I Took Shopify’s AI-Agent Prompt Somewhere It Wasn’t Meant to Go

Shopify recently published a guide to starting and running a store with AI agents. The pitch: your first step can be a conversation with an AI.

OK. Fine. But Shopify is easy mode. I mean, everything is pre-built, themed, and already wired together.

So I took the one complete prompt in their article and gave it to an AI agent working with Crystallize, a headless PIM and commerce platform. "Headless" usually comes with a warning label: you'll need a developer.

The thing is, I am not that developer. My last real code was changing PHP in WordPress and Shopify templates a few years ago. Some React and Next. js knowledge and a ton of assisted AI help.

I am not picking on Shopify. I wanted to check whether someone can, like me, get from Shopify's prompt to a working store if an agent does the heavy lifting? How far has the AI thing gone?

Disclosure: I work at Crystallize. In marketing, not engineering. Which is kind of the point.

What I used
• Shopify's own prompt, the oversized organic t-shirt one, word for word (shared in this article)
• Crystallize: headless PIM and commerce, starting from an empty tenant
• Crystallize MCP server and agent skills: how the AI learns the platform instead of guessing
• Claude for planning, Claude Code in VS Code for doing
• Mass operations: one JSON file that creates the whole catalogue in one go
• Catalogue API (GraphQL): what the storefront reads
• Next.js (App Router), running locally on localhost:3000
• Windows, VS Code, and a terminal I am still slightly scared of

Step by step

1. Get an empty tenant and an access token

Create a tenant, then create an access token in your account settings for a user with read and write access. Follow the docs; it is boring, in a good way.

2. Connect the agent

Crystallize runs a hosted MCP server. One command registers it with Claude Code, Cursor, Claude Desktop, and friends:

npx add-mcp https://mcp.crystallize.com/mcp \
  --header "X-Crystallize-Access-Token-Id: YOUR_TOKEN_ID" \
  --header "X-Crystallize-Access-Token-Secret: YOUR_TOKEN_SECRET"

It also ships skills: docs that teach the agent how shapes, prices, and imports work. Without them, the agent improvises, and you do not want an improvising agent near your product data.

3. Paste Shopify's prompt

Here is the prompt from Shopify's article, unchanged.

This is where it got interesting. The agent did not just make three "collections". It worked out what a heavyweight tee actually is and turned it into data: fabric weight as a number in gsm, composition, certifications from a fixed list (GOTS, OCS, OEKO-TEX), and a taxonomy for fabric, weight, fit, cut, and feel. "Heavyweight" stopped being a label someone maintains by hand. It became "240 gsm and up".

4. Let it write one import file, then run it

The agent wrote a single mass operation file: shapes, topics, folders, 8 products, 93 size and color variants, prices in EUR and USD. 60 operations. 60 succeeded, in about 12 seconds.

Then I opened the Crystallize app and checked with my own eyes, because of trust issues🤪

5. Build the storefront

New empty folder in VS Code, same prompt, the storefront half. What I wrapped around it boiled down to:

Build a storefront for this catalogue from Shopify's prompt.
Next.js App Router, run locally only. Read-only access to Crystallize.
Collections, plus filters by fabric, weight, fit, cut, and feel.
Product pages show gsm, composition, and certifications.
No photos exist: use placeholders. Log every step.

About 15 minutes later, I had a home page, "Shop all", three collections, filters with their own links, and product pages with EUR and USD prices. Under the hood, it is plain GraphQL against the Catalogue API:

{
  catalogue(path: "/collections/heavyweight", language: "en") {
    name
    children { name path }
  }
}

6. Run it

npm run dev

Open localhost:3000. Click everything. It worked. I was mildly offended.

The asterisk

• No cart or checkout. "Add to bag" is intentionally switched off. Payments are their own adventure, no matter what platform you run your business on.
• No photos, no SEO titles. Grey boxes that say "Photo to come". Very minimalist. Very intentional. If I'd taken more time, I could have made fake AI photos. But I was more interested in what they had in mind. Still, you can use this for a small t-shirt business.
• One size and colour per tee, for now. Products created through the newer import path came back without a variant list in the public feed. This test helped our engineers find a bug that had already been fixed, but it wasn't deployed when I wrote this.
• Filtering happens in the app. Fine for 8 tees. A real catalogue would use Crystallize's Discovery API instead.
• A terminal was involved. Twice. I survived.

Verdict

Claim holds, and not only on Shopify. That's the slightly crazy part. A few years ago this was a sprint and an agency invoice.

Today one prompt gets a structured catalogue and a working Next.js storefront on a headless backend, built for someone whose comfort zone is a WordPress theme file.

Which backend, which frontend? It matters less every month. The agent will figure out either.

What still matters is what always mattered: the product. An agent can build the shelf in 12 seconds. It can't make the T-shirt worth buying, write the fabric details you never collected, or answer the customer whose order went missing. A mediocre product and mediocre service now get a mediocre store, faster.

And this is the point of this post (in a way). The upside of AI: people with brilliant ideas no longer need gatekeepers.
The downside: people with garbage ideas don’t have them either.

Now look at what happens when actual engineers get involved

My store took an afternoon, one prompt, and a healthy amount of luck. The superfast.shop demos are what you get when people who build this for a living use the same platform: real photos, proper data models, and pages that load before you finish blinking. Like obe of these👇
[Fashion Universe] :: [Block Builder Universe] :: [3D Kitchen Universe]

That gap is the honest part of this experiment. My agent got me from zero to "huh, that actually works". Engineers get you from "that works" to "that sells, at scale, on Black Friday". They know where a data model breaks at 50,000 products, what to cache, what checkout should do when a payment fails, and which clever shortcut will cost you six months later. No prompt I wrote knew any of that.

The best setup isn't agents instead of engineers. It's engineers designing the foundation and agents doing the repetitive heavy lifting on top of it.

Shameless bit, since I work there: we package that engineering experience into ready-made prompts for setting up Crystallize, tailored per industry, so your agent starts from a data model someone experienced designed, not from my t-shirt experiments.