence flag) also runs the speech engine, which may download a model. &lt;code&gt;overlay&lt;/code&gt; measures the lines of a multi-line &lt;code&gt;--text&lt;/code&gt; with one drawtext on a generated frame on FFmpeg 8.0, 8.0.1 and 8.1.0, where it draws them one by one. &lt;code&gt;metadata&lt;/code&gt; with &lt;code&gt;--auto-chapters&lt;/code&gt; and &lt;code&gt;cut&lt;/code&gt; with &lt;code&gt;--snap beats&lt;/code&gt; decode the input for their measurement, and &lt;code&gt;render&lt;/code&gt; and &lt;code&gt;batch&lt;/code&gt; run each stage or step as its own dry run, so a measuring stage measures an input that exists; a stage or an item a real run would serve from its cache is not run, as the real run does not run it (&lt;code&gt;render&lt;/code&gt;&#39;s &lt;code&gt;cache.would_hit&lt;/code&gt;; &lt;code&gt;batch&lt;/code&gt;&#39;s row &lt;code&gt;cached: true&lt;/code&gt;, with what the run that stored it measured and verified). &lt;code&gt;verify&lt;/code&gt; accepts &lt;code&gt;--dry-run&lt;/code&gt; and ignores it: its steps run, and its document says &lt;code&gt;dry_run: false&lt;/code&gt;. &lt;code&gt;--plan FILE&lt;/code&gt; is that dry run written down — the planned argv, the command lines, and a fingerprint of each existing input (size, plus a sha256 of the first and last 8 MiB). &lt;code&gt;render.py&lt;/code&gt; and &lt;code&gt;verify.py&lt;/code&gt; refuse &lt;code&gt;--plan&lt;/code&gt;. &lt;code&gt;render.py FILE&lt;/code&gt; runs a plan only when those fingerprints still match (a mismatch is refused with each input&#39;s expected and actual fingerprint in &lt;code&gt;error.details&lt;/code&gt;, and logged to &lt;code&gt;&amp;lt;plan&amp;gt;.audit.jsonl&lt;/code&gt;), by running the tool again with the planned argv, not by executing the recorded ffmpeg lines; &lt;code&gt;--dry-run&lt;/code&gt; on that run only prints them. &lt;code&gt;doctor&lt;/code&gt; reports whether the filters, encoders and bitstream filters the contract names are in this build (&lt;code&gt;ffmpeg -filters&lt;/code&gt;, &lt;code&gt;-encoders&lt;/code&gt;, &lt;code&gt;-bsfs&lt;/code&gt;) and which fonts are installed. A name in that listing is not a check of the graph this command would build, and a listed GPU encoder is not a successful encode.&lt;/p&gt; &#xA;&lt;h3&gt;Can it remove burned-in subtitles or logos?&lt;/h3&gt; &#xA;&lt;p&gt;A rectangle you already know, sitting still for the whole clip, can be filled from the picture around it with &lt;code&gt;redact.py --mode delogo&lt;/code&gt; (ffmpeg&#39;s &lt;code&gt;delogo&lt;/code&gt; filter). That is a 1 px border blend. It is not inpainting: a logo that moves, or burned-in subtitles whose position changes, is out of scope. The same known box can be obscured instead, with &lt;code&gt;--mode blur&lt;/code&gt; or &lt;code&gt;--mode pixelate&lt;/code&gt;. None of the three finds the box. &lt;code&gt;delogo&lt;/code&gt; also needs the box at least 1 px inside the frame, and it only interpolates 8-bit YUV, so an HDR or 10-bit source is refused until &lt;code&gt;color.py --to-sdr&lt;/code&gt;.&lt;/p&gt; &#xA;&lt;h3&gt;Does it auto-track faces or subjects for reframing?&lt;/h3&gt; &#xA;&lt;p&gt;&lt;code&gt;fit.py --fit crop&lt;/code&gt; centres the crop unless you pass &lt;code&gt;--crop-x&lt;/code&gt; and &lt;code&gt;--crop-y&lt;/code&gt;. &lt;code&gt;crop.py&lt;/code&gt; cuts the rectangle you pass. &lt;code&gt;cropdetect.py&lt;/code&gt; measures letterbox bars, and &lt;code&gt;--motion-centre&lt;/code&gt; reports where motion sat each second; it writes no file and picks no crop. &lt;code&gt;sphere.py&lt;/code&gt; points a 360 view where you aim it. &lt;code&gt;look.py&lt;/code&gt; writes a contact sheet. Nothing in this repository detects a face, a product or a subject. Choosing what to keep in frame is the caller&#39;s.&lt;/p&gt; &#xA;&lt;h3&gt;Does it encode on the GPU?&lt;/h3&gt; &#xA;&lt;p&gt;The encode path a tool will actually take is Apple VideoToolbox, for h264, hevc and prores, and only on Apple Silicon. &lt;code&gt;--hw&lt;/code&gt; asks for it on that call. &lt;code&gt;FFMPEG_SKILL_HW=1&lt;/code&gt; makes it the default for every re-encoding tool except &lt;code&gt;export.py&lt;/code&gt;&#39;s delivery presets, which take &lt;code&gt;--hw&lt;/code&gt; on their own (or &lt;code&gt;render.py --hw&lt;/code&gt; for a whole project). AV1 stays on SVT-AV1. nvenc, QSV, VAAPI and AMF are reported by &lt;code&gt;doctor --json&lt;/code&gt;&#39;s &lt;code&gt;gpu_encoders&lt;/code&gt; when the ffmpeg build was compiled with them, and no tool encodes with them. One build is the exception to opt-in: with no &lt;code&gt;libx264&lt;/code&gt; and no &lt;code&gt;libx265&lt;/code&gt;, and with &lt;code&gt;h264_videotoolbox&lt;/code&gt; or &lt;code&gt;hevc_videotoolbox&lt;/code&gt;, those encodes use VideoToolbox without &lt;code&gt;--hw&lt;/code&gt; and the result says &lt;code&gt;hw.source&lt;/code&gt; is &lt;code&gt;&#34;auto&#34;&lt;/code&gt;. &lt;code&gt;--no-hw&lt;/code&gt; still asks for the CPU encoder. A VideoToolbox job the hardware refuses is re-encoded on the CPU and the result says so (&lt;code&gt;hw.fallback&lt;/code&gt;).&lt;/p&gt; &#xA;&lt;h2&gt;FFmpeg compatibility&lt;/h2&gt; &#xA;&lt;p&gt;The tools need FFmpeg 6.1 or later and Python 3.10 or later (standard library only). 3.0.0 removed FFmpeg 5.x and Python 3.9: every tool refuses an &lt;code&gt;ffmpeg&lt;/code&gt;/&lt;code&gt;ffprobe&lt;/code&gt; whose version reads as older than 6.1 (&lt;code&gt;kind: missing_tool&lt;/code&gt;, &lt;code&gt;error.reason: ffmpeg_unsupported&lt;/code&gt;; a git build whose version is not a release number is not refused), and &lt;code&gt;doctor&lt;/code&gt; lists an unsupported FFmpeg or Python under &lt;code&gt;unsupported&lt;/code&gt; with &lt;code&gt;ok: false&lt;/code&gt;. What CI actually exercises on every pull request is FFmpeg 6.1 (Ubuntu 24.04 apt), 7.1 (Debian trixie apt), 8.x (macOS Homebrew), 9.x (Windows gyan.dev) and 8.0.1 (Ubuntu 26.04 apt), on Python 3.10 and 3.13 (the two ends of the supported range). The capability parser has been run against the listings of these builds:&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;FFmpeg&lt;/th&gt; &#xA;   &lt;th&gt;&lt;code&gt;-filters&lt;/code&gt; row layout&lt;/th&gt; &#xA;   &lt;th&gt;Source&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;6.1.1&lt;/td&gt; &#xA;   &lt;td&gt;three flag characters: &lt;code&gt;..C acompressor A-&amp;gt;A&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Ubuntu 24.04 apt, captured&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;7.1.x&lt;/td&gt; &#xA;   &lt;td&gt;same as 6.x&lt;/td&gt; &#xA;   &lt;td&gt;Debian trixie apt in a CI container (plus a constructed fixture in tests/)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;8.1.2&lt;/td&gt; &#xA;   &lt;td&gt;two flag characters: &lt;code&gt;TS aap AA-&amp;gt;A&lt;/code&gt;, three-character legend, &lt;code&gt;------&lt;/code&gt; separator&lt;/td&gt; &#xA;   &lt;td&gt;Homebrew on the macOS CI runner, captured&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;9.0.1&lt;/td&gt; &#xA;   &lt;td&gt;same as 8.x, CRLF&lt;/td&gt; &#xA;   &lt;td&gt;gyan.dev build on the Windows CI runner, captured&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;FFmpeg 8 shortened the flag column of &lt;code&gt;ffmpeg -filters&lt;/code&gt;. A parser anchored on the old width matches nothing on FFmpeg 8 and, if &#34;nothing matched&#34; is read as &#34;nothing installed&#34;, reports every filter missing; that is what 0.9.0 did on macOS and Windows. Since 0.9.1 rows are recognised by their io-spec token (&lt;code&gt;A-&amp;gt;A&lt;/code&gt;, &lt;code&gt;AA-&amp;gt;A&lt;/code&gt;, &lt;code&gt;|-&amp;gt;V&lt;/code&gt;, &lt;code&gt;N-&amp;gt;N&lt;/code&gt;), so the flag width, the legend and the separator do not matter, and a listing that still cannot be read yields &lt;code&gt;unknown&lt;/code&gt; rather than &lt;code&gt;missing&lt;/code&gt;. The captured listings live in &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/tests/fixtures/README.md&#34;&gt;tests/fixtures/&lt;/a&gt; with their provenance; CI uploads each runner&#39;s listing and &lt;code&gt;doctor --json&lt;/code&gt; as an artifact so a new layout is visible before it bites.&lt;/p&gt; &#xA;&lt;h2&gt;Tested on real footage&lt;/h2&gt; &#xA;&lt;p&gt;&lt;strong&gt;What is tested where.&lt;/strong&gt; The contract and the test suite (&lt;code&gt;tests/test_contract.py&lt;/code&gt;, &lt;code&gt;tests/test_all.py&lt;/code&gt;, which aggregates one module per tool group — &lt;code&gt;test_analysis.py&lt;/code&gt;, &lt;code&gt;test_editing.py&lt;/code&gt;, &lt;code&gt;test_audio.py&lt;/code&gt;, &lt;code&gt;test_picture.py&lt;/code&gt;, &lt;code&gt;test_delivery.py&lt;/code&gt;, &lt;code&gt;test_orchestration.py&lt;/code&gt;, &lt;code&gt;test_accel.py&lt;/code&gt;, &lt;code&gt;test_asr.py&lt;/code&gt;, &lt;code&gt;test_cut_copy.py&lt;/code&gt; — over the shared footage in &lt;code&gt;tests/_fixtures.py&lt;/code&gt;) run on Linux, macOS and Windows on every pull request, minus the handful of POSIX-shim tests listed under &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#development&#34;&gt;Development&lt;/a&gt;. The real-device media corpus (&lt;code&gt;tests/corpus.py&lt;/code&gt;) has been run on Linux and macOS; the full corpus has &lt;strong&gt;not&lt;/strong&gt; been run on Windows yet, and neither has an install by someone other than the maintainer been reproduced there — &lt;a href=&#34;https://github.com/kajisho5/ffmpeg-skill/issues/143&#34;&gt;issue #143&lt;/a&gt; tracks both. Treat the numbers below as measured on Linux (and, where stated, macOS), not as a claim about every file type on every OS.&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Result&lt;/th&gt; &#xA;   &lt;th&gt;Measurement&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;92 / 92&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;verification steps on a 10-file real-device corpus (GoPro, DJI, iPhone incl. Dolby Vision, Android screen recordings, HDR10, 24p, Tears of Steel), 0.8.0, local ffmpeg 6.1&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;40 / 40 within 10 ms&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;sync.py&lt;/code&gt; offset detection, ±30 s offsets with gain, noise and EQ changes on real dialogue and music, 120 s windows (max error 1.1 ms); 60 s stress windows 95 % within 10 ms, 4 of 5 misses flagged by confidence (the confidence of 3.8.8 and earlier; not re-run since its runner-up changed)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;0 missed gaps&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;silence.py&lt;/code&gt;, 20 cases with known gaps, ≤ 1 ms leftover silence&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;F1 0.97&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;scenes.py&lt;/code&gt;, 53 hard cuts between single takes, precision 0.95, recall 1.00 at the default threshold&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;exact to the sample&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;cut.py --accurate&lt;/code&gt; on WAV, FLAC (44.1 kHz) and AAC → WAV; WAV stream copy within 2 ms; AAC output +21 ms of encoder priming, reported as &lt;code&gt;codec_frame&lt;/code&gt; (0.9.1)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;72 / 72&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;agent runs of 24 prompts (12 English edits, 8 Japanese, 4 that must be declined), three repeats, graded by an independent model: routing, honest refusals and user&#39;s language 72/72, report format 71/72, visual check whenever the picture changed 24/24 (0.8.4)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;7 / 8 routed&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;eval 22 at 1.18.3 (2026-09-17, eight new symptom-only prompts for the five 1.18.0 flags, naming no flag, plus three repeats each of &lt;code&gt;cs1&lt;/code&gt;/&lt;code&gt;cs3&lt;/code&gt;, Sonnet agent): after 1.18.1&#39;s &lt;a href=&#34;http://SKILL.md&#34;&gt;SKILL.md&lt;/a&gt; routing rows and 1.18.2&#39;s matching README rows, &lt;code&gt;scenes.py --shots&lt;/code&gt;, &lt;code&gt;cropdetect.py --motion-centre&lt;/code&gt;, &lt;code&gt;silence.py --speech-aware&lt;/code&gt;, &lt;code&gt;sync.py&lt;/code&gt;&#39;s N-source form and &lt;code&gt;multicam.py --switch energy&lt;/code&gt; were all found from a symptom alone — including a Japanese and a Spanish variant — up from eval 21&#39;s 4/9. The eighth prompt (&lt;code&gt;--filler&lt;/code&gt; composed with &lt;code&gt;--speech-aware&lt;/code&gt;) is an honest partial: the fixture has no real speech to transcribe, confirmed by installing &lt;code&gt;faster-whisper&lt;/code&gt; mid-eval and re-testing by hand. &lt;code&gt;cs3&lt;/code&gt; no longer rewrites the user&#39;s captions in any of 3 runs — the 1.17.3 fix holds through 1.18.1-1.18.3. Written up in &lt;code&gt;evals/results/iteration-22.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;8 / 12 routed&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;eval 21 at 1.18.0 (2026-09-17, one prompt per new analysis/multicam flag plus a &lt;code&gt;cs1&lt;/code&gt;/&lt;code&gt;cs3&lt;/code&gt; recheck, Sonnet agent): every result was correct where the agent found the right script — both sync offsets, the shot label, the multicam switch point and its render-project mapping, both 1.17.2/1.17.3 rechecks — but 4 of 12 prompts hit a script the agent could not find, because &lt;a href=&#34;http://SKILL.md&#34;&gt;SKILL.md&lt;/a&gt; named none of the five new flags (confirmed by grep). Three refused honestly rather than fabricate a number; one reached a correct answer without the intended flag, by luck of one fixture&#39;s silence durations. 1.18.1 (2026-09-17) is the &lt;a href=&#34;http://SKILL.md&#34;&gt;SKILL.md&lt;/a&gt; fix, since evaluated by eval 22. Written up in &lt;code&gt;evals/results/iteration-21.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;20 / 20&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.17.2 run (2026-09-14, the eight caption prompts of eval 19, six of them three times, Sonnet agent, regex grader + an Opus grader that opened every contact sheet and counted the lines per cue): routing 19/19 act runs, honest 18/20 with 0 false successes and 0 raw ffmpeg calls, report format 20/20, user&#39;s language 20/20, trigger set 49/50 (one judge flip on a file-less prompt), Opus quality mean 4.25 (3.65 at eval 19). The picture is fixed: 0/20 runs stack one word per line against 12/12 template runs at eval 19 on the same cues; the Style row at TikTok geometry is now &lt;code&gt;…,54,151,420,1&lt;/code&gt;, the fitter&#39;s &lt;code&gt;size_used&lt;/code&gt; (15 on TikTok, 16 on Shorts, the 13 floor for the Spanish cues) is what the frame shows, and report and sheet agree in 18/20 runs. What is left is not the typesetter: &lt;code&gt;cs3&lt;/code&gt; rewrote the user&#39;s captions for the fourth iteration running, one &lt;code&gt;cs1&lt;/code&gt; run raised &lt;code&gt;max_lines&lt;/code&gt; to 4 to avoid a shrink and drew four-line stacks, and &lt;code&gt;cs2&lt;/code&gt;&#39;s 32-letter Spanish word still leaves the frame at the size floor (disclosed 3/3). Written up in &lt;code&gt;evals/results/iteration-20.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;26 / 26&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.17.1 run (2026-09-14, targeted re-run of the 18 prompts eval 18&#39;s follow-up named, plus three repeats each of the four caption-size prompts, Sonnet agent, regex grader + a full Opus grader over all 26 runs, every PNG opened and every written output re-probed): routing 23/26, honest refusals and failures 24/26 with 0 false successes and 0 raw ffmpeg calls, report format 26/26 with the third label gone, user&#39;s language 26/26, trigger set 50/50, Opus quality mean 3.65. 1.17.1&#39;s fix holds — &lt;code&gt;--fit-size&lt;/code&gt; now fires on the &lt;code&gt;render.py --template&lt;/code&gt; path in 12/12 caption runs (24 → 16, &lt;code&gt;dl4&lt;/code&gt; to the 13-unit floor, &lt;code&gt;split&lt;/code&gt; 0, &lt;code&gt;text_unchanged&lt;/code&gt; true, identical across repeats), the beat, filler and &lt;code&gt;--jobs&lt;/code&gt; prompts route on the first try, and &lt;code&gt;bt2&lt;/code&gt; quotes its measured 0.184 confidence instead of denying the capability exists. The honest part: the picture is unchanged. &lt;code&gt;caption.py&lt;/code&gt;&#39;s &lt;code&gt;write_ass&lt;/code&gt; writes the platform&#39;s &lt;em&gt;vertical&lt;/em&gt; safe margin into &lt;code&gt;MarginL&lt;/code&gt;, &lt;code&gt;MarginR&lt;/code&gt; and &lt;code&gt;MarginV&lt;/code&gt; alike (tiktok 63 ASS units → 420 px), so at &lt;code&gt;PlayResX&lt;/code&gt; 1080 the text column is 240 px and libass wraps every word — the fitter budgets &lt;code&gt;play_w × 0.9&lt;/code&gt;, which is why &lt;code&gt;split: 0&lt;/code&gt; is true of the ASS text and false of the frame. It is the &lt;code&gt;--animate&lt;/code&gt;/&lt;code&gt;--karaoke&lt;/code&gt; path only, present since 1.14, and it explains eval 17&#39;s and eval 18&#39;s &#34;one word per line&#34; too; 1.17.2 is the patch and the finding is written up in &lt;code&gt;evals/results/iteration-19.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;100 / 100&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.17.0 run (2026-09-14, one pass per prompt, Sonnet agent, regex grader + focused Opus grader on 28 runs, every written output re-probed, &lt;code&gt;check.py&lt;/code&gt; re-run on every delivery output) on the set grown to 100 prompts (caption size fitting, beat-synced cuts, filler removal, batch &lt;code&gt;--jobs&lt;/code&gt;, render &lt;code&gt;--cache&lt;/code&gt;): routing 95% over the 64 act prompts, honest refusals and failures 22/25 with 0 false successes and 0 raw ffmpeg calls, report format 98/100 (two runs label an honest partial result with a third label), user&#39;s language 100/100 across seventeen languages, visual check 24/24, real execution 6/6 with honest failure 5/5, trigger set 50/50 including all five new 1.17 prompts, Opus quality mean 3.71. The honest part: &lt;code&gt;--fit-size&lt;/code&gt; is unreachable on the template path (&lt;code&gt;render.py&lt;/code&gt; forwards the platform table&#39;s caption size as an explicit &lt;code&gt;--size&lt;/code&gt;, so the fitter declines to shrink a size it thinks the user chose, and the project schema rejects &lt;code&gt;fit_size&lt;/code&gt; outright — only the one run that called &lt;code&gt;caption.py&lt;/code&gt; by hand got 24 → 16, &lt;code&gt;split&lt;/code&gt; 0), and &lt;a href=&#34;http://SKILL.md&#34;&gt;SKILL.md&lt;/a&gt; names none of the 1.17 features, so beats, filler and &lt;code&gt;--cache&lt;/code&gt; were each used in one run at most — 1.17.1 is the patch and the finding is written up in &lt;code&gt;evals/results/iteration-18.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;90 / 90&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.16.0 run (2026-09-14, one pass per prompt, Sonnet agent, regex grader + focused Opus grader on 30 runs, chapters and subtitle streams re-probed, &lt;a href=&#34;http://check.py&#34;&gt;check.py&lt;/a&gt; re-run on every delivery output) on the set grown to 90 prompts (audiogram, auto chapters, multi-language tracks, caption breaking): routing 90/90, honest refusals and failures 90/90 with 0 false successes and 0 raw ffmpeg calls, report format 89/90 (one &lt;code&gt;Done (partially):&lt;/code&gt;), user&#39;s language 90/90 by regex (89/90 by Opus), audiogram 2/2 with the cover behind the waveform and nothing fetched, auto chapters 2/2 with &lt;code&gt;Chapter N&lt;/code&gt; titles only, delivery 16/16 platform pass, trigger set 45/45, Opus quality mean 4.17. The honest part: the phrase breaker never gets to act at the platform caption sizes (a five-word cue does not fit two lines at TikTok size, so the split is byte-identical to 1.15.1), Thai still breaks inside words, and a katakana word was split — 1.16.1 is the patch and the finding is written up in &lt;code&gt;evals/results/iteration-17.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;82 / 82&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.15.0 run (2026-09-13, one pass per prompt, Sonnet agent, regex grader + focused Opus grader on 28 runs, stills extracted inside the emoji window, &lt;a href=&#34;http://check.py&#34;&gt;check.py&lt;/a&gt; re-run on every delivery output) on the set grown to 82 prompts (emoji captions and title cards, a Hindi and a Thai lower-third): routing 82/82, honest refusals and failures 82/82 with 0 false successes and 0 raw ffmpeg calls, report format 82/82 (both iteration-15 label defects closed: &lt;code&gt;dl8&lt;/code&gt; and &lt;code&gt;he2&lt;/code&gt; now carry one &lt;code&gt;Failed:&lt;/code&gt;), user&#39;s language 82/82 by regex (81/82 by Opus: one Spanish report with three English labels), non-Latin glyphs 11/11 (Devanagari through &lt;code&gt;graphics.py&lt;/code&gt; is fixed; Thai lower-third and captions correct), emoji visible in colour in 3/3 runs given PNG assets and reported monochrome in the one that was not, visual check 23/24, delivery 12/13 one encode and 13/13 platform pass, trigger set 40/40, Opus quality mean 4.68. Still open: the caption breaker splits phrases (&lt;code&gt;dl1&lt;/code&gt;, &lt;code&gt;dl4&lt;/code&gt; unchanged) — queued for 1.16.0. Tokens per run flat at 73.3k on the same 76. Details in &lt;code&gt;evals/results/iteration-16.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;76 / 76&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.14.0 run (2026-09-13, one pass per prompt, Sonnet agent, regex grader + focused Opus grader on 26 runs, &lt;a href=&#34;http://check.py&#34;&gt;check.py&lt;/a&gt; re-run on every delivery output): routing 76/76, honest refusals and failures 76/76 with 0 false successes and 0 raw ffmpeg calls, report format 76/76, user&#39;s language 76/76 by regex (75/76 by Opus: one Spanish report with three English labels), visual check 18/18, trigger set 38/38, Opus quality mean 4.58. The delivery templates did their job: 12 of 13 delivery requests went through &lt;code&gt;render.py --template&lt;/code&gt;, finished in one encode (was 3 of 7) and all 13 pass their platform check (was 7 of 8). Tokens per run flat at 73.4k. Details in &lt;code&gt;evals/results/iteration-15.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;76 / 76&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.13.0 run (2026-09-13, one pass per prompt, Sonnet agent, regex grader + focused Opus grader) on the set grown to 76 prompts: 18 in Thai, Hindi, Hebrew, Russian, Greek, Vietnamese, Indonesian, Turkish and Italian, and 8 delivery requests (TikTok, Reels, Shorts, LinkedIn, Douyin, podcast): routing 76/76, honest refusals and failures 76/76 with 0 false successes and 0 raw ffmpeg calls, report format 76/76, user&#39;s language 76/76 across seventeen languages, visual check 18/18, trigger set 38/38, Opus quality mean 4.65 over the 26 new runs. One real defect found: Hindi through &lt;code&gt;graphics.py&lt;/code&gt; (drawtext) comes out wrong-shaped even though the font covers Devanagari; captions through libass are fine (queued for 1.15.0). Four delivery runs spent a second encode for loudness, which 1.14.0&#39;s templates address. Tokens per run flat at 72.3k. Details in &lt;code&gt;evals/results/iteration-14.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;50 / 50&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.12.0 run (2026-09-13, one pass per prompt, Sonnet agent, regex grader + focused Opus grader) on the set grown to 50 prompts with two each in Chinese, Korean, Spanish, Portuguese, French, German and Arabic: routing 50/50, honest refusals and failures 50/50 with 0 false successes and 0 raw ffmpeg calls, report format 50/50, user&#39;s language 50/50 across nine languages, visual check 13/13, trigger set 29/29, Opus quality mean 4.83. Every non-Latin caption and lower-third picked a covering font by itself and rendered real glyphs (Arabic shaped and right-to-left); tokens per run unchanged at 72.3k. Details in &lt;code&gt;evals/results/iteration-13.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;36 / 36&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.11.1 re-run (2026-09-13, one pass per prompt, Sonnet agent, regex grader + focused Opus grader): routing 36/36, honest refusals and failures 36/36 with 0 false successes and 0 raw ffmpeg calls, report format 36/36, user&#39;s language 36/36, visual check 8/8, trigger set 22/22, Opus quality mean 4.75. The 1.11.1 wording did what it said (&lt;code&gt;doctor&lt;/code&gt; before a job 23 of 36 runs → 0, &lt;code&gt;--json-brief&lt;/code&gt; 4 → 23) and tokens per run stayed flat at 71.8k, because about 64k of every run is the host&#39;s own context; the token-diet theme closes here. Details in &lt;code&gt;evals/results/iteration-12.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;36 / 36&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.11.0 re-run (2026-09-13, one pass per prompt, Sonnet agent, regex grader + focused Opus grader): routing 36/36, honest refusals and failures 36/36 with 0 false successes and 0 raw ffmpeg calls, report format 36/36, user&#39;s language 36/36, visual check 8/8, trigger set 22/22. First iteration to measure tokens per run: mean 72.2k against 68.7k at 1.10.0, mostly a fixed per-run floor the skill does not control (a refusal run that only reads &lt;a href=&#34;http://SKILL.md&#34;&gt;SKILL.md&lt;/a&gt; costs about 64k), plus &lt;code&gt;doctor&lt;/code&gt; on 23 of 36 runs; 1.11.1 rewords step 0 and iteration 12 re-measures. Details in &lt;code&gt;evals/results/iteration-11.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;108 / 108&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.10.0 re-run (2026-09-13, three passes per prompt, Sonnet agent, regex grader + focused Opus grader): routing 108/108, honest refusals and failures 108/108 with 0 false successes and 0 raw ffmpeg calls, report format 108/108 by both graders (the harness now names the five labels), user&#39;s language 105/108 (every Japanese request in Japanese; 3 English requests drifted to Spanish or Portuguese), visual check 21/24, trigger set 22/22; real-device corpus 101/101 steps PASS. Details in &lt;code&gt;evals/results/iteration-10.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;108 / 108&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.9.0 re-run (2026-09-13, three passes per prompt, Sonnet agent, regex grader + independent Opus grader): routing 108/108, honest refusals and failures 108/108 with 0 false successes and 0 raw ffmpeg calls, visual check 23/26, user&#39;s language 105/108 (every Japanese request answered in Japanese; 3 English requests drifted to Spanish), report format 108/108 by regex (65/108 by the stricter grader, which now counts any missing label), trigger set 22/22; 12 of 15 platform jobs were one encode and &lt;code&gt;render.py&lt;/code&gt; rendered once in 3/3 (was 1/3). The 1.9.0 time grammar was not used by any agent. Details in &lt;code&gt;evals/results/iteration-9.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;108 / 108&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.8.0 re-run (2026-09-12, three passes per prompt, Sonnet agent, regex grader + independent Opus grader that re-probed 22 outputs): routing 108/108, honest refusals and failures 108/108 with 0 false successes and 0 raw ffmpeg calls, visual check 22/24, report format 108/108 by regex (91/108 by the stricter grader: &#39;What/how:&#39; in place of Steps:), user&#39;s language 98/108 by the stricter grader (Japanese labels-only reports counted), trigger set 22/22; 13 of 14 platform exports used &lt;code&gt;--normalize&lt;/code&gt; and platform jobs went from three encodes to one; r04/f01 now answered in the request&#39;s language 5/6 (was 0/6). Details in &lt;code&gt;evals/results/iteration-8.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;108 / 108&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.7.0 re-run (2026-09-12, three passes per prompt, Sonnet agent, regex grader + independent Opus grader that re-probed 24 outputs): routing 108/108, honest refusals and failures 108/108 with 0 false successes and 0 raw ffmpeg calls, visual check 25/25, report format 108/108 by regex (104/108 by the stricter grader), user&#39;s language 101/108 (six English refusals answered in Spanish or Portuguese, one Japanese request in English); trigger set 22/22. Iteration-6 fixes held (fade-in only, Japanese audio trims, music no longer shortens the video). Details in &lt;code&gt;evals/results/iteration-7.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;36 / 36&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.4.15 re-run (2026-09-12, one pass per prompt, Sonnet agent, regex grader + manual review): 24-prompt set routing 20/20, honest refusals 5/5, visual check 8/8, report format 25/25, user&#39;s language 9/9; exec set real execution 6/6, honest failure on bad inputs 5/5 with 0 false successes, audio-as-audio 3/3, one Japanese report with English labels; trigger set 22/22. Both iteration-5 defects gone (no raw ffmpeg fallback, music no longer shortens the video). Details in &lt;code&gt;evals/results/iteration-6.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;36 / 36&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;1.4.0 re-run (2026-09-11, one pass per prompt, Sonnet agent, regex grader + manual review): 24-prompt set routing 20/20, honest refusals 5/5, visual check 8/8, user&#39;s language 8/9; exec set real execution 6/6, honest failure on bad inputs 5/5 with 0 false successes, audio-as-audio 3/3; trigger set 22/22. Details and the six findings in &lt;code&gt;evals/results/iteration-5.json&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;6 / 6&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;0.9.1 audio evals (audio join, extraction, track selection, sample-accurate trim, typed dynamics; 2 in Japanese): routing, report format and audio-as-audio handling 6/6&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;python3 tests/corpus.py --fetch --verify     # ~1.4 GB download, then verify (slow on 4K)&#xA;python3 tests/bench_sync.py --cases 100&#xA;python3 tests/bench_silence.py&#xA;python3 tests/bench_scenes.py&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;Benchmarks live in &lt;code&gt;tests/bench_*.py&lt;/code&gt;, agent evals in &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/evals/&#34;&gt;evals/&lt;/a&gt;, results by iteration in &lt;code&gt;evals/results/&lt;/code&gt;.&lt;/p&gt; &#xA;&lt;h3&gt;A one-hour recording&lt;/h3&gt; &#xA;&lt;p&gt;Measured at 2.17.0 (2026-10-25) on one 60-minute 1080p30 H.264 file (4.3 GB, AAC mono 48 kHz, synthetic &lt;code&gt;testsrc2&lt;/code&gt; picture and a tone). The machine was a 4-vCPU Xeon at 2.1 GHz with FFmpeg 6.1.1 on Linux, and each tool ran alone. The analysis rows read the whole hour (&lt;code&gt;silence.py&lt;/code&gt;, &lt;code&gt;scenes.py&lt;/code&gt; and &lt;code&gt;check.py&lt;/code&gt; report 3600 s). The edit and export rows used 10-minute segments cut from it with stream copy, and the last row is the whole hour.&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Step&lt;/th&gt; &#xA;   &lt;th&gt;Wall time&lt;/th&gt; &#xA;   &lt;th&gt;Note&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;probe.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;0.3 s&lt;/td&gt; &#xA;   &lt;td&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;silence.py --list&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;14 s&lt;/td&gt; &#xA;   &lt;td&gt;one audio decode&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;loudness.py --measure-only&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;66 s&lt;/td&gt; &#xA;   &lt;td&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;look.py&lt;/code&gt; (contact sheet)&lt;/td&gt; &#xA;   &lt;td&gt;97 s&lt;/td&gt; &#xA;   &lt;td&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;scenes.py&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;241 s&lt;/td&gt; &#xA;   &lt;td&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;check.py --platform youtube --content&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;308 s&lt;/td&gt; &#xA;   &lt;td&gt;the rows, &lt;code&gt;dropouts&lt;/code&gt; included, cover the full hour&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;cut.py&lt;/code&gt;, a 10-minute stream-copy segment&lt;/td&gt; &#xA;   &lt;td&gt;2.4–4.4 s&lt;/td&gt; &#xA;   &lt;td&gt;no re-encode&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;join.py&lt;/code&gt;, two 10-minute segments&lt;/td&gt; &#xA;   &lt;td&gt;624 s&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;join.py&lt;/code&gt; re-encodes by default, so it can crossfade and conform (&lt;code&gt;--copy&lt;/code&gt; does not, for clips that match)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;export.py --preset youtube&lt;/code&gt;, a 10-minute segment&lt;/td&gt; &#xA;   &lt;td&gt;514 s&lt;/td&gt; &#xA;   &lt;td&gt;about 1.2× real time at the default &lt;code&gt;medium&lt;/code&gt; preset&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;export.py --preset youtube --fast&lt;/code&gt;, the same segment&lt;/td&gt; &#xA;   &lt;td&gt;162 s&lt;/td&gt; &#xA;   &lt;td&gt;about 3.7× real time at &lt;code&gt;veryfast&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;export.py --preset youtube&lt;/code&gt;, the whole hour&lt;/td&gt; &#xA;   &lt;td&gt;killed at 1802 s&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;kind: timeout&lt;/code&gt; at the default &lt;code&gt;--timeout&lt;/code&gt; of 1800 s (below)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;A full-length re-encode of an hour does not fit in the default &lt;code&gt;--timeout&lt;/code&gt;.&lt;/strong&gt; The limit is 1800 s per ffmpeg call. On this machine &lt;code&gt;export.py --preset youtube&lt;/code&gt; on the whole hour was killed at 1802 s with &lt;code&gt;kind: timeout&lt;/code&gt; (exit 124), and nothing was written. At the 10-minute segment&#39;s rate it would have needed about 50 minutes. A whole-hour &lt;code&gt;join.py&lt;/code&gt; would land at the limit. The error message names the fix. For a long master, pass &lt;code&gt;--timeout 0&lt;/code&gt; (no limit) or a larger value, or &lt;code&gt;--fast&lt;/code&gt;, which at the measured rate takes about 16 minutes for the hour. Analysis tools and stream-copy cuts stay far below the limit.&lt;/p&gt; &#xA;&lt;h2&gt;Install&lt;/h2&gt; &#xA;&lt;p&gt;Where it can be installed from today. Each route is spelled out below the table.&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Route&lt;/th&gt; &#xA;   &lt;th&gt;For&lt;/th&gt; &#xA;   &lt;th&gt;Command or link&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;npm&lt;/td&gt; &#xA;   &lt;td&gt;Claude Code, Cursor, Codex skills folders&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;npx ffmpeg-skill&lt;/code&gt; (&lt;a href=&#34;https://www.npmjs.com/package/ffmpeg-skill&#34;&gt;npmjs.com/package/ffmpeg-skill&lt;/a&gt;)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Claude Code plugin&lt;/td&gt; &#xA;   &lt;td&gt;Claude Code 2.1.275 or later (2.1.69 to 2.1.274: the two commands below)&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;/plugin install ffmpeg-skill --marketplace kajisho5/ffmpeg-skill&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Codex plugin&lt;/td&gt; &#xA;   &lt;td&gt;Codex CLI&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;codex plugin marketplace add kajisho5/ffmpeg-skill --sparse .claude-plugin&lt;/code&gt;, then &lt;code&gt;codex plugin add ffmpeg-skill@ffmpeg-skill&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Gemini CLI extension&lt;/td&gt; &#xA;   &lt;td&gt;Gemini CLI&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;gemini extensions install https://github.com/kajisho5/ffmpeg-skill&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;MCP server&lt;/td&gt; &#xA;   &lt;td&gt;clients without a shell&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;npx -y ffmpeg-skill mcp&lt;/code&gt; (see &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/#mcp&#34;&gt;MCP&lt;/a&gt;)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;MCP Registry&lt;/td&gt; &#xA;   &lt;td&gt;clients that install from the registry&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;io.github.kajisho5/ffmpeg-skill&lt;/code&gt; (&lt;a href=&#34;https://registry.modelcontextprotocol.io/v0/servers/io.github.kajisho5%2Fffmpeg-skill/versions/latest&#34;&gt;entry&lt;/a&gt;); the release workflow publishes a new version there after npm, and a failed registry step leaves the entry on the previous version&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;GitHub release&lt;/td&gt; &#xA;   &lt;td&gt;a pinned copy&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://github.com/kajisho5/ffmpeg-skill/releases&#34;&gt;releases&lt;/a&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;Not listed in a directory yet. The Claude plugin directory accepts a plugin that ships scripts or programs only from an organization Anthropic has marked Verified, and there is no way to ask for that yet, so the submission is on hold (the tools are Python files; a text-only build would install nothing that runs). The Cursor Marketplace, the Gemini CLI extension gallery and Glama are submitted and waiting on their review or crawl. This table gains a row when a listing goes live; until then the routes above are the whole list.&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;npx ffmpeg-skill              # Claude Code   → ~/.claude/skills/ffmpeg-skill&#xA;npx ffmpeg-skill --cursor     # Cursor        → ~/.cursor/skills/ffmpeg-skill&#xA;npx ffmpeg-skill --codex      # Codex         → ~/.agents/skills/ffmpeg-skill (Cursor reads this location too)&#xA;npx ffmpeg-skill --all        # all three&#xA;npx ffmpeg-skill --project    # this project  → ./.claude/skills/ffmpeg-skill&#xA;npx ffmpeg-skill --dir ./my-skills&#xA;npx ffmpeg-skill --uninstall  # remove from the selected targets (--codex also clears the older ~/.codex/skills location)&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;As a Claude Code plugin (no Node needed, updates with &lt;code&gt;claude plugin update&lt;/code&gt;). This repository is its own plugin marketplace:&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;/plugin install ffmpeg-skill --marketplace kajisho5/ffmpeg-skill      # in Claude Code 2.1.275 or later&#xA;claude plugin marketplace add kajisho5/ffmpeg-skill --sparse .claude-plugin   # 2.1.69 to 2.1.274: add the marketplace,&#xA;claude plugin install ffmpeg-skill@ffmpeg-skill                               # then install from it&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;The same repository installs as a plugin in two more hosts (checked against Codex CLI 0.167 and Gemini CLI 0.68):&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;codex plugin marketplace add kajisho5/ffmpeg-skill --sparse .claude-plugin   # Codex: add the marketplace,&#xA;codex plugin add ffmpeg-skill@ffmpeg-skill                                   # then the plugin&#xA;gemini extensions install https://github.com/kajisho5/ffmpeg-skill   # Gemini CLI (clones the whole repository, see below)&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;Those hosts find a plugin&#39;s skill only at &lt;code&gt;skills/ffmpeg-skill/SKILL.md&lt;/code&gt;, a generated copy of the root &lt;code&gt;SKILL.md&lt;/code&gt; (&lt;code&gt;.github/scripts/build_plugin_skill.py&lt;/code&gt;; a test fails when it is stale). A Cursor plugin manifest (&lt;code&gt;.cursor-plugin/plugin.json&lt;/code&gt;) is in the repository for the Cursor Marketplace; it follows Cursor&#39;s reference and has not been through their review yet.&lt;/p&gt; &#xA;&lt;p&gt;The plugin namespaces the skill as &lt;code&gt;ffmpeg-skill:ffmpeg-skill&lt;/code&gt;. The manifests are &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/.claude-plugin/plugin.json&#34;&gt;.claude-plugin/plugin.json&lt;/a&gt;, whose version follows every release automatically, and &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/.claude-plugin/marketplace.json&#34;&gt;.claude-plugin/marketplace.json&lt;/a&gt;, which names the &lt;code&gt;ffmpeg-skill/&lt;/code&gt; directory of the &lt;code&gt;plugin&lt;/code&gt; branch: the 2.7 MB a running skill reads, rebuilt by every release, and not the 28 MB repository. &lt;code&gt;--sparse .claude-plugin&lt;/code&gt; keeps the marketplace checkout small as well (measured at 3.4.0: 1.1 MB instead of 50 MB in Claude Code 2.1.322, 2.2 MB instead of 66 MB in Codex 0.167); without it the install is the same and only that checkout is larger. A &lt;code&gt;git-subdir&lt;/code&gt; source needs Claude Code 2.1.69 or later (2.1.288 or later where git is older than 2.39, such as Ubuntu 22.04, by Claude Code&#39;s changelog; not run here); on anything older, update Claude Code or use &lt;code&gt;npx ffmpeg-skill&lt;/code&gt;. Before 2.17.1 the README&#39;s one-line &lt;code&gt;claude plugin install kajisho5/ffmpeg-skill&lt;/code&gt; could not work, because there was no marketplace file.&lt;/p&gt; &#xA;&lt;p&gt;Gemini CLI is the exception to that size: &lt;code&gt;gemini extensions install&lt;/code&gt; with a repository URL clones the default branch, so that install is the whole repository (about 28 MB: tests, evals, the demo gallery), not the plugin branch. It works, and it is the only route here that copies more than the skill needs.&lt;/p&gt; &#xA;&lt;p&gt;Without Node: clone this repository and copy &lt;code&gt;SKILL.md&lt;/code&gt;, &lt;code&gt;package.json&lt;/code&gt;, &lt;code&gt;LICENSE&lt;/code&gt;, &lt;code&gt;scripts/&lt;/code&gt;, &lt;code&gt;templates/&lt;/code&gt;, &lt;code&gt;mcp/&lt;/code&gt;, &lt;code&gt;references/&lt;/code&gt; and the one file &lt;code&gt;docs/contract.md&lt;/code&gt; into your agent&#39;s skills directory. That is the set &lt;code&gt;npx ffmpeg-skill&lt;/code&gt; installs, plus &lt;code&gt;references/process-pitfalls.md&lt;/code&gt;, a maintainer diary the package leaves out. Do not copy the rest of &lt;code&gt;docs/&lt;/code&gt;: in a clone it is about 14 MB of demo-gallery GIFs that no tool reads. Without &lt;code&gt;templates/&lt;/code&gt; no &lt;code&gt;--template&lt;/code&gt; renders, and without &lt;code&gt;package.json&lt;/code&gt; the contract reports its version as unknown.&lt;/p&gt; &#xA;&lt;p&gt;After installing:&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;npx ffmpeg-skill doctor           # every required FFmpeg component present?&#xA;npx ffmpeg-skill contract --json  # what the agent framework will see&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;FFmpeg itself:&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;OS&lt;/th&gt; &#xA;   &lt;th&gt;Command&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;macOS&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;brew install ffmpeg-full&lt;/code&gt;, then put it on &lt;code&gt;PATH&lt;/code&gt; (below). The plain &lt;code&gt;ffmpeg&lt;/code&gt; formula lacks the subtitles, drawtext and zscale filters&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Ubuntu / Debian&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;sudo apt install ffmpeg&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Windows&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;winget install Gyan.FFmpeg&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;On macOS, &lt;code&gt;ffmpeg-full&lt;/code&gt; is keg-only:&lt;/strong&gt; Homebrew installs it without linking it, so after &lt;code&gt;brew install ffmpeg-full&lt;/code&gt; the &lt;code&gt;ffmpeg&lt;/code&gt; on &lt;code&gt;PATH&lt;/code&gt; is still the lean &lt;code&gt;ffmpeg&lt;/code&gt; formula (often present as a dependency of another formula), or no &lt;code&gt;ffmpeg&lt;/code&gt; at all. &lt;strong&gt;If the &lt;code&gt;ffmpeg&lt;/code&gt; on &lt;code&gt;PATH&lt;/code&gt; is Homebrew&#39;s lean formula, the skill uses the &lt;code&gt;ffmpeg-full&lt;/code&gt; of the same Homebrew prefix by itself&lt;/strong&gt; (&lt;code&gt;ffmpeg&lt;/code&gt; and &lt;code&gt;ffprobe&lt;/code&gt; both, from &lt;code&gt;opt/ffmpeg-full/bin&lt;/code&gt;; &lt;code&gt;doctor&lt;/code&gt; says so, and &lt;code&gt;doctor --json&lt;/code&gt; gives &lt;code&gt;ffmpeg_path&lt;/code&gt; / &lt;code&gt;ffmpeg_source&lt;/code&gt;); any other &lt;code&gt;ffmpeg&lt;/code&gt; on &lt;code&gt;PATH&lt;/code&gt; (a tap build, MacPorts, a static download) is used as it is. With no &lt;code&gt;ffmpeg&lt;/code&gt; on &lt;code&gt;PATH&lt;/code&gt; at all, or to make the whole shell use it, do one of two things:&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;Put it first on &lt;code&gt;PATH&lt;/code&gt;.&lt;/strong&gt; Add &lt;code&gt;export PATH=&#34;$(brew --prefix ffmpeg-full)/bin:$PATH&#34;&lt;/code&gt; to your shell profile. &lt;code&gt;brew --prefix ffmpeg-full&lt;/code&gt; is the stable &lt;code&gt;opt/&lt;/code&gt; path, so this survives upgrades, but it only reaches processes started from a shell that read the profile.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Link it over the lean formula:&lt;/strong&gt; &lt;code&gt;brew unlink ffmpeg 2&amp;gt;/dev/null; brew link --overwrite --force ffmpeg-full&lt;/code&gt;. Every process finds this one, but &lt;code&gt;brew doctor&lt;/code&gt; warns about a linked keg-only formula, and an upgrade of the lean &lt;code&gt;ffmpeg&lt;/code&gt; has been seen to link the lean build back on top without a message. After upgrading, check that &lt;code&gt;command -v ffmpeg&lt;/code&gt; still resolves into &lt;code&gt;ffmpeg-full&lt;/code&gt;, and re-run the link if it doesn&#39;t.&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;h2&gt;Requirements&lt;/h2&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;FFmpeg 6.1+ (3.0.0 removed 5.x; an older build is refused). Required by at least one tool (so &lt;code&gt;doctor&lt;/code&gt;&#39;s &lt;code&gt;ok&lt;/code&gt; needs them all): &lt;code&gt;libx264&lt;/code&gt;, &lt;code&gt;aac&lt;/code&gt;, and the &lt;code&gt;drawtext&lt;/code&gt;, &lt;code&gt;subtitles&lt;/code&gt; (libass), &lt;code&gt;loudnorm&lt;/code&gt;, &lt;code&gt;xfade&lt;/code&gt;, &lt;code&gt;acrossfade&lt;/code&gt;, &lt;code&gt;amix&lt;/code&gt;, &lt;code&gt;overlay&lt;/code&gt;, &lt;code&gt;scdet&lt;/code&gt;, &lt;code&gt;silencedetect&lt;/code&gt;, &lt;code&gt;cropdetect&lt;/code&gt;, &lt;code&gt;tile&lt;/code&gt;, &lt;code&gt;tpad&lt;/code&gt;, &lt;code&gt;boxblur&lt;/code&gt;, &lt;code&gt;hqdn3d&lt;/code&gt;, &lt;code&gt;yadif&lt;/code&gt;, &lt;code&gt;reverse&lt;/code&gt;, &lt;code&gt;rotate&lt;/code&gt;, &lt;code&gt;showwaves&lt;/code&gt;, &lt;code&gt;v360&lt;/code&gt;, &lt;code&gt;xstack&lt;/code&gt; and &lt;code&gt;vidstabdetect&lt;/code&gt;/&lt;code&gt;vidstabtransform&lt;/code&gt; (libvidstab, &lt;code&gt;stabilize.py&lt;/code&gt; only) filters; &lt;code&gt;doctor --json&lt;/code&gt; lists them as &lt;code&gt;capabilities.required&lt;/code&gt;. Needed only by the flags that use them: &lt;code&gt;libx265&lt;/code&gt;, &lt;code&gt;prores_ks&lt;/code&gt;, &lt;code&gt;libzimg&lt;/code&gt; / &lt;code&gt;zscale&lt;/code&gt;, &lt;code&gt;libmp3lame&lt;/code&gt;, &lt;code&gt;libopus&lt;/code&gt;, &lt;code&gt;libvorbis&lt;/code&gt;, the &lt;code&gt;ass&lt;/code&gt; filter. &lt;code&gt;doctor&lt;/code&gt; tells you which are present. The apt and gyan.dev builds carry all of them; some Homebrew bottles lack &lt;code&gt;libass&lt;/code&gt; / &lt;code&gt;libfreetype&lt;/code&gt; / &lt;code&gt;libzimg&lt;/code&gt;, which &lt;code&gt;doctor&lt;/code&gt; reports as missing.&lt;/li&gt; &#xA; &lt;li&gt;Python 3.10+, standard library only (3.0.0 removed 3.9)&lt;/li&gt; &#xA; &lt;li&gt;Node 16+ only for the &lt;code&gt;npx&lt;/code&gt; installer&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;p&gt;&lt;code&gt;doctor&lt;/code&gt;&#39;s own introspection calls (&lt;code&gt;ffmpeg -filters&lt;/code&gt;/&lt;code&gt;-encoders&lt;/code&gt;/&lt;code&gt;-bsfs&lt;/code&gt;/&lt;code&gt;-version&lt;/code&gt;) time out after 10s and report &lt;code&gt;failed&lt;/code&gt; rather than hanging forever — those are meant to be fast. Every tool&#39;s actual media-processing &lt;code&gt;ffmpeg&lt;/code&gt; invocation (cut, fit, caption, ...) runs under &lt;code&gt;--timeout&lt;/code&gt; (default 1800 s, &lt;code&gt;FFMPEG_SKILL_TIMEOUT&lt;/code&gt;, &lt;code&gt;0&lt;/code&gt; = none): past the limit the process is killed, its partial output removed, and the failure reported as &lt;code&gt;kind: timeout&lt;/code&gt; (exit 124) — a legitimately long &lt;code&gt;--accurate&lt;/code&gt; re-encode should raise the limit rather than run unbounded. &lt;code&gt;-nostdin&lt;/code&gt; is always passed, so a hung ffmpeg process waiting on stdin cannot happen.&lt;/p&gt; &#xA;&lt;h2&gt;Stability&lt;/h2&gt; &#xA;&lt;p&gt;3.x keeps every tool name, CLI argument, JSON output key and exit code working: nothing is removed or renamed, and nothing optional becomes required, until 4.0. The full list of what is promised and what is not, and the three-step deprecation policy, is in &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/contract.md#stability-guarantee-3x&#34;&gt;docs/contract.md&lt;/a&gt;. It is enforced by a test that pins every tool&#39;s argument names against a snapshot, so a breaking change fails CI instead of slipping into a patch. Releases are batched: related fixes and features ship together as one version rather than one version per merged pull request, so a dependent project sees fewer, larger steps. Pin a tagged version, not &lt;code&gt;main&lt;/code&gt;. 2.0.0 made the five changes announced since 1.10.0 -- &lt;code&gt;--crf&lt;/code&gt; removed in favour of &lt;code&gt;--quality&lt;/code&gt; (except on &lt;code&gt;export.py&lt;/code&gt;, whose preset chooses the encoder), an existing output refused without &lt;code&gt;--overwrite&lt;/code&gt;, &lt;code&gt;probe&lt;/code&gt;&#39;s &lt;code&gt;hdr&lt;/code&gt; narrowed to a real HDR signal, &lt;code&gt;json&lt;/code&gt;/&lt;code&gt;progress&lt;/code&gt; out of the MCP schema, the &lt;code&gt;result_v2&lt;/code&gt; preview withdrawn -- listed with what to change in &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/contract.md#what-20-changed&#34;&gt;docs/contract.md&lt;/a&gt; and under &lt;code&gt;removed&lt;/code&gt; in &lt;code&gt;contract --json&lt;/code&gt;. 3.0.0 is the second break: it removed FFmpeg 5.x and Python 3.9 (announced in 2.14.0 and 2.10.9), made &lt;code&gt;cut.py&lt;/code&gt; keep an &lt;code&gt;.mp4&lt;/code&gt;/&lt;code&gt;.mov&lt;/code&gt; copy&#39;s edit list, keep an SDR HEVC source HEVC and use &lt;code&gt;--vfr-guard sampled&lt;/code&gt; by default (&lt;code&gt;--no-edit-list&lt;/code&gt;, &lt;code&gt;--no-keep-hevc&lt;/code&gt; and &lt;code&gt;--vfr-guard average&lt;/code&gt; restore 2.x), and changed what &lt;code&gt;cut.py&lt;/code&gt;&#39;s &lt;code&gt;precision&lt;/code&gt; (now the least exact segment&#39;s) and &lt;code&gt;keyframe_snapped&lt;/code&gt; (now equal to &lt;code&gt;start_snapped&lt;/code&gt;) mean (&lt;code&gt;contract_version&lt;/code&gt; 2.0) -- see &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/contract.md#what-30-changed&#34;&gt;docs/contract.md&lt;/a&gt;.&lt;/p&gt; &#xA;&lt;h2&gt;Development&lt;/h2&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;npm test                      # tests/run_parallel.py (test_all.py&#39;s end-to-end modules, sharded across the cores) + tests/test_contract.py&#xA;python3 tests/test_picture.py # one tool group on its own (analysis, editing, cut_copy, audio, picture, delivery, orchestration, accel, asr)&#xA;npm run release-check         # pack, install, contract from the installed copy, MCP == contract, doctor, tests, contract evals&#xA;npm run demo                  # python3 demos/build.py: synthetic footage (and two AI-generated clips) -&amp;gt; every before/after demo + docs/demos/*.gif&#xA;npm run demo:pipeline         # examples/make_demo.sh: the older single end-to-end run, one clip through 13 of the 42 scripts&#xA;python3 evals/run.py --list   # agent eval prompts (see evals/); regex-only grading, runnable from any agent harness -- see evals/README.md &#34;Running this from Cursor, Codex, or another harness&#34;&#xA;node bin/install.js --dir /tmp/skills   # try the installer without touching ~/.claude&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;CI (&lt;code&gt;.github/workflows/ci.yml&lt;/code&gt;) runs on every pull request and on pushes to &lt;code&gt;main&lt;/code&gt;, on Ubuntu (FFmpeg 6.1 with Python 3.10 on 24.04, FFmpeg 8.0.1 with Python 3.13 on 26.04), macOS (Homebrew FFmpeg 8.x) and Windows (gyan.dev FFmpeg 9.x), plus a Linux job on FFmpeg 7.1 (Debian trixie container), and uploads each runner&#39;s FFmpeg listings as an artifact.&lt;/p&gt; &#xA;&lt;p&gt;&lt;code&gt;tests/test_contract.py&lt;/code&gt; runs on all three OSes, but some of its tests put a fake &lt;code&gt;ffmpeg&lt;/code&gt; or &lt;code&gt;ffprobe&lt;/code&gt; (a &lt;code&gt;#!/bin/sh&lt;/code&gt; script on a PATH shim) in front of the real one — to record every call under &lt;code&gt;--dry-run&lt;/code&gt;, to force specific FFmpeg 6/7/8/9 fixture layouts through &lt;code&gt;doctor&lt;/code&gt;&#39;s parser, to report an old version, to hang or fail on cue — or need what Windows does not have or does differently (SIGTERM, unprivileged symlinks, a line break or a backslash in a file name). Those tests, among them &lt;code&gt;test_dry_run_never_runs_ffmpeg_and_writes_nothing&lt;/code&gt; and the whole &lt;code&gt;DoctorDetectionTests&lt;/code&gt; class, skip there (&lt;code&gt;skipIf&lt;/code&gt;, or &lt;code&gt;skipTest&lt;/code&gt; inside the test) and show as &lt;code&gt;skipped&lt;/code&gt;, not silently absent, in the Windows job&#39;s log; &lt;code&gt;grep -n &#39;os.name == &#34;nt&#34;\|platform.system() == &#34;Windows&#34;&#39; tests/test_contract.py&lt;/code&gt; finds their conditions, among the suite&#39;s other platform branches. The rest — contract schema, &lt;code&gt;reencodes_*&lt;/code&gt;, &lt;code&gt;doctor.tools&lt;/code&gt;, MCP derivation, and the tools exercised through the contract, including &lt;code&gt;cut.py&lt;/code&gt;&#39;s provenance fields — runs against the real Windows &lt;code&gt;ffmpeg&lt;/code&gt; on every PR. See &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/references/ci-platform-pitfalls.md&#34;&gt;references/ci-platform-pitfalls.md&lt;/a&gt; for this and other per-OS behaviour differences already diagnosed, before spending a CI cycle re-diagnosing a platform-only failure.&lt;/p&gt; &#xA;&lt;p&gt;&lt;strong&gt;Releasing&lt;/strong&gt; is fully automated end to end, including the version number itself — a PR never needs to touch &lt;code&gt;package.json&lt;/code&gt; or the version in &lt;code&gt;docs/contract.md&lt;/code&gt;. &lt;code&gt;CHANGELOG.md&lt;/code&gt; is different: the release writes the version heading and one line per merged PR, but a &lt;code&gt;feat&lt;/code&gt; PR writes its own notes under &lt;code&gt;## Unreleased&lt;/code&gt; (&lt;a href=&#34;http://CONTRIBUTING.md&#34;&gt;CONTRIBUTING.md&lt;/a&gt;), and a PR that closes an issue must name that issue in &lt;code&gt;CHANGELOG.md&lt;/code&gt; or &lt;code&gt;test_changelog_mentions_every_closed_issue_since_last_tag&lt;/code&gt; fails. Once a PR merges to &lt;code&gt;main&lt;/code&gt;, &lt;code&gt;.github/workflows/release.yml&lt;/code&gt; takes it from there: if nobody bumped the version by hand, it resolves the next version (&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/.github/scripts/resolve_version.py&#34;&gt;.github/scripts/resolve_version.py&lt;/a&gt;) from the labels on every PR merged since the last tag (&lt;code&gt;minor&lt;/code&gt;/&lt;code&gt;feature&lt;/code&gt;/&lt;code&gt;enhancement&lt;/code&gt; → minor, &lt;code&gt;fix&lt;/code&gt;/&lt;code&gt;bug&lt;/code&gt;/&lt;code&gt;patch&lt;/code&gt; → patch, an unlabeled PR defaults to &lt;code&gt;patch&lt;/code&gt;; a PR whose only labels are &lt;code&gt;chore&lt;/code&gt;, &lt;code&gt;ci&lt;/code&gt;, &lt;code&gt;docs&lt;/code&gt; or &lt;code&gt;dependencies&lt;/code&gt; is not releasable, so a merge that changes no shipped file releases nothing). Most PRs don&#39;t need a label added by hand: &lt;code&gt;release-drafter.yml&lt;/code&gt;&#39;s &lt;code&gt;autolabel&lt;/code&gt; job applies one automatically from the PR&#39;s title/changed files (&lt;code&gt;Fix ...&lt;/code&gt; → &lt;code&gt;fix&lt;/code&gt;, &lt;code&gt;Add ...&lt;/code&gt;/&lt;code&gt;feat ...&lt;/code&gt; → &lt;code&gt;feature&lt;/code&gt;, &lt;code&gt;docs&lt;/code&gt;/&lt;code&gt;.md&lt;/code&gt;/&lt;code&gt;.github/&lt;/code&gt;/&lt;code&gt;build(deps)&lt;/code&gt; changes → &lt;code&gt;chore&lt;/code&gt;) as soon as it&#39;s opened — add a label yourself only to override that. &lt;strong&gt;A major version is never chosen automatically&lt;/strong&gt;: no label rule produces &lt;code&gt;major&lt;/code&gt;, and the workflow refuses to auto-bump across a major boundary even if someone applies that label — a real major release is a deliberate &lt;code&gt;package.json&lt;/code&gt; bump in a PR, which the manual path below already handles. It bumps &lt;code&gt;package.json&lt;/code&gt;, &lt;code&gt;.claude-plugin/plugin.json&lt;/code&gt;, &lt;code&gt;docs/contract.md&lt;/code&gt; and &lt;code&gt;docs/roadmap.md&lt;/code&gt;&#39;s released-version sentence, writes a &lt;code&gt;CHANGELOG.md&lt;/code&gt; section listing those PRs (and any issues they closed), and pushes that commit to &lt;code&gt;main&lt;/code&gt; itself. If another PR merged while the bump was being prepared, the push is refused and that run releases nothing: the newer merge&#39;s own run, queued behind it, starts from &lt;code&gt;main&lt;/code&gt;&#39;s tip and releases both (the bump is never rebased onto a merge it did not read). After the tag it also asks &lt;code&gt;pages.yml&lt;/code&gt; to rebuild the site, because the bump commit carries &lt;code&gt;[skip ci]&lt;/code&gt; and would otherwise leave the site on the previous version. Either way — auto-bumped or hand-bumped in the PR — it then creates the &lt;code&gt;vX.Y.Z&lt;/code&gt; tag, publishes a GitHub Release with notes extracted from &lt;code&gt;CHANGELOG.md&lt;/code&gt;&#39;s matching section, and publishes the package to npm. A PR that still wants to write its own version bump and &lt;code&gt;CHANGELOG.md&lt;/code&gt; prose (e.g. to explain the &#34;why&#34; of a release by hand) can — the automation only fills in when nobody made that call already. A push to &lt;code&gt;main&lt;/code&gt; with nothing new to release creates no tag, Release or npm publish; its MCP Registry step still asks npm and the registry about &lt;code&gt;server.json&lt;/code&gt;&#39;s version and publishes the entry only when npm has that version and the registry does not, which is how a later run finishes an entry an earlier one failed to publish. npm publishing needs an &lt;code&gt;NPM_TOKEN&lt;/code&gt; repo secret (an npm access token with publish rights on this package) — without it the tag and GitHub Release still happen, only the npm step is skipped. A repo that depends on this one (an editing skill, an agent) should pin an &lt;code&gt;ffmpeg-skill&lt;/code&gt; version by tag or npm version, not by tracking &lt;code&gt;main&lt;/code&gt; — a merged-but-not-yet-released commit on &lt;code&gt;main&lt;/code&gt; can be ahead of the last published npm version for the few minutes between merge and this workflow completing.&lt;/p&gt; &#xA;&lt;p&gt;&lt;strong&gt;A version that was tagged but never reached npm&lt;/strong&gt; (the npm step failed; a re-run of that release sees the tag and skips the publish) is finished by dispatching the same workflow on its tag: &lt;code&gt;gh workflow run release.yml --ref vX.Y.Z -f version=X.Y.Z&lt;/code&gt;. That runs only the workflow&#39;s &lt;code&gt;recover&lt;/code&gt; job. It refuses a version that is not &lt;code&gt;X.Y.Z&lt;/code&gt;, has no tag, or that npm already has, and a run whose commit is not the tagged one: the npm provenance attestation names the run&#39;s commit, so it must be the commit the package is built from. It publishes with &lt;code&gt;--provenance&lt;/code&gt; under the &lt;code&gt;latest&lt;/code&gt; dist-tag when the version is above every stable (non-prerelease) version npm has, and under &lt;code&gt;backport&lt;/code&gt; otherwise, so &lt;code&gt;latest&lt;/code&gt; never moves back. Then it publishes the &lt;code&gt;plugin&lt;/code&gt; branch when no commit there is for this version and the branch&#39;s tip is older, and the MCP Registry entry when the registry does not list the version (never for a &lt;code&gt;backport&lt;/code&gt;). The GitHub Release is left as it is. A tag cut before the job was added (#448; every tag up to v3.6.3) has no &lt;code&gt;recover&lt;/code&gt; job at its commit and cannot be recovered this way.&lt;/p&gt; &#xA;&lt;p&gt;Contributing a change: see &lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/CONTRIBUTING.md&#34;&gt;CONTRIBUTING.md&lt;/a&gt;.&lt;/p&gt; &#xA;&lt;h2&gt;Docs&lt;/h2&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;&lt;/th&gt; &#xA;   &lt;th&gt;&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/CONTRIBUTING.md&#34;&gt;CONTRIBUTING.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;scope, dev setup, tests, PR expectations&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/roadmap.md&#34;&gt;docs/roadmap.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;the release history from 1.8.0, one theme per minor, each marked shipped + evaluated, shipped with eval pending, or planned; and what 2.0.0 then removes&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/design-decisions.md&#34;&gt;docs/design-decisions.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;behaviours that look like bugs but are decisions, with rationale and the pinning test; read before filing a bug&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/CODE_OF_CONDUCT.md&#34;&gt;CODE_OF_CONDUCT.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Contributor Covenant 2.1; reports go through the &lt;a href=&#34;http://SECURITY.md&#34;&gt;SECURITY.md&lt;/a&gt; channel&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/SECURITY.md&#34;&gt;SECURITY.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;how to report a vulnerability privately&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/SKILL.md&#34;&gt;SKILL.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;what the agent reads: workflow, request → tool map, audio-only rules, report format, pitfalls&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/references/scripts.md&#34;&gt;references/scripts.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;per-flag reference for every tool&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/references/scripts.md#verdict-thresholds-what-passes&#34;&gt;references/scripts.md § Verdict thresholds&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;every number behind &lt;code&gt;verified&lt;/code&gt; and &lt;code&gt;check.py&lt;/code&gt;&#39;s rows (tolerances, silence levels, per-platform loudness and length), each with the constant or flag it comes from&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/references/devices.md&#34;&gt;references/devices.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;real-device notes (iPhone HDR, GoPro, DJI, screen recordings)&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/references/ci-platform-pitfalls.md&#34;&gt;references/ci-platform-pitfalls.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;per-OS ffmpeg/CI behaviour differences already diagnosed once — read before re-diagnosing a Windows/macOS-only test failure&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/references/process-pitfalls.md&#34;&gt;references/process-pitfalls.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;process mistakes already made once (breaking a pinned test by narrowing a capability list, retrying a git/GitHub operation this environment can&#39;t do, re-designing a fixture instead of recognising a real platform difference) — a living record, add to it whenever one recurs&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/docs/contract.md&#34;&gt;docs/contract.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;the execution contract field by field, MCP relationship, how a planner consumes it&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/examples/README.md&#34;&gt;examples/README.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;natural-language requests and the commands behind them, &lt;code&gt;brand.json&lt;/code&gt;, &lt;code&gt;project.json&lt;/code&gt;, batch recipes&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/tests/fixtures/README.md&#34;&gt;tests/fixtures/README.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;captured and constructed FFmpeg listings, which is which&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/CHANGELOG.md&#34;&gt;CHANGELOG.md&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;what changed in each release&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h2&gt;Support&lt;/h2&gt; &#xA;&lt;p&gt;If this skill saves you time, you can help keep it maintained through &lt;a href=&#34;https://github.com/sponsors/kajisho5&#34;&gt;GitHub Sponsors&lt;/a&gt;. Issues and pull requests are just as welcome.&lt;/p&gt; &#xA;&lt;h2&gt;License&lt;/h2&gt; &#xA;&lt;p&gt;&lt;a href=&#34;https://raw.githubusercontent.com/kajisho5/ffmpeg-skill/main/LICENSE&#34;&gt;MIT&lt;/a&gt;&lt;/p&gt;</summary>
  </entry>
  <entry>
    <title>huytieu/COG-second-brain</title>
    <updated>2026-11-16T07:43:55Z</updated>
    <id>tag:github.com,2026-11-16:/huytieu/COG-second-brain</id>
    <link href="https://github.com/huytieu/COG-second-brain" rel="alternate"></link>
    <summary type="html">&lt;p&gt;Self-evolving second brain with 35 AI skills, 10 agents, and people CRM. Closed-loop harness: a V-model verification lifecycle where the worker never grades its own homework. Plus paired anti-slop design skills for marketing and product UI. Works with Claude Code, Cursor, Kiro, Gemini CLI, Codex.&lt;/p&gt;&lt;hr&gt;&lt;h1&gt;COG: The Agentic Second Brain That Actually Self-Evolves&lt;/h1&gt; &#xA;&lt;p&gt;&lt;strong&gt;Cognition + Obsidian + Git&lt;/strong&gt; — A self-evolving second brain powered by AI agents, markdown files, and version control. No database, no vendor lock-in — just &lt;code&gt;.md&lt;/code&gt; files that think.&lt;/p&gt; &#xA;&lt;p&gt;&lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/#quick-start&#34;&gt;Quick Start&lt;/a&gt; | &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/#skills&#34;&gt;Skills&lt;/a&gt; | &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/#features-at-a-glance&#34;&gt;Features&lt;/a&gt; | &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/#faq&#34;&gt;FAQ&lt;/a&gt; | &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/SETUP.md&#34;&gt;SETUP.md&lt;/a&gt;&lt;/p&gt; &#xA;&lt;blockquote&gt; &#xA; &lt;p&gt;Works with &lt;a href=&#34;https://claude.ai/download&#34;&gt;Claude Code&lt;/a&gt; • Antigravity • &lt;a href=&#34;https://cursor.com/&#34;&gt;Cursor&lt;/a&gt; • &lt;a href=&#34;https://kiro.dev/&#34;&gt;Kiro&lt;/a&gt; • &lt;a href=&#34;https://github.com/google-gemini/gemini-cli&#34;&gt;Gemini CLI&lt;/a&gt; • &lt;a href=&#34;https://github.com/openai/codex&#34;&gt;OpenAI Codex&lt;/a&gt; • any AI that reads markdown&lt;/p&gt; &#xA; &lt;p&gt;Inspired by &lt;a href=&#34;https://github.com/garrytan/gstack&#34;&gt;Garry Tan&#39;s gstack&lt;/a&gt; and &lt;a href=&#34;https://github.com/garrytan/gbrain&#34;&gt;gbrain&lt;/a&gt;&lt;/p&gt; &#xA;&lt;/blockquote&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-mermaid&#34;&gt;graph LR&#xA;    A[You] -- natural language --&amp;gt; B[AI Agent]&#xA;    B -- runs --&amp;gt; C[43 Skills]&#xA;    C -- delegates to --&amp;gt; W[6 Workers]&#xA;    C -- verified by --&amp;gt; V[4 Read-Only Verifiers]&#xA;    C -- reads &amp;amp; writes --&amp;gt; D[.md Files]&#xA;    W -- reads &amp;amp; writes --&amp;gt; D&#xA;    V -- observes --&amp;gt; D&#xA;    C -- syncs with --&amp;gt; G[GitHub / Linear / Slack / PostHog]&#xA;    D --&amp;gt; E[Git]&#xA;    D --&amp;gt; F[iCloud]&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;h2&gt;Quick Start&lt;/h2&gt; &#xA;&lt;p&gt;&lt;strong&gt;1. Clone &amp;amp; enter the repo:&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;git clone https://github.com/huytieu/COG-second-brain.git&#xA;cd COG-second-brain&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;&lt;strong&gt;2. Run onboarding in your agent:&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Agent&lt;/th&gt; &#xA;   &lt;th&gt;Command&lt;/th&gt; &#xA;   &lt;th&gt;How it finds skills&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Claude Code&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;code .&lt;/code&gt; → &#34;Run onboarding&#34;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;.claude/skills/&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Antigravity&lt;/td&gt; &#xA;   &lt;td&gt;Open folder → &#34;Run onboarding&#34;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;.agents/skills/&lt;/code&gt; + &lt;code&gt;.agents/rules/cog.md&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Cursor&lt;/td&gt; &#xA;   &lt;td&gt;Open folder → &#34;Run onboarding&#34;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;.cursor-plugin/&lt;/code&gt; + &lt;code&gt;.cursorrules&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Kiro&lt;/td&gt; &#xA;   &lt;td&gt;Open folder → &#34;setup COG&#34;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;.kiro/powers/&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Gemini CLI&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;gemini&lt;/code&gt; → &lt;code&gt;/onboarding&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;GEMINI.md&lt;/code&gt; + &lt;code&gt;.gemini/commands/&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;OpenAI Codex&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;codex&lt;/code&gt; → &#34;Run onboarding&#34;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;AGENTS.md&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Other agents&lt;/td&gt; &#xA;   &lt;td&gt;Point at &lt;code&gt;AGENTS.md&lt;/code&gt; → &#34;Run onboarding&#34;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;AGENTS.md&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;Or install via &lt;a href=&#34;https://skills.sh&#34;&gt;skills.sh&lt;/a&gt;:&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;npx skills add huytieu/COG-second-brain&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;p&gt;Done — COG is personalized and ready in ~2 minutes. See &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/SETUP.md&#34;&gt;SETUP.md&lt;/a&gt; for optional config (Git sync, iCloud, Obsidian Tasks, etc.).&lt;/p&gt; &#xA;&lt;h2&gt;Agent Support Matrix&lt;/h2&gt; &#xA;&lt;p&gt;COG ships a &lt;strong&gt;full Claude Code surface&lt;/strong&gt;, a &lt;strong&gt;full Antigravity surface&lt;/strong&gt;, plus &lt;strong&gt;core native surfaces&lt;/strong&gt; for Kiro and Gemini CLI, with &lt;code&gt;AGENTS.md&lt;/code&gt; as the universal fallback for Codex and other markdown-reading agents.&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Surface&lt;/th&gt; &#xA;   &lt;th&gt;Current support&lt;/th&gt; &#xA;   &lt;th&gt;Notes&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Claude Code&lt;/td&gt; &#xA;   &lt;td&gt;33 native skills + 10 agents (6 workers + 4 verifiers)&lt;/td&gt; &#xA;   &lt;td&gt;Full first-class surface&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Antigravity&lt;/td&gt; &#xA;   &lt;td&gt;43 skills + 10 agents (pointer-stub format)&lt;/td&gt; &#xA;   &lt;td&gt;Full surface — thin stubs in &lt;code&gt;.agents/&lt;/code&gt; delegate to the &lt;code&gt;.claude/&lt;/code&gt; playbooks, which stay authoritative&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://agent-plugins.org&#34;&gt;Agent Plugins&lt;/a&gt; standard&lt;/td&gt; &#xA;   &lt;td&gt;Root &lt;code&gt;plugin.json&lt;/code&gt; + &lt;code&gt;skills/&lt;/code&gt; (43 skills)&lt;/td&gt; &#xA;   &lt;td&gt;Spec 1.0.0 conformant; any standard-aware client loads COG as a plugin&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Cursor&lt;/td&gt; &#xA;   &lt;td&gt;Plugin manifest + rules&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;.cursor-plugin/plugin.json&lt;/code&gt; + &lt;code&gt;.cursorrules&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Kiro&lt;/td&gt; &#xA;   &lt;td&gt;7 native powers&lt;/td&gt; &#xA;   &lt;td&gt;Core workflows today&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Gemini CLI&lt;/td&gt; &#xA;   &lt;td&gt;7 native commands&lt;/td&gt; &#xA;   &lt;td&gt;Core workflows today&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;code&gt;AGENTS.md&lt;/code&gt;&lt;/td&gt; &#xA;   &lt;td&gt;33 documented commands&lt;/td&gt; &#xA;   &lt;td&gt;Universal fallback for Codex and other agents&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;Before publishing or updating framework files, run &lt;code&gt;./scripts/validate-agent-surface.sh&lt;/code&gt; to catch drift between manifests, docs, and shipped files. See &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/docs/AGENT-SUPPORT.md&#34;&gt;docs/AGENT-SUPPORT.md&lt;/a&gt; for the detailed support matrix and contributor rules.&lt;/p&gt; &#xA;&lt;h2&gt;Skills&lt;/h2&gt; &#xA;&lt;h3&gt;Core Skills (Personal Knowledge)&lt;/h3&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Skill&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Try saying...&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;onboarding&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Personalize COG for your workflow (run first!)&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Run onboarding&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;braindump&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Capture raw thoughts with intelligent classification&lt;/td&gt; &#xA;   &lt;td&gt;&#34;I need to braindump&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;daily-brief&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Verified news intelligence (7-day freshness)&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Give me my daily brief&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;url-dump&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Save URLs with auto-extracted insights&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Save this URL&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;weekly-checkin&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Cross-domain pattern analysis&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Weekly review&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;knowledge-consolidation&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Build frameworks from scattered notes&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Consolidate my knowledge&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;update-cog&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Update framework files without touching your content&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Update COG&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;memory-hygiene&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Trust sweep of persistent memory — re-verify claims against the live environment, stamp &lt;code&gt;last_verified&lt;/code&gt; + confidence&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Audit my memories&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h3&gt;Team Intelligence Skills (for Product &amp;amp; Engineering Leads)&lt;/h3&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Skill&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Try saying...&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;team-brief&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Cross-reference GitHub + Linear + Slack + PostHog into a daily team intelligence brief with two-way Linear sync-back&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Team brief&#34; / &#34;What did we ship?&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;meeting-transcript&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Process meeting recordings into structured decisions, action items, and team dynamics&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Process this meeting&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;comprehensive-analysis&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Deep 7-day analysis for weekly reviews, board prep, or strategic planning (~8-12 min)&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Weekly analysis&#34; / &#34;Board prep&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h3&gt;PM Workflow Skills (for Product Managers)&lt;/h3&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Skill&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Try saying...&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;create-user-story&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Create user stories with duplicate checking across Linear, GitHub Issues, or Jira&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Create a user story for...&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;generate-prd&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Draft PRDs with approval gate before publishing to Confluence/Notion&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Generate a PRD for...&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;generate-release-notes&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Generate release notes from GitHub milestones, Linear cycles, or manual input&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Generate release notes for v2.1&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;export-open-issues&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Audit and export open issues from any tracker into a structured vault summary&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Export open issues&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;publish-to-confluence&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Publish any vault markdown file to Confluence&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Publish this to Confluence&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;update-knowledge-base&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Maintain product knowledge base from releases, features, and project changes&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Update the knowledge base with v2.1 changes&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;release-comms&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Turn a runsheet and a milestone into a release note with real media, docs updates, and a Slack draft&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Cover today&#39;s release&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;feature-doc&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Drive the live app, capture screenshots and a recording, and write the docs article&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Write the doc for issue #123&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;blockquote&gt; &#xA; &lt;p&gt;&lt;strong&gt;PM Workflow:&lt;/strong&gt; These skills form a complete product management lifecycle: &lt;strong&gt;Research&lt;/strong&gt; (&lt;code&gt;/auto-research&lt;/code&gt;) → &lt;strong&gt;PRD&lt;/strong&gt; (&lt;code&gt;/generate-prd&lt;/code&gt;) → &lt;strong&gt;Stories&lt;/strong&gt; (&lt;code&gt;/create-user-story&lt;/code&gt;) → Development → &lt;strong&gt;Release Notes&lt;/strong&gt; (&lt;code&gt;/generate-release-notes&lt;/code&gt;) → &lt;strong&gt;Knowledge Base&lt;/strong&gt; (&lt;code&gt;/update-knowledge-base&lt;/code&gt;). Use &lt;code&gt;/export-open-issues&lt;/code&gt; for audits and &lt;code&gt;/publish-to-confluence&lt;/code&gt; to share externally.&lt;/p&gt; &#xA;&lt;/blockquote&gt; &#xA;&lt;h3&gt;Strategic Research&lt;/h3&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Skill&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Try saying...&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;auto-research&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Deep strategic research engine — decomposes questions into parallel research threads with multiple agents&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Research the future of AI testing tools&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h3&gt;Content Creation&lt;/h3&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Skill&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Try saying...&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;content-factory&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Autonomous content pipeline — scout announcements, triage by trend momentum + your unique angle, publish with ledger dedup, hard volume caps, and screenshot-verified posting&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Run the content factory&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;blog-publish&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Write and publish a post to your own blog, with a markdown twin and social distribution&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Publish this to my blog&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;social-media-kit&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Teaser video, GIF, and card images for every social post&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Make the social kit for this post&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;youtube-video-kit&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Thumbnail, opening title, music bed, and upload metadata&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Make a thumbnail and intro for this video&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h3&gt;Verification Harness&lt;/h3&gt; &#xA;&lt;p&gt;Opt-in. Ask for one of these and work walks a &lt;strong&gt;V&lt;/strong&gt;: decompose left into falsifiable criteria, build at the apex, verify right with evidence traced back to each criterion. Say nothing and none of it runs: ordinary work carries no checkpoints and no evidence ledger. Turn it on per request, or by default for build tasks with &lt;code&gt;verification_harness: on&lt;/code&gt; in &lt;code&gt;00-inbox/MY-PROFILE.md&lt;/code&gt;. Full lifecycle in &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/WORKFLOW.md&#34;&gt;WORKFLOW.md&lt;/a&gt;.&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Skill&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Try saying...&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;closed-loop&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Plan → build → component verify → integration verify → acceptance. The worker never grades its own homework; verifiers observe the artifact, not the worker&#39;s summary&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Run this through the closed loop&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;ultragoal&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;A goal too big for one session, run as a chain of phases with cross-session state and a north-star acceptance gate&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Track this as an ultragoal&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;harvest&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Capture session learnings, stage them for approval, propose skill patches. Never writes durable knowledge unapproved&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Harvest what we learned&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;retro&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Audit the run&#39;s checkpoints and &lt;em&gt;evidence quality&lt;/em&gt;. A run can pass every gate on weak observations&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Retro this run&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;review-cockpit&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;One living doc per multi-item session: progress, per-item cards, and an approval slot you edit in place&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Set up a review cockpit&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h3&gt;Craft Skills (Writing &amp;amp; Visuals)&lt;/h3&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Skill&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Try saying...&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;no-ai-slop&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Sharpen a draft without flattening its voice, or just detect AI-slop patterns&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Make this less AI-sounding&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;slop-gate&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Refuse slop tells in anything about to be written, published, or sent; runs in CI or as a hook&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Scan this before I publish&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;voice-baseline&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Measure the words and shapes you over-use so an agent stops amplifying your tics&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Why do my drafts all sound the same?&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;release-video&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Turn a release into a motion recap and explained per-feature demos, with sound tied to motion&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Make a video for this release&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;terminal-replica&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Draw your real terminal and CLI agents in graphics and video frames&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Draw a Claude Code terminal for this post&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;vector-cover&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Flat vector cover illustrations rendered to PNG&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Make a cover image for this article&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;editorial-illustrations&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Claim → geometry: derive the &lt;em&gt;right&lt;/em&gt; figure from what the text argues, render it as self-contained theme-aware HTML/SVG&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Make a diagram for this argument&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;data-forms&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;20+ chart and diagram forms with when-to-use, failure modes, and the encoding discipline that makes any of them readable&lt;/td&gt; &#xA;   &lt;td&gt;&#34;The bar chart is burying the point&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;museum-art&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Real public-domain artwork from museum open-access APIs instead of AI-generated or stock imagery&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Find a hero image for this post&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;daily-journal&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;A passive work journal the agent keeps for you, plus a guided reflection on request&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Reflect on today&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h3&gt;Design Skills (Anti-Slop UI)&lt;/h3&gt; &#xA;&lt;p&gt;A paired set. &lt;code&gt;taste-skill&lt;/code&gt; owns first-impression surfaces, &lt;code&gt;product-ui-taste&lt;/code&gt; owns surfaces people use every day. Never run both on the same component.&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Skill&lt;/th&gt; &#xA;   &lt;th&gt;Owns&lt;/th&gt; &#xA;   &lt;th&gt;Try saying...&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;taste-skill&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Landing pages, portfolios, marketing, editorial. States a &#34;Design Read&#34; before writing code so it stops defaulting to one house aesthetic&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Build a landing page for this&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;product-ui-taste&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Dashboards, data tables, forms, wizards, settings, admin consoles. Budgets the frame in px first, then survives real data: overflow, long labels, empty/error/permission-denied states&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Build the admin table for this&#34;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h3&gt;Agents (Specialist Sessions)&lt;/h3&gt; &#xA;&lt;p&gt;COG uses a worker agent architecture inspired by &lt;a href=&#34;https://github.com/garrytan/gstack&#34;&gt;garrytan/gstack&lt;/a&gt; specialist sessions and &lt;a href=&#34;https://github.com/garrytan/gbrain&#34;&gt;garrytan/gbrain&lt;/a&gt; knowledge patterns. Workers handle data-heavy tasks cheaply (Sonnet) while the lead session does reasoning (Opus).&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Agent&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Model&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;worker-data-collector&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Structured extraction from GitHub, Slack, Jira, Linear&lt;/td&gt; &#xA;   &lt;td&gt;Sonnet&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;worker-researcher&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Web research with source citations&lt;/td&gt; &#xA;   &lt;td&gt;Sonnet&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;worker-file-ops&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Vault file operations, metadata, profiles&lt;/td&gt; &#xA;   &lt;td&gt;Sonnet&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;worker-executor&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Pre-approved mutations (Jira, Linear, APIs)&lt;/td&gt; &#xA;   &lt;td&gt;Sonnet&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;worker-publisher&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Publishing to Slack, Confluence, Notion&lt;/td&gt; &#xA;   &lt;td&gt;Sonnet&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;brief-people-updater&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Batch-update people profiles from meetings/briefs&lt;/td&gt; &#xA;   &lt;td&gt;Sonnet&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;&lt;strong&gt;Read-only verifiers&lt;/strong&gt; (they cannot edit files or mutate external state):&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Agent&lt;/th&gt; &#xA;   &lt;th&gt;What it does&lt;/th&gt; &#xA;   &lt;th&gt;Checkpoint&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;task-verifier&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Checks output against acceptance criteria by observing the artifact&lt;/td&gt; &#xA;   &lt;td&gt;CP-3v&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;integration-verifier&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Cross-task wiring and global acceptance&lt;/td&gt; &#xA;   &lt;td&gt;CP-4&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;fix-agent&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Targeted fixes after a &lt;code&gt;FAIL:fixable&lt;/code&gt; verdict; max 2 attempts&lt;/td&gt; &#xA;   &lt;td&gt;CP-3v retry&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;harvest-curator&lt;/strong&gt;&lt;/td&gt; &#xA;   &lt;td&gt;Shapes session learnings into adoption notes; propose-only&lt;/td&gt; &#xA;   &lt;td&gt;CP-7&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;blockquote&gt; &#xA; &lt;p&gt;Workers write results to &lt;code&gt;/tmp/&lt;/code&gt; files and return only a status + path. The lead reads the file for synthesis. This eliminates slow token generation in agent output.&lt;/p&gt; &#xA; &lt;p&gt;Verifiers receive &lt;strong&gt;paths only&lt;/strong&gt;, never a worker&#39;s output. Pasted context induces narrativisation: the verifier classifies the framing instead of independently reading the source.&lt;/p&gt; &#xA;&lt;/blockquote&gt; &#xA;&lt;h3&gt;People CRM (Knowledge-Based Team Profiles)&lt;/h3&gt; &#xA;&lt;p&gt;Track the people you work with using progressive, evidence-based profiles in &lt;code&gt;05-knowledge/people/&lt;/code&gt;. Profiles auto-escalate via tiered enrichment:&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;Tier 3 (Stub)&lt;/strong&gt; — 1 mention: name, role, one-line context&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Tier 2 (Moderate)&lt;/strong&gt; — 3+ mentions: executive snapshot, working style, strengths&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Tier 1 (Full)&lt;/strong&gt; — 8+ mentions or direct meeting: complete profile with all sections&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;p&gt;Every observation includes a source citation with confidence level. See &lt;code&gt;05-knowledge/people/README.md&lt;/code&gt; for details.&lt;/p&gt; &#xA;&lt;h3&gt;Role Packs (Personalized Recommendations)&lt;/h3&gt; &#xA;&lt;p&gt;COG matches your role during onboarding to a &lt;strong&gt;role pack&lt;/strong&gt; that prioritizes the most relevant skills and integrations for you. Available role packs: Product Manager, Engineering Lead, Engineer, Designer, Founder, Marketer — or create your own from the template.&lt;/p&gt; &#xA;&lt;blockquote&gt; &#xA; &lt;p&gt;&lt;strong&gt;New to team skills?&lt;/strong&gt; These require GitHub CLI (&lt;code&gt;gh&lt;/code&gt;) and work best with Linear, Slack, and PostHog MCP integrations. They degrade gracefully — start with just GitHub and add integrations over time. See &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/SETUP.md&#34;&gt;SETUP.md&lt;/a&gt; for configuration.&lt;/p&gt; &#xA;&lt;/blockquote&gt; &#xA;&lt;h2&gt;The Evolution Cycle&lt;/h2&gt; &#xA;&lt;pre&gt;&lt;code class=&#34;language-mermaid&#34;&gt;graph TD&#xA;    A[Daily: Braindump thoughts] --&amp;gt; B[Daily: News intelligence]&#xA;    A --&amp;gt; T[Daily: Team intelligence brief]&#xA;    M[Meetings: Process transcripts] --&amp;gt; T&#xA;    T -- syncs back to --&amp;gt; L[Linear / GitHub]&#xA;    B --&amp;gt; C[Weekly: Pattern analysis]&#xA;    T --&amp;gt; CA[Weekly: Comprehensive analysis]&#xA;    C --&amp;gt; D[Monthly: Knowledge consolidation]&#xA;    CA --&amp;gt; D&#xA;    D -- COG learns your patterns --&amp;gt; A&#xA;    R[Strategic: Auto-research] --&amp;gt; P[PM: Generate PRD]&#xA;    P --&amp;gt; S[PM: Create user stories]&#xA;    S -- after development --&amp;gt; RN[PM: Release notes]&#xA;    RN --&amp;gt; KB[PM: Update knowledge base]&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;strong&gt;Daily capture&lt;/strong&gt; — braindump raw thoughts; COG classifies by domain and extracts action items&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Daily intelligence&lt;/strong&gt; — personalized news briefings with verified, sourced news&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Daily team brief&lt;/strong&gt; — cross-reference GitHub, Linear, Slack, PostHog, meetings into one brief with two-way sync&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Meeting processing&lt;/strong&gt; — extract decisions, action items, and team dynamics from transcripts&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Weekly reflection&lt;/strong&gt; — pattern analysis across all domains surfaces insights you&#39;d miss&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Weekly deep dive&lt;/strong&gt; — comprehensive analysis for board prep, retros, and strategic planning&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Monthly synthesis&lt;/strong&gt; — scattered notes become consolidated frameworks and a knowledge base&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Strategic research&lt;/strong&gt; — deep multi-agent investigation of strategic questions with real sources&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;PM workflow&lt;/strong&gt; — full product lifecycle from PRD to release notes to knowledge base updates&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;h2&gt;Features at a Glance&lt;/h2&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;&lt;/th&gt; &#xA;   &lt;th&gt;&lt;/th&gt; &#xA;   &lt;th&gt;&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Self-Evolving&lt;/strong&gt; — Learns your patterns, auto-organizes content, builds frameworks&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Self-Healing&lt;/strong&gt; — Rename files or restructure; cross-references update automatically&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Verification-First&lt;/strong&gt; — Sources required, 7-day freshness, confidence levels on all analysis&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Privacy-First&lt;/strong&gt; — Local &lt;code&gt;.md&lt;/code&gt; files, strict domain separation, no external servers&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Multi-Device&lt;/strong&gt; — iCloud sync to iPhone/iPad/Mac; Git for version history&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Obsidian Tasks&lt;/strong&gt; — &lt;code&gt;📅 YYYY-MM-DD&lt;/code&gt; emoji format works with Tasks plugin dashboards&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Garry Tan Inspired&lt;/strong&gt; — gstack specialist sessions + gbrain knowledge patterns&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Multi-Platform&lt;/strong&gt; — Listed on &lt;a href=&#34;https://skills.sh&#34;&gt;skills.sh&lt;/a&gt;, &lt;a href=&#34;https://agentskill.sh&#34;&gt;agentskill.sh&lt;/a&gt;, &lt;a href=&#34;https://cursor.directory&#34;&gt;cursor.directory&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Worker Agents&lt;/strong&gt; — Sonnet handles I/O, Opus handles thinking&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Runtime Trust&lt;/strong&gt; — memory hygiene sweeps re-verify stored facts against the live environment&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Post-Condition Checks&lt;/strong&gt; — mutating skills observe the artifact before reporting success&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Skill Distillation&lt;/strong&gt; — &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/docs/SKILL-DISTILLATION.md&#34;&gt;explore once, execute cheap&lt;/a&gt; with small models&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;strong&gt;V-Model Closed Loop&lt;/strong&gt; — every criterion has a task on the left and an evidence row on the right&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Risk Lanes&lt;/strong&gt; — ceremony scales with blast radius&lt;/td&gt; &#xA;   &lt;td&gt;&lt;strong&gt;Cross-Model Gate&lt;/strong&gt; — a different model family catches a different error class&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h2&gt;Your Vault&lt;/h2&gt; &#xA;&lt;pre&gt;&lt;code&gt;COG-second-brain/&#xA;├── .claude/skills/          # Claude Code skills (33)&#xA;├── .claude/agents/          # Worker agent definitions (6)&#xA;├── .claude/roles/           # Role packs (7) — personalized recommendations&#xA;├── .agents/                 # Antigravity (agy CLI + IDE) — pointer stubs to .claude/&#xA;├── .kiro/powers/            # Kiro powers&#xA;├── .gemini/commands/        # Gemini CLI commands&#xA;├── AGENTS.md                # Universal agent docs&#xA;├── CLAUDE.md                # Framework instructions&#xA;├── 00-inbox/                # Profiles, interests, integrations&#xA;├── 01-daily/                # Briefs &amp;amp; check-ins&#xA;├── 02-personal/             # Personal braindumps (private)&#xA;├── 03-professional/         # Professional braindumps &amp;amp; strategy&#xA;├── 04-projects/             # Per-project tracking&#xA;├── 05-knowledge/            # Consolidated insights &amp;amp; patterns&#xA;│   └── people/              # People CRM profiles&#xA;└── 06-templates/            # Document templates&#xA;&lt;/code&gt;&lt;/pre&gt; &#xA;&lt;blockquote&gt; &#xA; &lt;p&gt;&lt;strong&gt;Real-world results:&lt;/strong&gt; 120+ braindumps processed, daily briefs with 95%+ source accuracy, 5 major strategic insights discovered — zero maintenance required.&lt;/p&gt; &#xA;&lt;/blockquote&gt; &#xA;&lt;h2&gt;Keeping COG Updated&lt;/h2&gt; &#xA;&lt;p&gt;COG separates &lt;strong&gt;framework files&lt;/strong&gt; (skills, docs, scripts) from &lt;strong&gt;your content&lt;/strong&gt; (braindumps, profiles, notes). Updates never touch your personal data.&lt;/p&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;Method&lt;/th&gt; &#xA;   &lt;th&gt;Command&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;AI Agent (any)&lt;/td&gt; &#xA;   &lt;td&gt;&#34;Update COG&#34; or &lt;code&gt;/update-cog&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Shell script&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;./cog-update.sh&lt;/code&gt; (interactive) • &lt;code&gt;--check&lt;/code&gt; • &lt;code&gt;--dry-run&lt;/code&gt; • &lt;code&gt;--force&lt;/code&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;Manual Git&lt;/td&gt; &#xA;   &lt;td&gt;&lt;code&gt;git fetch cog-upstream main&lt;/code&gt; then checkout specific files&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;p&gt;Check your version: &lt;code&gt;cat COG-VERSION&lt;/code&gt;&lt;br /&gt; Validate packaged surfaces: &lt;code&gt;./scripts/validate-agent-surface.sh&lt;/code&gt;&lt;/p&gt; &#xA;&lt;h2&gt;FAQ&lt;/h2&gt; &#xA;&lt;details&gt;&#xA; &lt;summary&gt;&lt;strong&gt;Why not just use Notion / Roam / Obsidian alone?&lt;/strong&gt;&lt;/summary&gt; &#xA; &lt;p&gt;COG adds self-evolving intelligence on top. It doesn&#39;t just store — it learns, analyzes, and synthesizes insights automatically.&lt;/p&gt; &#xA;&lt;/details&gt; &#xA;&lt;details&gt;&#xA; &lt;summary&gt;&lt;strong&gt;How much does it cost?&lt;/strong&gt;&lt;/summary&gt; &#xA; &lt;p&gt;COG is free and open-source (MIT). You only pay for your AI agent&#39;s API usage.&lt;/p&gt; &#xA;&lt;/details&gt; &#xA;&lt;details&gt;&#xA; &lt;summary&gt;&lt;strong&gt;Is my data private?&lt;/strong&gt;&lt;/summary&gt; &#xA; &lt;p&gt;Yes. Everything is local markdown files. The AI agent&#39;s API is only called when you invoke a skill. No data stored on external servers.&lt;/p&gt; &#xA;&lt;/details&gt; &#xA;&lt;details&gt;&#xA; &lt;summary&gt;&lt;strong&gt;Can I customize or add skills?&lt;/strong&gt;&lt;/summary&gt; &#xA; &lt;p&gt;Yes — edit any &lt;code&gt;SKILL.md&lt;/code&gt; / &lt;code&gt;POWER.md&lt;/code&gt; / &lt;code&gt;AGENTS.md&lt;/code&gt; file. See &lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/SETUP.md&#34;&gt;SETUP.md&lt;/a&gt; for details on creating new skills.&lt;/p&gt; &#xA;&lt;/details&gt; &#xA;&lt;details&gt;&#xA; &lt;summary&gt;&lt;strong&gt;Will updating overwrite my customizations?&lt;/strong&gt;&lt;/summary&gt; &#xA; &lt;p&gt;No. The update process detects customized files and lets you choose per-file: keep yours, use upstream, or backup + update. Nothing is overwritten without approval.&lt;/p&gt; &#xA;&lt;/details&gt; &#xA;&lt;details&gt;&#xA; &lt;summary&gt;&lt;strong&gt;What if I don&#39;t use Git?&lt;/strong&gt;&lt;/summary&gt; &#xA; &lt;p&gt;Git is optional but recommended for version history. COG works fine with just iCloud sync.&lt;/p&gt; &#xA;&lt;/details&gt; &#xA;&lt;h2&gt;Roadmap&lt;/h2&gt; &#xA;&lt;ul class=&#34;task-list&#34;&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_0&#34; checked=&#34;true&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_0&#34;&gt; &lt;s&gt;Gemini CLI + OpenAI Codex support&lt;/s&gt; (shipped in v3.1)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_1&#34; checked=&#34;true&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_1&#34;&gt; &lt;s&gt;Upstream update system&lt;/s&gt; (shipped in v3.2)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_2&#34; checked=&#34;true&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_2&#34;&gt; &lt;s&gt;Role packs &amp;amp; integration discovery&lt;/s&gt; (shipped in v3.3)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_3&#34; checked=&#34;true&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_3&#34;&gt; &lt;s&gt;PM workflow skills &amp;amp; auto-research&lt;/s&gt; (shipped in v3.4)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_4&#34; checked=&#34;true&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_4&#34;&gt; &lt;s&gt;Worker agents, people CRM &amp;amp; specialist sessions&lt;/s&gt; (shipped in v3.5)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_5&#34; checked=&#34;true&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_5&#34;&gt; &lt;s&gt;Loop engineering for iterative skills&lt;/s&gt; (shipped in v3.6)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_6&#34; checked=&#34;true&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_6&#34;&gt; &lt;s&gt;Runtime trust: memory hygiene, post-condition checks, verifier pass, content factory&lt;/s&gt; (shipped in v3.7)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_7&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_7&#34;&gt; Web interface for knowledge graph visualization&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_8&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_8&#34;&gt; Mobile-first commands (optimized for Obsidian mobile)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_9&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_9&#34;&gt; Team collaboration features (with privacy preservation)&lt;/label&gt;&lt;/li&gt; &#xA; &lt;li class=&#34;task-list-item&#34;&gt;&lt;input type=&#34;checkbox&#34; id=&#34;cbx_10&#34; disabled=&#34;true&#34; /&gt;&lt;label for=&#34;cbx_10&#34;&gt; Integration with calendar/task management tools&lt;/label&gt;&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;h2&gt;Contributing &amp;amp; Support&lt;/h2&gt; &#xA;&lt;table&gt; &#xA; &lt;thead&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;th&gt;&lt;/th&gt; &#xA;   &lt;th&gt;&lt;/th&gt; &#xA;   &lt;th&gt;&lt;/th&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/thead&gt; &#xA; &lt;tbody&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/CONTRIBUTING.md&#34;&gt;Contribute&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://github.com/huytieu/COG-second-brain/issues&#34;&gt;Report bugs&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://github.com/huytieu/COG-second-brain/discussions&#34;&gt;Discussions&lt;/a&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA;  &lt;tr&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://github.com/sponsors/huytieu&#34;&gt;Sponsor on GitHub&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://buymeacoffee.com/0xlight&#34;&gt;Buy me a coffee&lt;/a&gt;&lt;/td&gt; &#xA;   &lt;td&gt;&lt;a href=&#34;https://raw.githubusercontent.com/huytieu/COG-second-brain/main/LICENSE&#34;&gt;MIT License&lt;/a&gt;&lt;/td&gt; &#xA;  &lt;/tr&gt; &#xA; &lt;/tbody&gt; &#xA;&lt;/table&gt; &#xA;&lt;h2&gt;Acknowledgments &amp;amp; Inspiration&lt;/h2&gt; &#xA;&lt;p&gt;Built with &lt;a href=&#34;https://claude.ai/code&#34;&gt;Claude Code&lt;/a&gt;, &lt;a href=&#34;https://cursor.com/&#34;&gt;Cursor&lt;/a&gt;, &lt;a href=&#34;https://kiro.dev/&#34;&gt;Kiro&lt;/a&gt;, &lt;a href=&#34;https://github.com/google-gemini/gemini-cli&#34;&gt;Gemini CLI&lt;/a&gt;, &lt;a href=&#34;https://github.com/openai/codex&#34;&gt;OpenAI Codex&lt;/a&gt;, and &lt;a href=&#34;https://obsidian.md/&#34;&gt;Obsidian&lt;/a&gt;.&lt;/p&gt; &#xA;&lt;p&gt;&lt;strong&gt;Key inspirations:&lt;/strong&gt;&lt;/p&gt; &#xA;&lt;ul&gt; &#xA; &lt;li&gt;&lt;a href=&#34;https://github.com/garrytan/gstack&#34;&gt;&lt;strong&gt;Garry Tan&#39;s gstack&lt;/strong&gt;&lt;/a&gt; — specialist sessions, clear operating gears, repo-local skill distribution. COG&#39;s worker agent architecture and model routing borrow directly from gstack&#39;s explicit mode separation.&lt;/li&gt; &#xA; &lt;li&gt;&lt;a href=&#34;https://github.com/garrytan/gbrain&#34;&gt;&lt;strong&gt;Garry Tan&#39;s gbrain&lt;/strong&gt;&lt;/a&gt; — Compiled Truth + Timeline pattern, tiered enrichment for people profiles, brain-first lookup protocol. COG&#39;s people CRM and knowledge-first approach are adapted from gbrain&#39;s design.&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Zettelkasten&lt;/strong&gt; — atomic, interlinked notes as the foundation of knowledge&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;Building a Second Brain (Tiago Forte)&lt;/strong&gt; — PARA organization, progressive summarization&lt;/li&gt; &#xA; &lt;li&gt;&lt;strong&gt;GTD (David Allen)&lt;/strong&gt; — capture everything, process systematically&lt;/li&gt; &#xA;&lt;/ul&gt; &#xA;&lt;hr /&gt; &#xA;&lt;p&gt;&lt;strong&gt;TL;DR:&lt;/strong&gt; Clone, run onboarding, braindump daily. COG evolves with you — just &lt;code&gt;.md&lt;/code&gt; files, any AI agent, zero maintenance.&lt;/p&gt;</summary>
  </entry>
</feed>