A field note from the autonomous Claude Code agent I run every day on one Windows PC. The numbers come from its own ledgers, not from memory.
Free to read first: a PDF sample of the book behind this setup — its first 3 chapters: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code-free-sample
In the last note I found Korean text left inside the code blocks of my translated book, and I added a check that refuses to upload a PDF with too much of it. That check went into one store first. A few hours later I looked at the second store that sells the same editions.
The second store had no check at all
Its replace step had no Korean count. It also deletes the old file before it uploads the new one. So a wrong file would not sit next to the right one; it would take its place.
I added the same count, with the same limit, read from the same constant, so the two stores can't drift apart. Then I ran it in read-only mode over the builds that were on sale. All eight were refused, with 2,734 to 4,201 Korean characters each.
Three editions had cleaned builds ready. I uploaded those, downloaded each one back from the store, and compared it with the file I had sent. All three matched. The other five stay as they are until translation work resumes; the check refuses them, which is what it's for.
What I counted, and what I didn't
I had measured the eight language editions, because that's what the code loops over: a list of eight language codes.
The store's ledger lists ten products. One is the Korean original's free edition, which is Korean on purpose. The other is an English free sample, thirteen pages long, offered as the way into the paid English edition.
The independent review I run before calling anything done asked about that one. The sample PDF in my build folder, built on October 18, had 446 Korean characters. The first file a new English reader would open was the one I hadn't checked.
I rebuilt the sample after the cleanup: zero Korean characters, still 13 pages. I uploaded it, downloaded it back, and it matched.
The part I still can't answer
Before replacing anything, I archive the old file. For the sample, the archived copy turned out to be a different size from the file the store listed. It came from my build folder, not from the store. So the 446 belongs to my October 18 build, and I never measured the file that had actually been for sale. It may have been similar. I don't know.
The rules
- Count products, not code paths. My loop had eight entries; the ledger had ten products. The ledger is what a customer sees.
- When you leave something out of a check, write down why. The Korean free edition is excluded on purpose. The sample was excluded by accident, and nothing recorded the difference.
- Archive from where the thing is sold, not from where it was built. Otherwise "before" is a copy of your intentions.
Where this comes from. Every post here comes from one setup I run daily: a CLAUDE.md, memory files the agent reads before it touches anything, and a separate auditor agent that returns PASS or FAIL. The first 3 chapters of the book that walks through it are free as a PDF: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code-free-sample
The full edition is 11 chapters plus 4 ready-to-use templates (CLAUDE.md starter, memory files, auditor checklist, measurement guide) and a hands-on section for every chapter, $19 as a PDF: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code
Questions about the setup are welcome in the comments — I'll answer with what actually happened, not theory.
Top comments (0)