agentsmd-generator（AgentMD 生成器）

Agent Context Generator

What You'll Do

• 🔍 Inventory the repository's structure, capture a .gitignore-aware tree output, and record automation entry points (preferring just/make tasks when available)
• 🧭 Capture coding conventions, directory ownership, testing expectations, and review workflows so future agents can navigate confidently
• 🧩 Produce an AGENTS.md file following the opinionated section order below, honoring scope rules for nested directories
• ✅ Embed universal wrap-up tasks: ensure the README is updated after significant code changes and summarize changes per conventional commits while resolving any open questions with the developer

Phase 1 · Understand the Repository

1. Check for existing AGENTS.md

- Use find alternative (glob or repo tree) to discover current files. Determine scope inheritance so you can update or extend instead of duplicating.

1. Read Core Docs

- Skim README.md, CONTRIBUTING.md, and other onboarding docs for project philosophy, setup, and workflows. - If docs/ or documentation/ exists, scan for architectural or process references worth surfacing.

1. Survey Project Layout

- Note primary directories, languages, build targets, and ownership (e.g., "src/ui maintained by Frontend team"). - Check for plans/, docs/, or other knowledge directories. Flag must-read files (ADR indexes, architecture overviews, runbooks) to reference later in AGENTS.md.

1. Build a Git-aware Tree

- Use the tree command with the --gitignore flag (tree ≥ 2.0) so ignored paths stay hidden: tree --gitignore -a -L 3 > tmp/tree.txt. - If your tree build lacks --gitignore, run tree -a -L 3 --prune and manually prune any ignored directories noted in .gitignore, or install an updated version via your package manager. - Capture or trim the output before placing it in AGENTS.md (focus on the top 2–3 levels, and note when you omitted details for brevity).

1. Identify Automation Runners

- If Justfile exists, run just --list (or just --list --unsorted for extra notes). - If Makefile exists (and just does not), run make help or inspect phony targets for canonical tasks. - Record which commands are recommended for linting, testing, building, syncing data, etc. Link the definitive task names you surface in your notes for inclusion later.

1. Catalog Tooling & Environment

- List required runtimes, package managers, env vars, secrets handling, and local services. - Note down any .env.example, config/, or secrets documentation that agents must review.

1. Clarify Testing & Quality Gates

- Identify test suites, coverage expectations, linting, formatting, and CI workflows.

1. Resolve Ambiguities Early

- Whenever conventions, ownership, or workflows seem unclear, prompt the developer with focused questions before drafting the guide. - Ask explicitly whether existing plans/ or documentation directories are authoritative or stale, and clarify what canon to reference.

Outcome: A structured notes list describing layout, tooling, commands, testing, release process, documentation references, pending questions, and update expectations.

Phase 2 · Plan the AGENTS.md Structure

Follow this opinionated order to keep files consistent and scannable:

1. Header — Title + short purpose statement.
2. Quick Facts — Table or bullet summary (languages, package manager, key scripts, CI).
3. Repository Tour — High-level directory map with responsibilities and ownership hints.
4. Tooling & Setup — Required runtimes, package managers, environment variables, secrets.
5. Common Tasks — Lint/test/build/deploy commands. Prefer listing just recipes first, then make targets, then raw commands.
6. Testing & Quality — When and how to run tests, linting, formatting, coverage, and CI expectations.
7. Workflow Expectations — Branching model, review norms, feature flagging, deployment cadence.
8. Documentation Duties — When to update README.md, architecture diagrams, or other docs.
9. Finish the Task — Mandatory wrap-up checklist for every agent task.

For deeper directories (e.g., services/api/), include a "Scope" note at the top clarifying inheritance from parent AGENTS instructions. Always confirm with the developer before drafting new per-directory AGENTS files so you do not duplicate existing guidance or create unnecessary overhead.

Phase 3 · Compose AGENTS.md

Use the template below and adapt each section to the project:

# Project Agent Guide

> Scope: Root project (applies to all subdirectories unless overridden)

## Quick Facts
- **Primary language:**
- **Package manager:**
- **Entrypoints:**
- **CI/CD:**

## Repository Tour
- `path/` — description & owner

## Tooling & Setup
- Install instructions (per OS)
- Required environment variables (with purpose)
- Secrets management notes

## Common Tasks
- `just <task>` — what it does (preferred)
- `make <target>` — what it does
- Raw command fallback when automation missing

## Testing & Quality Gates
- Unit/integration test commands
- Lint/format commands
- Coverage expectations & thresholds
- CI status command or dashboard link

## Workflow Expectations
- Branch naming and review rules
- Feature toggles or release cadence
- Any approval or ticket linkage requirements

## Documentation Duties
- Update `README.md` when features, setup steps, or developer ergonomics change materially
- List other docs to refresh (architecture, ADRs, etc.)

## Finish the Task Checklist
- [ ] Update relevant docs (& `README.md` if significant changes landed)
- [ ] Summarize changes in conventional commit format (e.g., `feat: ...`, `fix: ...`)

Subdirectory Template (Use Only with Developer Approval)

# <Directory Name> Agent Guide

> Scope: ./path/to/directory (inherits root AGENTS.md unless noted)

## Purpose
- What lives here
- Who owns it (team/contact)

## Key Files
- `file_or_folder/` — why it matters

## Common Tasks
- `just <task>` / `make <target>` / command snippets scoped to this directory

## Testing & Quality
- Specific tests, linters, or data fixtures for this directory

## Hand-off Notes
- Docs or runbooks to reference
- Open questions captured during discovery

Only create these per-directory guides after confirming with the developer which areas need dedicated context and what information should be emphasized.

Writing Notes:

• Keep language direct and actionable. Agents should follow commands verbatim.
• Mention the preferred order of operations (e.g., "Always run just format before opening a PR").
• When referencing scripts, include relative paths so agents can jump quickly (e.g., scripts/bootstrap.sh).
• Incorporate a trimmed tree --gitignore snapshot (or link to the saved artifact) so readers grasp layout quickly.
• In the Repository Tour, highlight where plans/, docs/, design docs, or ADRs live if present.
• Call out any unanswered questions as action items, and confirm with the developer before creating any per-directory AGENTS overlays.
• If the project mixes languages/platforms, add subsections per component but keep global guidance first.

Phase 4 · Validate & Wrap Up

1. Self-review

- Does the file respect AGENTS scope rules? (Mention inheritance or overrides.) - Are all critical commands documented, especially automation entry points? - Is the README update expectation explicit? - Did you obtain developer approval before adding any per-directory AGENTS files, and is that approval reflected in the write-up? - Does the "Finish the Task" checklist include the conventional commit summary reminder?

1. Formatting

- Ensure headings use Title Case, commands are wrapped in backticks, and lists are concise. - Keep sections under ~8 bullets unless a table is clearer.

1. Handoff Summary

- When delivering the AGENTS.md to the user, include: - A short summary of major sections added/updated. - Confirmation that README and conventional commit reminders are present. - Any follow-up suggestions (e.g., missing tests or outdated scripts).

Use this skill whenever a repo lacks AGENTS context or when existing instructions are incomplete or outdated. The goal is to leave future agents with a single, trustworthy map of the project, its tooling, and the expectations for finishing tasks responsibly.