| name | agents-md-authoring |
| description | Write AGENTS.md files that serve as operating contracts for AI agents. Use when creating or updating AGENTS.md for a project. Trigger terms: AGENTS.md, agent instructions, agent contract, agent configuration, coding agent setup. |
AGENTS.md Authoring
AGENTS.md is the entry point for AI agents working in a codebase. It is not a README — it is a contract between the author and machines-with-judgment. Its job is to encode decisions an agent should not have to rediscover.
Core Principles
-
Agents are executors, not readers. Every line competes for context budget. Omit what any competent agent would do by default; state only what is specific to this project.
-
Encode done, not just started. Verification criteria are the most important content. If the agent cannot tell whether its work is complete, it will stop too early or keep going forever.
-
Actionable over aspirational. "Write clean code" is noise. "No any types; run npm run check after every code change; fix all errors before committing" is a contract.
-
Progressive disclosure. Inline what's frequently needed; reference what's rarely needed. Link to .docs/ files with a one-line description of when to read them.
Recommended Structure
# AGENTS.md
## Project
<!-- 2-4 sentences: what it does, primary focus, non-obvious terminology -->
## Architecture
<!-- Key directories, package boundaries, how pieces connect -->
## Commands
<!-- Exact commands. Not "run the tests" — the actual invocation. -->
<!-- Include: what to run after code changes, what is FORBIDDEN -->
## Code Standards
<!-- Language-specific rules that deviate from defaults -->
<!-- Only rules that are enforced — aspirational conventions erode trust -->
## Workflow
<!-- When to ask vs proceed, commit/push rules, session completion -->
## Testing
<!-- Framework, how to run specific tests, what to do after modifying tests -->
## Verifying Work
<!-- What "done" means. Exact commands, expected output, what failure looks like. -->
## References
<!-- Links to .docs/ files with WHEN to read them -->
Size Guidelines
Effective range: 100-300 lines. The 530-line ceiling works only for large monorepos where every line is load-bearing.
Heuristic: If the agent needs it on any non-trivial task → inline. If only for a specific task class → link to .docs/. If the agent wouldn't need it → delete.
Audience Separation
| File | Audience | Content |
|---|
| README.md | Humans | Project purpose, setup, usage |
| AGENTS.md | AI agents | Operating contract, commands, verification |
Never duplicate between them. Content written for one should not appear in the other.
Anti-Patterns
- Too long — 500 lines of background context the agent rarely needs. Front-load high-frequency content.
- Too vague — "Follow good engineering practices" produces vague behavior.
- Duplicating README — wastes context and creates maintenance fork.
- No verification criteria — agent cannot tell if output is acceptable.
- Aspirational conventions not enforced by tooling — teaches agent to claim compliance without evidence.
- Stale autonomy directives — blanket "never ask" left from a different task type.
Sources
- best_practices/docs/practices/agents-md-authoring.md (internal)
- Observed across 10+ real AGENTS.md files: 84-530 lines effective range
