AgentSkillsCN

agents-md

创建并编辑 CLAUDE.md 或 AGENTS.md 文件——AI 代理指令文档——的指南。当用户希望新建 CLAUDE.md/AGENTS.md 文件、编辑现有代理文档、为代码库设置 AI 指令,或询问代理文档结构的最佳实践时,可使用本指南。

SKILL.md
--- frontmatter
name: agents-md
description: Guide for creating and editing CLAUDE.md or AGENTS.md files - AI agent instruction documents. Use when the user wants to create a new CLAUDE.md/AGENTS.md, edit existing agent documentation, set up AI instructions for a repository, or asks about best practices for agent docs structure.

Agents MD

Overview

This skill provides knowledge and templates for creating and editing CLAUDE.md and AGENTS.md files - structured documentation that guides AI coding agents working with a codebase.

Output Format Selection

Let the user choose the output format:

FormatUse Case
CLAUDE.mdClaude Code specific (supports @ imports, hierarchical memory)
AGENTS.mdCross-tool compatible (Codex, Cursor, Copilot, etc.)

For format-specific features, see:

  • references/claude-md-features.md - Claude Code specific features
  • references/agents-md-spec.md - AGENTS.md specification

Document Structure

Use this 8-section structure. See assets/template.md for the full template.

1. Project Name & Overview

Start with # [Project Name] followed by a brief description (1-3 sentences) explaining WHY this project exists, then list key features/capabilities as bullet points. No separate section heading needed.

2. Architecture Overview

Key architectural points as bullet list - important components, data flow, design decisions. Answer "WHAT are the key things to know?" Keep it concise (3-5 bullets). If docs/ARCHITECTURE.md or similar exists, link to it for details.

3. Directory Structure

Tree view of the repository layout with brief descriptions. Answer "WHERE is everything?"

4. Core Principles

Coding principles to follow. Use these defaults as-is; append project-specific items rather than replacing:

markdown
## Core Principles

- Prefer existing patterns and terminology over general best practices
- Avoid adding new dependencies unless necessary; remove when possible
- Follow Conventional Commits for commit messages

5. Commands

Build, run, lint, and test commands. Example:

markdown
## Commands

- `npm install` - Install dependencies
- `npm run dev` - Start development server
- `npm run build` - Production build
- `npm test` - Run all tests
- `npm run lint` - Run linter

6. Testing

Testing strategy and policies (not commands). Use these defaults as-is; append project-specific items rather than replacing:

markdown
## Testing

- Prefer integration tests over unit tests
- Use in-source testing for unit tests covering edge cases
- Avoid mocks; use them only for external communication or resource fetching

7. Language Policy

Programming and natural language rules. Use these defaults as-is; append project-specific items rather than replacing:

markdown
## Language Policy

- Follow the user's language for comments, commits, and tests
- Write the following files in English:
  - CLAUDE.md, AGENTS.md
  - Files under `.claude/`
  - Files under `.github/` (except comments)

8. Additional Resources

References to detailed documentation. Format:

markdown
## Additional Resources

- `docs/ARCHITECTURE.md`: Detailed architecture decisions
- `docs/agents/[FILE].md`: [Description]
- `docs/adr/adr-001-[slug].md`: [ADR title]

Best Practices

For detailed guidelines on writing effective agent documentation, see references/best-practices.md.

Key points:

  • Keep under 300 lines (ideally under 100)
  • Be specific, not generic
  • Use Progressive Disclosure - link to detailed docs instead of including everything
  • Let linters handle code style, not agent docs