Coding Agents · Power user

Write an AGENTS.md / CLAUDE.md for your project

You're starting with a coding agent and want it to understand your project and its limits before it touches anything.

  • Grok Build
  • Claude Code

For: Tech person, Owner · 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.

Prompt Park Sessions · Three-Minute AI for Small Business

Video walkthrough

Video coming later: Write It Down Once: One AGENTS.md That All Four Coding Agents Read (P3-02). Until then, watch the free episodes that are out now.

Also covered in: P3-06 · CLAUDE.md vs AGENTS.md: Which File Should You Write? (Part 3: Coding Agents Compared) · coming later

The prompts

Copy the version for the tool you use, then replace every [bracket].

GROK BUILD
Run from your project folder in plan mode so you approve the file before it's written.
Read this project and propose an AGENTS.md. Include: what the project is, how to run and test it, folder layout, coding conventions you observe, and a "Never do" list: never commit secrets, never delete files outside [folder], never deploy, ask before installing packages. Only state what you saw in the files; mark guesses as "to confirm". Show me the file as a plan first. Don't write it until I approve.
CLAUDE CODE
Claude Code reads CLAUDE.md as project memory (and can read AGENTS.md).
Draft a CLAUDE.md for this repo: purpose, build and test commands, structure, conventions, and a safety section (no secrets, no deploys, no force-push, confirm before deleting anything). Only state what you saw in the files; mark guesses as "to confirm". Keep it under 60 lines. Show the diff before saving.

Fill in these [brackets]

BracketWhat to put there
[folder]Where the agent may delete files, if anywhere. e.g. src/pages/ and data/

Real example (ChatGPT, free plan, tested Oct 25, 2026, given a project snapshot in a web chat as a stand-in for a coding agent):

5. Never do

  • Never commit secrets. Do not commit API keys, passwords, tokens, credentials, or real secret values. Keep .env files and sensitive configuration out of version control.
  • Never delete files outside src/pages/ and data/. …
  • Never deploy. Do not upload dist, publish the site, or trigger a production deployment. Maya handles deployment manually.
  • Ask before installing packages. Obtain explicit approval before installing new packages or making dependency changes.

…

What I’d verify before finalizing

The snapshot gives enough information to draft the file, but not to confirm every implementation detail. I would avoid claiming that a particular CSS pattern, link-checking behavior, or build configuration has been verified beyond what you supplied.

Answer given with our fictional inputs (a made-up restaurant website project). Excerpt trimmed; “…” marks cuts. AI answers vary, so yours will be different. Check it before you use it.

Watch for: In the same test, Claude stated “There is no automated test suite” as fact from a short snapshot. The prompts above now ask it to mark guesses as “to confirm”.

Check before you send

  • Build and test commands actually run
  • The "Never do" list matches your real risks
  • No passwords, API keys, or customer data in the file

Privacy note: what not to paste

Never paste API keys or passwords into a prompt or into AGENTS.md. Use environment variables.