AgentSkillsCN

create-agents-md

生成 AGENTS.md 文件,为在代码库中工作的 AI 代理提供被动式上下文信息,包括项目结构、编码规范、技术栈以及精简版文档索引。

SKILL.md
--- frontmatter
name: create-agents-md
description: Creates AGENTS.md files that provide passive context for AI agents working in codebases, including project structure, coding conventions, tech stack, and compressed documentation indexes.
version: 1.0.0
author: jdwork
category: documentation
tags: [agents, context, onboarding]

Skill: Create AGENTS.md

Description

Creates AGENTS.md files—the authoritative instruction file for AI agents working in a codebase. This file is added to the agent's context automatically, providing 100% reliable passive context without requiring tool calls or retrieval decisions.

When to Use

  • Setting up a new project for AI-assisted development
  • Onboarding AI agents to an existing codebase
  • Documenting project conventions that agents should follow
  • Creating a compressed index for large documentation sets

Important: AGENTS.md is for knowledge (conventions, structure, rules). For actions (deploy, test, migrate), create a Skill instead using the create-claude-skill skill.

Instructions

1. Pre-flight Analysis

Before generating, examine the project:

  • Structure: Run ls -la and check directory organization
  • Tech stack: Check package.json, Cargo.toml, pyproject.toml, go.mod, etc.
  • Existing docs: Read README.md for project context
  • Conventions: Look at existing code for style patterns
  • CI/CD: Check .github/workflows/, Makefile, etc. for build commands

2. Structure the File

Use this template structure:

markdown
# AGENTS.md

> This file provides context for AI agents working in this codebase.
> IMPORTANT: Prefer retrieval-led reasoning. Read local doc files before answering framework-specific questions.

## Project Context
- **Project**: [Name and brief purpose]
- **Tech Stack**: [Key technologies]
- **Architecture**: [Monolith/Microservices/Monorepo/etc.]

## Critical Rules
- [Rule 1: Focus on "don'ts" and constraints]
- [Rule 2: Project-specific conventions]
- [Rule 3: Common mistakes to avoid]

## Commands
- **Build**: `[command]`
- **Test**: `[command]`
- **Lint**: `[command]`

## Documentation Index
> Read files from this index when relevant to your task.
[Root]: ./.docs/
[category]: {file1.md, file2.md}

## File-Specific Notes
- **[pattern]**: [convention or note]

3. Write Critical Rules

Focus on constraints and gotchas:

  • What NOT to do (more useful than obvious "do's")
  • Project-specific conventions that differ from defaults
  • Common mistakes agents make in this codebase
  • Security considerations (what not to commit, etc.)

Example rules:

markdown
## Critical Rules
- Never use `any` type in TypeScript—use `unknown` and narrow
- Always use `server-only` import for server utilities
- Run `pnpm check` before committing—it runs lint + typecheck + test
- Don't modify files in `generated/`—they're auto-generated

4. Add Compressed Documentation Index (if applicable)

For projects with extensive documentation (frameworks, large codebases):

  1. Store docs in a .docs/ folder (or reference existing docs location)
  2. Create a compressed index mapping categories to files
  3. Add the retrieval instruction

Format:

markdown
## Documentation Index
> Agents: Read the file at the path below if related to your task.

[Root]: ./.docs/
01-getting-started: {installation.md, project-structure.md}
02-api: {endpoints.md, authentication.md, error-handling.md}
03-database: {schema.md, migrations.md, queries.md}

5. Interoperability

Recommend symlinking to support multiple tools:

bash
# Create symlinks for other AI tools
ln -s AGENTS.md CLAUDE.md

Note: Only recommend symlinks the user has tools for. Common conventions:

  • CLAUDE.md - Claude Code
  • .cursorrules - Cursor
  • COPILOT.md - GitHub Copilot

Anti-Patterns

  • DON'T dump entire documentation files—link to them instead
  • DON'T state obvious rules ("write clean code")
  • DON'T duplicate README content verbatim
  • DON'T include frequently-changing information (versions, etc.)
  • DON'T add action workflows—those belong in Skills

Examples

Minimal AGENTS.md (small project)

markdown
# AGENTS.md

## Project Context
- **Project**: Personal blog built with Astro
- **Stack**: Astro, TypeScript, Tailwind CSS, MDX

## Rules
- Content lives in `src/content/`—follow the existing frontmatter schema
- Don't modify `src/components/ui/`—they're from shadcn

## Commands
- **Dev**: `pnpm dev`
- **Build**: `pnpm build`

With Documentation Index (framework project)

markdown
# AGENTS.md

> IMPORTANT: Read local doc files before answering framework questions.

## Project Context
- **Project**: E-commerce platform
- **Stack**: Next.js 15, TypeScript, Prisma, PostgreSQL
- **Architecture**: Monorepo with apps/ and packages/

## Documentation Index
[Root]: ./docs/
architecture: {overview.md, data-flow.md}
api: {rest-endpoints.md, graphql-schema.md}
database: {schema.md, migrations.md}

## Critical Rules
- Use Server Actions for mutations, not API routes
- All database queries go through `packages/db`
- Never import from `apps/` into `packages/`

Cross-Reference

For repeatable workflows and actions, create a Skill instead using the create-claude-skill skill. Skills are better for:

  • Deployment procedures
  • Database migrations
  • Test suite execution
  • Code generation workflows