Open an unfamiliar codebase in a coding agent and the first few prompts are rarely about the feature you actually want. They're spent on reconnaissance: where does the data live, what's the token system called, is this a server component by default, which file would a human on this team touch to add a page. A generic template leaves the agent to guess at all of it, and a wrong guess costs you a revert.
What ships in every template
- CLAUDE.md - read by Claude Code on open. Commands, routes, where things live, the design-token system, conventions, how to add a page, and what not to touch.
- AGENTS.md - the same instructions in the shared format a growing list of agents read, OpenAI Codex among them.
- .cursorrules - the same instructions again, as the rules file Cursor loads for the project.
All three carry identical content. They're generated from one shared structure per template, so the conventions stay consistent across the whole catalog rather than drifting from one template to the next depending on who wrote the docs that week. Open the project in whichever agent you use, and it loads the file it already knows how to read. Nothing to paste in, nothing to remember to attach.
What's actually in the file
Not a generic AI disclaimer. Here's a trimmed, real excerpt from Stockwell's CLAUDE.md, our largest storefront template:
## Commands
- `npm run dev` - local dev server
- `npm run build` - production build (run before finishing any change)
- `npm run typecheck` - TypeScript check
## Where things live
- Routes: /, /account, /c/[category], /cart, /checkout, /deals, /p/[slug], /shop, /wishlist
- Data: The whole catalogue is in src/lib/catalog.ts - categories, products, reviews, Q&A, sample orders.
- Design tokens (colours, fonts, radii): src/app/globals.css
## Payments and AI
The shopping assistant is offline and rule-based. src/lib/assistant.ts matches keywords
against the catalogue and returns a reply with product picks. It uses no model and needs
no API key.
That last section is the one we'd flag. An agent told the wrong thing about what's actually running in a template will confidently build on a false assumption, like wiring a real API key into a chat feature that's actually a keyword matcher. So the file says so, plainly, instead of staying quiet about it.
The honest AI-status table
That same honesty carries over to /ai-ready, a page that lists every template's actual AI status next to its agent files - which ones run a real model, which run a rule-based or canned-demo assistant, and which have no AI features at all. It exists because a few templates in the catalog do ship mock, rule-based, or demo-corpus AI for the preview experience, and a buyer's coding agent deserves to be told that directly rather than find out by reading the source after the fact.
A prompt cookbook, too
Each template also gets its own cookbook page at /templates/[slug]/ai-prompts - three ready-made prompts for the customizations people actually make first: changing the brand colour, adding a page, wiring up a real backend. Each prompt names the exact file it changes, so it's copy-paste, not guesswork.
The result: open a template in Claude Code, Cursor, or Codex, and it already knows the answers to the first five questions it would otherwise have had to ask you.
See every template's agent files and AI status, or browse the full catalog.
Top comments (0)