agents-md-author git:20260430.abe6505A

Immutable. This exact content never changes and is served at /api/v1/blob/e66870aeb9809486.

---
name: agents-md-author
description: "Bootstrap or update project AI-agent instruction files (AGENTS.md for Codex, CLAUDE.md for Claude Code, or both) from the current project's structure. Triggers on: write AGENTS.md, write CLAUDE.md, generate AGENTS.md, agent instructions, project conventions for AI, bootstrap agent docs, set up project memory for Codex/Claude, document dev/test/build commands for AI coding tools, single-source project instructions for any AI agent."
---

# agents-md-author

Generate cross-vendor project instruction files. The Anthropic Agent Skills open spec (Dec 2025) standardized SKILL.md across 32+ tools; this skill applies the same logic at the project-instructions layer where Codex reads `AGENTS.md` and Claude Code reads `CLAUDE.md`.

## When to activate

- User asks to "write AGENTS.md", "create CLAUDE.md", or "set up project instructions for AI"
- User wants the same project doc to work in Codex, Claude Code, Cursor, or any tool that follows the AGENTS.md convention
- User mentions onboarding a new AI tool to an existing repo and wants conventions captured

## Procedure

### 1. Detect project context (always do this first)

Read the project root and determine:
- **Project name** — from `package.json` `name`, `pyproject.toml` `[project].name`, `go.mod` `module`, `Cargo.toml` `[package].name`, or fallback to repo dir name
- **Language & runtime** — `package.json` (Node + lockfile flavor), `requirements.txt` / `pyproject.toml` (Python), `go.mod` (Go), `Cargo.toml` (Rust), `pom.xml` / `build.gradle` (JVM), `*.csproj` (.NET)
- **Common scripts** — Node `scripts.{dev,build,test,lint,typecheck}`; Python `tox.ini` / `Makefile`; Go/Rust task tools
- **Test framework** — vitest, jest, pytest, go test, cargo test, etc.
- **Repo conventions** — existing `.editorconfig`, `.prettierrc`, `eslint.config.*`, `pre-commit-config.yaml`

If the user provided context (e.g., "this is a Next.js + Prisma app"), use it; otherwise infer from files. Don't ask questions you can answer from disk.

### 2. Decide the target file(s)

| User wants | Write |
|---|---|
| Codex / OpenAI tools only | `AGENTS.md` |
| Claude Code only | `CLAUDE.md` |
| Both, kept in sync | `AGENTS.md` + `CLAUDE.md` (identical content), or `AGENTS.md` plus a one-line `CLAUDE.md` that says "See AGENTS.md" |
| Single source (recommended) | `AGENTS.md` + `CLAUDE.md` symlink → `AGENTS.md` |

Default to "both, identical content" unless the user specifies — it's the lowest-risk option and works everywhere.

### 3. Compose content

Use this canonical structure (load `references/agents-md-spec.md` if you need the field-by-field rationale or cross-vendor details):

```markdown
# <Project name>

<one-sentence description>

## Tech stack

- <language + version>
- <framework + version>
- <database / runtime / deployment>

## Dev commands

- Install: `<command>`
- Dev server: `<command>`
- Build: `<command>`
- Test: `<command>` (unit) / `<command>` (e2e)
- Lint / typecheck: `<command>`

## Code conventions

- <indent + line length>
- <import style>
- <naming conventions>
- <test colocation>
- <commit style>

## Architecture notes

<2-5 bullets on layout — where routes live, where business logic lives, what's shared>

## Out of scope for AI agents

<deliberate "don't touch" list — generated files, secrets, migrations, etc.>
```

Keep it under ~150 lines. AI agents skim — verbose docs are skipped or ignored.

### 4. Write files

- Use canonical line endings for the OS (LF on macOS/Linux)
- Don't add a trailing summary or "generated by" footer — the file is now the user's
- If a file already exists, **read it first**, propose a diff, and ask the user before overwriting

### 5. Verify

- Print the relative paths written
- If both `AGENTS.md` and `CLAUDE.md` were written, confirm they're byte-identical (or show the symlink)
- Suggest the user commit the files

## Anti-patterns

- Inventing test commands the project doesn't have — read `package.json.scripts` literally
- Writing a 500-line wall — agents skip it. Aim for ~100 lines.
- Embedding secrets, API keys, or environment-specific paths — those go in `.env`, not in agent docs
- Forking content between AGENTS.md and CLAUDE.md — they should be byte-identical or symlinked

## References

- `references/agents-md-spec.md` — full cross-vendor convention summary (Anthropic, OpenAI, Cursor, etc.) — load this on demand if you need the spec details
- Anthropic Agent Skills spec: https://github.com/anthropics/skills
- OpenAI Codex AGENTS.md guide: https://developers.openai.com/codex/guides/agents-md
- agents.md cross-vendor standard: https://agents.md