claude-md 精选

User Input

$ARGUMENTS

You MUST consider the user input before proceeding (if not empty). User may specify:

• create - Create new CLAUDE.md from scratch
• update - Improve existing CLAUDE.md
• audit - Analyze and report on current CLAUDE.md quality
• A specific path to create/update (e.g., src/api/CLAUDE.md for directory-specific instructions)

Core Principles

LLMs are stateless: CLAUDE.md is the only file automatically included in every conversation. It serves as the primary onboarding document for AI agents into your codebase.

The Golden Rules

1. Less is More: Frontier LLMs can follow ~150-200 instructions. Claude Code's system prompt already uses ~50. Keep your CLAUDE.md focused and concise.

1. Universal Applicability: Only include information relevant to EVERY session. Task-specific instructions belong in separate files.

1. Don't Use Claude as a Linter: Style guidelines bloat context and degrade instruction-following. Use deterministic tools (prettier, eslint, etc.) instead.

1. Never Auto-Generate: CLAUDE.md is the highest leverage point of the AI harness. Craft it manually with careful consideration.

Execution Flow

1. Project Analysis

First, analyze the current project state:

1. Check for existing CLAUDE.md files:

• Root level: ./CLAUDE.md or .claude/CLAUDE.md
• Directory-specific: **/CLAUDE.md
• Global user config: ~/.claude/CLAUDE.md

1. Identify the project structure:

• Technology stack (languages, frameworks)
• Project type (monorepo, single app, library)
• Development tools (package manager, build system, test runner)

1. Review existing documentation:

• README.md
• CONTRIBUTING.md
• package.json, pyproject.toml, Cargo.toml, etc.

2. Content Strategy (WHAT, WHY, HOW)

Structure CLAUDE.md around three dimensions:

WHAT - Technology & Structure

• Technology stack overview
• Project organization (especially important for monorepos)
• Key directories and their purposes

WHY - Purpose & Context

• What the project does
• Why certain architectural decisions were made
• What each major component is responsible for

HOW - Workflow & Conventions

• Development workflow (bun vs node, pip vs uv, etc.)
• Testing procedures and commands
• Verification and build methods
• Critical "gotchas" or non-obvious requirements

3. Progressive Disclosure Strategy

For larger projects, recommend creating an agent_docs/ folder:

agent_docs/
  |- building_the_project.md
  |- running_tests.md
  |- code_conventions.md
  |- architecture_decisions.md

In CLAUDE.md, reference these files with instructions like:

For detailed build instructions, refer to `agent_docs/building_the_project.md`

Important: Use file:line references instead of code snippets to avoid outdated context.

4. Quality Constraints

When creating or updating CLAUDE.md:

1. Target Length: Keep it under a few hundred lines; shorter is better
2. No Style Rules: Remove any linting/formatting instructions
3. No Task-Specific Instructions: Move to separate files
4. No Code Snippets: Use file references instead
5. No Redundant Information: Don't repeat what's in package.json or README

5. Essential Sections

A well-structured CLAUDE.md should include:

# Project Name

Brief one-line description.

## Tech Stack
- Primary language and version
- Key frameworks/libraries
- Database/storage (if any)

## Project Structure
[Only for monorepos or complex structures]
- `apps/` - Application entry points
- `packages/` - Shared libraries

## Development Commands
- Install: `command`
- Test: `command`
- Build: `command`

## Critical Conventions
[Only non-obvious, high-impact conventions]
- Convention 1 with brief explanation
- Convention 2 with brief explanation

## Known Issues / Gotchas
[Things that consistently trip up developers]
- Issue 1
- Issue 2

6. Anti-Patterns to Avoid

DO NOT include:

• Code style guidelines (use linters)
• Documentation on how to use Claude
• Long explanations of obvious patterns
• Copy-pasted code examples
• Generic best practices ("write clean code")
• Instructions for specific tasks
• Auto-generated content
• Extensive TODO lists

7. Validation Checklist

Before finalizing, verify:

• [ ] Kept under a few hundred lines; shorter is better
• [ ] Every line applies to ALL sessions
• [ ] No style/formatting rules
• [ ] No code snippets (use file references)
• [ ] Commands are verified to work
• [ ] Progressive disclosure used for complex projects
• [ ] Critical gotchas are documented
• [ ] No redundancy with README.md

Output Format

For create or default:

1. Analyze the project
2. Draft a CLAUDE.md following the structure above
3. Present the draft for review
4. Write to the appropriate location after approval

For update:

1. Read existing CLAUDE.md
2. Audit against best practices
3. Identify:

• Content to remove (style rules, code snippets, task-specific)
• Content to condense
• Missing essential information

1. Present changes for review
2. Apply changes after approval

For audit:

1. Read existing CLAUDE.md
2. Generate a report with:

• Current line count vs target
• Percentage of universally-applicable content
• List of anti-patterns found
• Recommendations for improvement

1. Do NOT modify the file, only report

AGENTS.md Handling

If the user requests AGENTS.md creation/update:

Since v2.1.277, Claude Code reads AGENTS.md directly as project instructions — but only when the working directory and every directory above it contain no CLAUDE.md, .claude/CLAUDE.md, or CLAUDE.local.md. ~/.claude/CLAUDE.md, managed CLAUDE.md, and .claude/rules/ do not count for that check and keep loading alongside. Which files are read is controlled by Project instructions in /config: claude-md-or-agents-md (default), claude-md-and-agents-md, claude-md, or managed-only. Where direct reading is unavailable — versions before v2.1.277 (before v2.1.281 on Bedrock/Vertex/Foundry, LLM gateways, or with telemetry disabled), or the first session after upgrading — fall back to importing it from CLAUDE.md with @AGENTS.md, or symlinking CLAUDE.md to it.

AGENTS.md is a cross-tool project-context file — the same category of document as CLAUDE.md, not an agent-definition format. It exists so several coding agents can share one set of project conventions:

• Build, test, and lint commands
• Code style and architectural conventions
• Repository layout and where things live

Subagents are defined separately, in .claude/agents/*.md — not in AGENTS.md.

Apply similar principles:

• Keep focused and concise
• Use progressive disclosure
• Reference external docs instead of embedding content

Notes

• Always verify commands work before including them
• When in doubt, leave it out - less is more
• The system reminder tells Claude that CLAUDE.md "may or may not be relevant" - the more noise, the more it gets ignored
• Monorepos benefit most from clear WHAT/WHY/HOW structure
• Directory-specific CLAUDE.md files should be even more focused

Last Updated: September 26, 2026

Claude Code Version: 2.1.283

Sources:

• https://code.claude.com/docs/en/skills
• https://code.claude.com/docs/en/memory#agents-md
• https://code.claude.com/docs/en/memory#when-agents-md-support-is-unavailable

Compatible Models: Claude Fable 5, Claude Opus 5, Claude Sonnet 5, Claude Sonnet 4.6, Claude Opus 4.8, Claude Haiku 4.5