| name | agents-md-author |
| description | Generate or maintain a minimal high-signal AGENTS.md for coding agents. Use whenever a repo needs agent instructions, an existing AGENTS.md is too long or vague, or project docs/scripts need a concise routing map for Codex. |
AGENTS.md Author
Purpose
Create a short AGENTS.md that routes agents to authoritative docs and commands without becoming a giant project knowledge dump.
Use nested-agents-md-author after the root file is healthy when apps, packages, or docs need scoped local instructions.
Inspect First
- existing
AGENTS.md or other agent instruction files
docs/00-index.md, README, architecture docs, testing docs, and command docs
- manifests and scripts to confirm standard commands
Procedure
-
Keep AGENTS.md short.
- Target under 120 lines.
- Put detailed knowledge in
docs/.
- Link to docs instead of repeating them.
-
Include only stable operating contract.
- Read-first docs.
- Standard commands.
- Non-negotiable repo rules.
- Architecture and quality gate pointers.
- Validation expectations.
-
Preserve local policy.
- If existing instructions contain project-specific constraints, keep them or move them to the right doc and link back.
- Remove stale, duplicated, or framework-generic prose.
-
Update related docs.
- If
AGENTS.md references a missing doc, create or update that doc through repo-knowledge-base-builder.
- If commands changed, update
docs/development/commands.md.
- If directory-specific rules are needed, keep root
AGENTS.md as the map and create scoped files through nested-agents-md-author.
- If control-loop artifacts exist, link to feature state, progress, decisions, startup readiness, definition of done, and session handoff docs.
- If instructions are long or duplicated, run or recommend
check_instruction_snr.py.
- Use
project-local-skill-generator when project-specific workflows belong in .codex/skills instead of root AGENTS.md.
Required Shape
Use this structure unless the repo already has a stronger convention:
# Agent Instructions
This file is a map. The authoritative project knowledge is in `docs/`.
## Read first
- `docs/00-index.md`
- `docs/architecture/overview.md`
- `docs/architecture/boundaries.md`
- `docs/development/commands.md`
- `docs/testing/strategy.md`
- `docs/quality/code-style.md`
## Standard commands
- Dev: `./scripts/dev`
- Lint: `./scripts/lint`
- Typecheck: `./scripts/typecheck`
- Test: `./scripts/test`
- Build: `./scripts/build`
- Full validation: `./scripts/validate`
## Rules
- Inspect existing code and docs before changing behavior.
- Prefer established patterns over new abstractions.
- Add or update tests for behavior changes.
- Update docs when commands, environment variables, routes, schemas, or architecture change.
- Run the narrowest relevant check first, then `./scripts/validate`.
- If validation cannot run, state the exact command, error, and likely cause.
## Architecture boundaries
See `docs/architecture/boundaries.md`.
Validation
- Check every linked doc exists.
- Check every listed command exists or is documented as intentionally absent.
- Check the file is concise and not duplicating architecture, product, or testing docs.
- Check root instructions route to state and handoff files when the repo uses them.
Completion Criteria
AGENTS.md is a routing file, not a manual.
- It contains no stale commands or missing links.
- It preserves project-specific constraints in the right location.