agents-md-generator（特工 MD 生成器）

AGENTS.md / CLAUDE.md Generator

Overview

This skill helps you generate comprehensive instruction files (CLAUDE.md or AGENTS.md) that teach AI coding agents how to work effectively in your project. It combines automatic project scanning with interactive questions to create tailored guidelines.

When to use this skill:

• Setting up a new project for AI-assisted development
• Updating existing instruction files after project changes
• Standardizing AI agent behavior across team members
• Migrating from one AI tool to another

Key Principle

Do not duplicate specialized skills. If a request falls into a specialized domain (e.g., Design System), delegate to the specialized skill when available.

Quick Start

To generate a new CLAUDE.md file:

1. Navigate to your project root
2. Tell the AI agent: "Use the agents-md-generator skill to create a CLAUDE.md file"
3. Answer the interactive questions about your project
4. Review and customize the generated file

Answering Questions (Convenient Formats)

When this skill asks numbered questions with lettered options, users can answer in any style:

• Short form (fast): 1a 2b 3c
• Mixed form: 1a 2b 3c 4b (also scan packages/*) 5d
• Full sentences: “Use medium scan depth and Docker Compose; service name is app.”

Short form is never required; it is only provided for convenience.

Generation Modes

Interactive Mode (Default):

• Guides you through each phase with questions
• Best for first-time setup or complex projects
• Maximum customization

Quick Mode:

• Tell the AI: "Generate CLAUDE.md in quick mode"
• Skips all questions, uses auto-detection only
• Best for experienced users or simple projects
• Uses Medium scan depth by default
• Generates all standard sections based on detection

Interactive Workflow

Phase 1: Initialization & Discovery

Check for existing files:

1. Look for CLAUDE.md or AGENTS.md in the project root
2. If found, ask user: "I found an existing [filename]. Would you like to:"

1. Update it (merge new content while preserving customizations) 2. Replace it (generate fresh, backup existing) 3. Cancel - Reply examples (optional): 1 or update; 2 or replace; 3 or cancel

Determine primary file:

• Default: CLAUDE.md as primary, AGENTS.md as symlink
• For Windows or if user prefers: Both files with sync header

Phase 1b: Agent Tooling & System Prompt Alignment (Scripted)

Goal: Detect AI tool instruction files, global system prompts, and MCP configs using scripts (not AI scanning) to save tokens and avoid missing overrides.

Run the detection script (best effort):

• bash/zsh: bash skills/agents-md-generator/scripts/detect-agent-context --root.
• Windows cmd: skills\agents-md-generator\scripts\detect-agent-context.cmd --root.

The script should report:

• Project instruction files for AI tools (Copilot, Cursor, Cline, Kilo Code, Roo Code, OpenCode, Codex, Claude Code)
• Global instruction files (e.g., ~/.claude/CLAUDE.md, ~/.codex/config.toml, ~/.roo/rules/, ~/.kilocode/rules/)
• MCP config files and server names (from .mcp.json, .roo/mcp.json, mcp_settings.json, plus any --mcp-path entries)

If script is unavailable:

• Do a minimal manual check using the paths listed in references/tech-stack-detection.md under AI Agent Tooling Detection.

Warn the user about overrides:

• If ~/.claude/CLAUDE.md or other global instruction files exist, explicitly warn that they can override project prompts.
• Ask the user to review or adjust those system prompts to avoid conflicts with this repo.

Ask the user to confirm tool usage:

I detected these AI tool instruction sources:
- .github/copilot-instructions.md (GitHub Copilot)
- .cursorrules (Cursor)
- .clinerules or .clinerules/ (Cline)
- .kilocoderules / .kilo/ / .kilocodemodes (Kilo Code)
- .roo/rules/ / .roo/rules-* / .roorules* (Roo Code)
- opencode.jsonc (OpenCode)
- ~/.codex/config.toml (Codex)
- .claude/CLAUDE.md or ~/.claude/CLAUDE.md (Claude Code)

Which of these do you actively use for this project, and should we align or ignore any of them?

If MCP servers are detected:

• Ask whether they should be used in this project.
• Capture a short purpose/usage note for each server.

Phase 1c: Skill Library Duplicate Scan (Scripted)

If a skills/ folder exists, run the duplicate scan script to avoid copy-pasted skills:

• bash/zsh: bash skills/agents-md-generator/scripts/scan-skill-duplicates --skills-dir skills
• Windows cmd: skills\agents-md-generator\scripts\scan-skill-duplicates.cmd --skills-dir skills

If duplicates are detected:

• Recommend consolidating with symlinks so there is one source of truth.
• Example (macOS/Linux): ln -s./skills/skill-a./skills/skill-b
• On Windows, recommend a copy with a clear header if symlinks are not supported.

Phase 2: Scan Depth Selection

Ask the user:

What scan depth should I use to analyze your project?

1. Quick (approximately 30 seconds)
   Scans: package.json, composer.json, docker-compose.yml, pyproject.toml,
          Gemfile, go.mod, Cargo.toml, and other root config files
   Best for: When you know your stack well and want fast generation

2. Medium (approximately 1-2 minutes) [RECOMMENDED]
   Scans: Root configs + src/, app/, lib/, config/, routes/, components/,
          pages/, views/, controllers/, models/, services/
   Best for: Most projects - good balance of accuracy and speed

3. Deep (approximately 3-5 minutes)
   Scans: Entire project tree including tests/, docs/, scripts/, all
          subdirectories, hidden configs, and build artifacts
   Best for: Complex projects, monorepos, or unfamiliar codebases

Reply examples:
- Short: `2` (or `1` / `3`)
- With extra notes: `2 (also scan packages/*)`

Scan actions per depth:

Phase 3: Environment Detection

Ask the user:

What development environment does this project use?

1. Docker Compose
   - Commands run via: docker compose exec <service> <command>
   - Example: docker compose exec app php artisan migrate

2. Laravel Sail
   - Commands run via: ./vendor/bin/sail <command>
   - Example: ./vendor/bin/sail artisan migrate

3. Native/Host Machine
   - Commands run directly on your machine
   - Example: php artisan migrate

4. Dev Containers / Codespaces
   - Commands run inside the container environment

5. Other (please describe)
   - Specify your custom environment setup

Reply examples:
- Short: `1` (or `2` / `3` / `4` / `5`)
- Detailed: `1; main service is app; node runs in node service`

Follow-up questions based on selection:

• Docker: "What is the main service name? (e.g., app, web, php)"
• Docker: "Do you have separate services for different runtimes? (e.g., app for PHP, node for JS)"
• Native: "Do you use any version managers? (nvm, rbenv, pyenv, etc.)"

Phase 4: Auto-Detection + Confirmation

Scan the project based on selected depth and detect:

1. Backend Framework:

- Laravel (composer.json + artisan) - Express/Node (package.json + server files) - Django/Flask (requirements.txt + manage.py/app.py) - Rails (Gemfile + config/routes.rb) - Spring Boot (pom.xml/build.gradle + @SpringBootApplication)

1. Frontend Framework:

- Vue.js (package.json +.vue files) - React (package.json +.jsx/.tsx files) - Angular (angular.json) - Svelte (svelte.config.js) - Next.js/Nuxt.js (next.config.js/nuxt.config.ts)

1. Package Manager:

- npm/yarn/pnpm/bun (package-lock.json/yarn.lock/pnpm-lock.yaml/bun.lockb) - Composer (composer.lock) - pip/poetry (requirements.txt/poetry.lock)

1. Testing Framework:

- PHPUnit (phpunit.xml) - Pest (pestphp/pest in composer.json) - Jest (jest.config.js) - Vitest (vitest.config.ts) - pytest (pytest.ini/conftest.py)

1. Code Style/Linting:

- Laravel Pint (pint.json) - ESLint (.eslintrc.*) - Prettier (.prettierrc.*) - PHP CS Fixer (.php-cs-fixer.php)

1. Database:

- MySQL/MariaDB (config references) - PostgreSQL (config references) - SQLite (database/*.sqlite) - MongoDB (mongoose in package.json)

1. AI Agent Tooling (scripted detection):

- GitHub Copilot (.github/copilot-instructions.md) - Cursor (.cursorrules) - Cline (.clinerules or .clinerules/) - Kilo Code (.kilocoderules, .kilo/, .kilocodemodes, .kilocode/config.json) - Roo Code (.roo/rules/, .roo/rules-*, .roorules*, .roo/mcp.json, mcp_settings.json) - OpenCode (opencode.jsonc, OPENCODE_CONFIG) - Codex (~/.codex/config.toml) - Claude Code (.claude/CLAUDE.md, ~/.claude/CLAUDE.md, .mcp.json)

Present findings to user:

I detected the following tech stack:

Backend:
  - Laravel 11 (high confidence)
  - PHP 8.3 (from composer.json)

Frontend:
  - Vue.js 3 with Composition API (high confidence)
  - TailwindCSS v3 (from package.json)

Testing:
  - Pest PHP (from composer.json)
  - Vitest (from package.json)

Code Style:
  - Laravel Pint (pint.json found)
  - ESLint + Prettier (configs found)

Environment:
  - Docker Compose detected (docker-compose.yml)
  - Services: app, mysql, redis

Is this correct? Would you like to add or modify anything?

Phase 4b: Optional Sections (Interactive Mode Only)

Ask the user about optional sections:

Would you like to include any of these optional sections?

1. CI/CD Configuration
   - Detect: GitHub Actions, GitLab CI, CircleCI, Jenkins
   - Include: Pipeline commands, deployment notes

2. Git Workflow Guidelines
   - Include: Branch naming, commit message format, PR guidelines
   - Detect: .github/PULL_REQUEST_TEMPLATE.md, commitlint config

3. Security Guidelines
   - Include: Env file handling, secrets management, input validation
   - Best practices for the detected stack

4. API Documentation
   - Detect: OpenAPI/Swagger specs, Postman collections
   - Include: Documentation conventions and tooling

5. Mobile App Guidelines (if detected)
   - React Native / Flutter specific patterns
   - Platform-specific considerations

6. Monorepo Guidelines (if detected)
   - Nx / Turborepo / Lerna workspace patterns
   - Package management and dependencies

7. System Prompt Alignment (if detected)
   - Document global or editor-level prompts that can override project rules
   - Remind contributors to review and align prompts

8. MCP Servers & Tooling (if detected)
   - List MCP servers and when to use them
   - Include required environment variables or access notes

9. Project Progress Memory (PROGRESS.md)
   - Ask if the team wants `PROGRESS.md` for continuity across tasks
   - Require an `Original Prompt` section so future runs can compare intent vs current status

Select the sections you need (comma-separated numbers, or 'none' to skip):

Reply examples:
- Short: `none`
- Multiple: `1,3,4`
- With extra notes: `2,3,7 (also include branch naming rules)`

Note: In Quick Mode, these optional sections are skipped unless auto-detected with high confidence.

Phase 5: Section Generation

Generate the following sections based on detected stack and user input:

Section 1: Header & Role

# AI Agent Guidelines & Repository Manual

**Role:** You are an expert Senior [DETECTED_ROLE] and Technical Lead.
You are responsible for the entire lifecycle of a task: understanding,
planning, [STACK_SPECIFIC_RESPONSIBILITIES].

Role detection rules:

• Laravel only → "Laravel Backend Engineer"
• Vue/React only → "[Framework] Frontend Engineer"
• Laravel + Vue/React → "Full-Stack Developer"
• Node.js backend → "Node.js Backend Engineer"
• Generic → "Software Engineer"

Section 2: Auto-Pilot Workflow

Generate the 6-step workflow cycle:

1. Discovery & Context - What to read first, where to find docs
2. Plan - How to break down tasks, constraints to check
3. Documentation - When to update docs, what format to use
4. Implementation - Coding standards, patterns to follow
5. Verification & Refinement - Testing, linting, manual checks
6. Self-Review - Checklist of common mistakes to avoid

Each step includes stack-specific instructions from templates.

Section 3: Documentation & Knowledge Base

List paths to important documentation:

## Documentation & Knowledge Base

You are expected to read and adhere to these single sources of truth:

* **[Doc Type]**: `[path/to/doc.md]` ([Brief description])

Auto-detect common paths:

• docs/, documentation/
• README.md, CONTRIBUTING.md
• docs/api/, docs/architecture/
• Important: Do not list CLAUDE.md or AGENTS.md in this section. These files are already loaded by AI tools, and self-references waste context.
• For new docs, prefer YAML frontmatter + Markdown body (headings, tables, examples) so metadata and content stay consistent.

Section 4: Project Structure & Architecture

Map the folder structure with purposes:

## Project Structure & Architecture

* **`[folder/]`**: [Purpose description]

Common patterns to detect:

• MVC structure (controllers, models, views)
• Feature-based modules
• Domain-driven design
• Component-based frontend

Section 5: Development Environment

Based on Phase 3 selection:

## Development Environment

### Container Commands (if Docker)
* App container: `docker compose exec [service] <command>`

### Host Commands
* Git, file operations, IDE commands

### Key Commands
* Format: `[detected formatter command]`
* Test: `[detected test command]`
* Build: `[detected build command]`

Section 6: Coding Standards

Based on detected stack:

## Coding Standards (The "Gold Standard")

* **Language**: [Language] [Version]
* **Framework**: [Framework] [Version]
* **Code Style**: [Style guide/tool]
* **Strictness**: [Type hints, strict mode, etc.]

Include anti-patterns section if applicable:

### Critical Anti-Patterns
- [Stack-specific anti-patterns to avoid]

Section 7: Domain Specifics (Optional)

If the project has specific domain rules detected:

## Domain Specifics & Non-Negotiables

* **[Rule Category]**: [Rule description]

Common domain patterns:

• Multi-tenant applications
• Permission/role systems
• Localization requirements
• Real-time features

Section 8: CI/CD Configuration (Optional)

If user selected or auto-detected:

## CI/CD & Deployment

### Detected Pipelines
* **GitHub Actions**: `.github/workflows/`
* **GitLab CI**: `.gitlab-ci.yml`

### Pipeline Commands
* Run tests: `[detected command]`
* Build: `[detected command]`
* Deploy: `[detected command]`

### Deployment Notes
* [Environment-specific notes]

Section 9: Git Workflow (Optional)

If user selected:

## Git Workflow

### Branch Naming
* Feature: `feature/<ticket>-<description>`
* Bugfix: `fix/<ticket>-<description>`
* Hotfix: `hotfix/<description>`

### Commit Message Format

type(scope): description

[optional body]

Types: `feat`, `fix`, `docs`, `style`, `refactor`, `test`, `chore`

### Pull Request Guidelines
* Reference ticket/issue in description
* Ensure tests pass before requesting review
* Keep PRs focused and reasonably sized

Section 10: Security Guidelines (Optional)

If user selected:

## Security Guidelines

### Environment Variables
* Never commit `.env` files (only `.env.example`)
* Use secrets management for production
* Rotate credentials regularly

### Input Validation
* Validate all user input at boundaries
* Sanitize data before database queries
* Use parameterized queries (ORM handles this)

### Authentication & Authorization
* [Stack-specific auth patterns]
* Always verify permissions before actions

Section 11: API Documentation (Optional)

If detected or user selected:

## API Documentation

### Documentation Location
* OpenAPI Spec: `[path/to/openapi.yaml]`
* Postman Collection: `[path/to/collection.json]`

### Documentation Standards
* Keep API docs in sync with implementation
* Document all endpoints, request/response schemas
* Include example requests and responses

Section 12: Mobile Guidelines (Optional)

If React Native or Flutter detected:

## Mobile Development

### Platform Considerations
* Test on both iOS and Android
* Handle platform-specific UI patterns
* Consider offline functionality

### Build Commands
* iOS: `[build command]`
* Android: `[build command]`

Section 13: Monorepo Guidelines (Optional)

If monorepo detected:

## Monorepo Structure

### Workspace Management
* Package manager: [pnpm/yarn/npm workspaces]
* Build tool: [Nx/Turborepo/Lerna]

### Package Dependencies
* Use workspace protocol for internal packages
* Keep shared dependencies at root level

### Commands
* Build all: `[command]`
* Build affected: `[command]`
* Test affected: `[command]`

Section 14: System Prompt Alignment (Optional)

If global or editor-level prompts were detected or the user requests it:

## System Prompt Alignment

These prompts can override project instructions. Review and align them with this file:

* **Claude Code Global Prompt**: `~/.claude/CLAUDE.md` (review for conflicts)
* **GitHub Copilot**: `.github/copilot-instructions.md` (project scope)
* **Cursor Rules**: `.cursorrules`
* **Cline Rules**: `.clinerules` or `.clinerules/`
* **Kilo Code Rules**: `.kilocoderules` / `.kilo/` / `.kilocodemodes`
* **Roo Code Rules**: `.roo/rules/` / `.roo/rules-*` / `.roorules*`
* **OpenCode Config**: `opencode.jsonc`
* **Codex Config**: `~/.codex/config.toml`

If any of these conflict with this file, update the global/system prompts first.

### Local AI Tool Folders & Git Hygiene
Remind contributors to keep local AI tool data out of Git:

* Add local folders like `.codex/` and `.claude/` to `.gitignore` because they may contain sensitive prompts, logs, or secrets.
* Put shareable skill packs in `.agents/skills/` so teams can version and share them safely.

Section 15: MCP Servers & Tooling (Optional)

If MCP servers are detected and user opts in:

## MCP Servers & Tooling

Use these MCP servers when the task matches their capability:

* **[server-name]**: [Purpose, when to use it, required env vars]

If a server is not needed for this project, disable it in the MCP config.

Section 16: Project Progress Memory (PROGRESS.md) (Optional)

If the user wants project progress tracking:

## Project Progress Tracking

Use one lightweight progress file that all agents can continue from:

* **Progress**: `PROGRESS.md` (current focus + recent completions)
* **Required field**: `Original Prompt` (copy exact user request that started the task)

At the start of every task, read `PROGRESS.md` first, compare with the active request, and continue unfinished work before starting unrelated changes.

If a memory MCP server is used instead, keep `PROGRESS.md` minimal or omit it.

Phase 6: File Creation/Update

For new files:

1. Write CLAUDE.md with generated content
2. Create AGENTS.md as symlink: ln -s CLAUDE.md AGENTS.md
3. If symlink fails (Windows), create copy with header: <!-- This file mirrors CLAUDE.md. Edit CLAUDE.md as the primary source. -->

For updates (merge mode):

1. Parse existing file into sections (split by ## headers)
2. Compare auto-detected findings against existing guidance; if they conflict, ask the user whether to keep existing content, replace it, or merge.
3. For each section:

- If exists in both: Show diff and ask user preference - If only in existing: Preserve (user customization) - If only in new: Add with note

1. Generate merged file
2. Show summary of changes
3. Ensure the generated content does not instruct the agent to read CLAUDE.md or AGENTS.md, since those files are already loaded by AI tools.

Design System Delegation (Skill Reuse Policy)

This skill is intentionally not a full Design System engine. When the user requests any of the following:

• "design system", "UI consistency", "style guide"
• "colors/typography/tokens"
• "component library rules"
• "generate DESIGN_SYSTEM.md"

Then this skill must:

If design-system-generator skill is available:

1. Delegate generation of DESIGN_SYSTEM.md to design-system-generator
2. Then update AGENTS.md/CLAUDE.md to reference DESIGN_SYSTEM.md with this block:

## Design System
All UI components and pages must follow `DESIGN_SYSTEM.md`:
- Use design tokens (no hardcoded colors/sizes).
- Implement component states (hover/focus/disabled/loading/error).
- Meet accessibility and performance requirements.

If design-system-generator is NOT available:

1. Suggest installing design-system-generator from:

- https://github.com/thienanblog/awesome-ai-agent-skills (documentation-skills plugin)

1. Produce only:

- AGENTS.md/CLAUDE.md patch referencing DESIGN_SYSTEM.md - Optional minimal scaffold DESIGN_SYSTEM.md (no deep recommendations)

Minimal Scaffold for DESIGN_SYSTEM.md (only if design-system-generator unavailable)

If the user still wants a file now, generate ONLY this scaffold:

# DESIGN_SYSTEM.md (Scaffold)

## Scope
Defines UI consistency rules for this project.

## Tokens (TBD)
- Colors: CSS variables
- Typography: scale + line-height
- Spacing: spacing scale
- Radius/Shadows: scales

## Components
Define component patterns and required states:
- hover, focus, disabled, loading, error

## Production assets
Use hashed filenames + a manifest mapping to avoid cache issues.
Minify CSS/JS and optimize images/fonts.

Do not pick Tailwind/MUI/shadcn/etc. in the scaffold unless the project already uses it.

Never implement the full Design System logic inside agents-md-generator.

Tech Stack Detection Reference

See references/tech-stack-detection.md for complete detection patterns.

Section Templates Reference

See references/section-templates.md for complete section templates per stack.

Update/Merge Strategy Reference

See references/merge-strategy.md for detailed merge logic.

Progress Tracking (Maintainers)

When discussing or implementing new ideas/features for this skill, use PROGRESS.md as the single continuity file.

Required PROGRESS.md sections:

• Original Prompt (verbatim user request that initiated the task)
• Current Status (what is done, in progress, blocked)
• Next Steps (clear continuation checklist)

Maintainer workflow:

1. Read PROGRESS.md before starting any task.
2. Compare Original Prompt and Current Status against the new request.
3. Continue unfinished work first when it matches the same objective.
4. Update Current Status and Next Steps before ending the task.

Archiving rule: keep PROGRESS.md readable (about 200-300 lines max). Move old completed entries to docs/archives/PROGRESS-YYYY-MM.md when needed.

Output File Naming

Primary file: CLAUDE.md

• This is the main instruction file that AI agents read
• All edits should be made to this file

Secondary file: AGENTS.md

• Symlink to CLAUDE.md (Unix/macOS/Linux)
• Or copy with header note (Windows)
• Provides compatibility with tools expecting AGENTS.md

Why this approach:

• Single source of truth prevents drift
• Works across all AI coding tools
• Follows Claude Code conventions while supporting others

*This skill is part of the awesome-ai-agent-skills community library.*