Create an AGENTS.md File

Write repo-root instructions so coding agents know the product, conventions, tools, and what needs approval.

Open in ChatGPT
Framer Agent
Sponsored

Best AI site builder for designers, PMs, and founders.

Try Framer free

Prompt

You write AGENTS.md files for coding agents. Create one for this project.

## Project context

- Product or project name: [name]
- What the product does: [description]
- Target users: [users]
- Product stage: [idea, MVP, growth, enterprise, etc.]
- Tech stack and tools: [stack/tools]
- Repository or workspace structure: [structure]
- Important files, folders, or documentation: [references]

## Team conventions

- Code, design, content, or documentation conventions: [conventions]
- Naming conventions: [rules]
- Design-system requirements: [rules]
- Testing and QA expectations: [expectations]
- Accessibility requirements: [requirements]
- Security, privacy, or compliance requirements: [requirements]

## Agent boundaries

- Actions the agent may take automatically: [actions]
- Actions that require approval: [actions]
- Actions the agent must never take: [actions]
- Available tools and integrations: [tools]
- Definition of done: [definition]

Write a short AGENTS.md with these sections:

1. Project purpose and target users
2. Workspace map
3. Source-of-truth hierarchy
4. Required conventions
5. How to work before making a change
6. Design and UX requirements
7. Tool-use rules
8. Approval and permission boundaries
9. Testing and verification requirements
10. Definition of done
11. Expected handoff or status-report format
12. Prohibited actions
13. Escalation rules for missing context, ambiguity, or risk

Use instructions you can test. Mark unknowns as [NEEDS TEAM INPUT]. Do not invent details.

How to use

  1. 1Fill product, stack, repo map, and what the agent may vs must not do.
  2. 2Generate AGENTS.md, put it at the project root, and link DESIGN.md or docs as sources of truth.
  3. 3Run one real agent task, then tighten any section the agent ignored.

Tips

  • Prefer AGENTS.md when more than one tool reads the repo. Use CLAUDE.md when the team lives in Claude Code.
  • Put approvals next to prohibited actions so they are hard to miss.
  • Point at DESIGN.md for UI rules instead of pasting the whole design system here.

New prompts & templates by email

Weekly copy-paste prompts, pattern notes, and AI UX resources on Substack - no spam, unsubscribe anytime.

Subscribe on Substack