Don't write any code yet. Interview me, then draft PROJECT-BRIEF.md for [what you're building]. Sections: Goal (one sentence), Users ([who uses it]), Must have, Not in this version, Done means ([done means]), Open questions. Ask me up to five questions first if anything is unclear. Don't fill in requirements I didn't give you; put them under Open questions. Show me the draft before saving.
Write the project brief before the agent writes code
You're about to ask a coding agent to build something, and you want it to know the goal, the limits, and what 'done' means before it touches a file.
- Claude Code
- Grok Build
For: Owner, Tech person · Updated October 25, 2026
A starting point: fill in the [brackets] with your own details, and check the answer before you use it. Tool screens change often, so yours may look a little different.
Video walkthrough
Video coming later: Four AI Coding Agents, One Bakery Website: Setup and What Each Costs (P3-01). Until then, watch the free episodes that are out now.
The prompts
Copy the version for the tool you use, then replace every [bracket].
Now draft AGENTS.md for this project: what it is (link @PROJECT-BRIEF.md), how to run and test it, folder layout, conventions, and a "Never do" list: [never do], never commit secrets, never deploy. Keep it under 60 lines. Then create a CLAUDE.md that contains only the line @AGENTS.md so Claude Code and other agents share one file. Show the diff before writing./plan Read PROJECT-BRIEF.md and AGENTS.md. Before building anything, tell me in five bullets what you understand the job to be, what you won't do, and how you'll prove it's done. List any conflicts between the two files.
Fill in these [brackets]
| Bracket | What to put there |
|---|---|
[what you're building] | e.g. a one-page catering order calculator for our restaurant website |
[who uses it] | e.g. people ordering office lunches or parties for 10 to 150 |
[done means] | Checks you can see with your own eyes. e.g. pick items and a headcount, see an estimated total, send the request by email; works on a phone |
[never do] | e.g. touch the live site, add a payment form, install packages without asking |
Why this works: Claude Code loads CLAUDE.md at the start of every session (and can read AGENTS.md), and Grok Build loads AGENTS.md and CLAUDE.md from the repo root down. Writing the brief once means you stop re-explaining the job, and the agent has something concrete to check its work against. Docs checked Oct 17, 2026.
Real example (Claude, free plan, tested Oct 25, 2026, step 1 run in a web chat as a stand-in for Claude Code):
- Fulfillment: Do you offer delivery, pickup, or both? Is there a delivery area, a minimum order, or a lead time (e.g., 48 hours’ notice) the form should enforce or mention?
…
Let customers build a catering order on one page of the Lantern Row Kitchen website, see an estimated total, and send the request to us by email.
…
- A running estimated total, labeled “plus tax”. Tax is not calculated.
- Pickup-only messaging (no delivery option shown).
…
- Sending email from a static site: Do we use a “mailto:” link (opens the customer’s email app) or a form service that emails us directly? This affects reliability on phones.
Answer given with our fictional inputs (Lantern Row Kitchen, a made-up restaurant). Excerpt trimmed; “…” marks cuts. AI answers vary, so yours will be different. Check it before you use it.
Check before you send
- 'Done means' items are things you can test without reading code
- The 'Never do' list covers your real risks (live site, money, customer data)
- No passwords or API keys anywhere in the files
- Run grok inspect or /context (Claude Code) to confirm the files loaded
Privacy note: what not to paste
Brief files get committed and shared. Put secrets in environment variables, never in PROJECT-BRIEF.md, AGENTS.md, or CLAUDE.md.