claude-md-author（Claude MD author 安全）

CLAUDE.md Authoring Guide

This skill guides the creation of effective CLAUDE.md files that provide Claude with essential project context.

Core Principle

LLMs are stateless—they have no memory between sessions. The only thing Claude knows about a codebase is the tokens provided to it. CLAUDE.md enters every conversation, making it the primary tool for context management.

The WHAT-WHY-HOW Framework

Structure CLAUDE.md content around three categories:

Critical Constraints

Instruction Budget

Frontier LLMs reliably follow approximately 150-200 instructions. Claude Code's system prompt already contains ~50 instructions, leaving limited budget for CLAUDE.md. Smaller models degrade exponentially as instructions increase.

Size Guidelines

• Target under 300 lines
• Aim for under 60 lines if possible (HumanLayer's CLAUDE.md achieves this)
• Every line should provide universal value across all sessions

What to Include

Essential elements:

• Build and test commands (the HOW)
• Project structure overview (the WHAT)
• Key architectural decisions (the WHY)
• Patterns and conventions specific to this codebase
• File references using file:line format (not code snippets)

Example build/test section:

## Commands
- Build: `npm run build`
- Test: `npm test`
- Lint: `npm run lint`
- Type check: `npx tsc --noEmit`

What to Avoid

Never Use as a Linter

"Never send an LLM to do a linter's job." Code style guidelines waste tokens and degrade performance. Use deterministic tools (ESLint, Prettier, Biome) instead. Configure hooks to run formatters automatically.

Avoid Auto-Generation

CLAUDE.md is one of the highest leverage points for Claude's effectiveness. Manually craft it rather than using /init commands. Auto-generated content tends to be verbose and generic.

Skip Universally Irrelevant Content

Claude may ignore CLAUDE.md entirely if content seems task-irrelevant. The system explicitly reminds Claude that "this context may or may not be relevant to your tasks."

No Duplicate Information

Avoid copying code snippets—they become outdated quickly. Use file:line pointers instead.

Progressive Disclosure Strategy

Store task-specific guidance in separate files rather than bloating CLAUDE.md:

.claude/
├── CLAUDE.md              # Core context (universal)
├── docs/
│   ├── testing.md         # Testing procedures
│   ├── deployment.md      # Deployment guide
│   └── architecture.md    # Architecture decisions
└── plans/                 # Implementation plans

Reference these in CLAUDE.md:

## Documentation
- Testing procedures: `.claude/docs/testing.md`
- Deployment guide: `.claude/docs/deployment.md`

Alternative Mechanisms

Consider using Claude Code's other features instead of cramming everything into CLAUDE.md:

Creation Workflow

To create an effective CLAUDE.md:

1. Audit the codebase: Identify build commands, test procedures, key patterns
2. Draft minimal content: Start with WHAT-WHY-HOW essentials only
3. Remove task-specific content: Move specialized instructions to .claude/docs/
4. Use pointers: Replace code snippets with file:line references
5. Test for universality: Ask "Will this help in every Claude session?"

Template

# Project Name

Brief one-line description of what this project does.

## Stack
- [Primary language/framework]
- [Database/storage]
- [Key dependencies]

## Commands
- Build: `command`
- Test: `command`
- Lint: `command`

## Structure
- `src/` - Application code
- `tests/` - Test files
- [Other key directories]

## Key Patterns
- [Pattern 1 with file:line reference]
- [Pattern 2 with file:line reference]

## Notes
- [Important architectural decision]
- [Gotcha or non-obvious behavior]

Additional Resources

For detailed examples and anti-patterns, consult:

• references/examples.md - Real CLAUDE.md examples
• references/anti-patterns.md - Common mistakes to avoid