claude-md-authoring（Claude MD authoring 搜索）

CLAUDE.md Authoring

Quickstart

1. Determine scope: User-level (~/.claude/) or project-level (.claude/)?
2. Apply WHAT/WHY/HOW framework (project-level only)
3. Keep instructions under budget: ~50 for user-level, ~100 for project-level
4. Never use Claude as a linter: Style rules belong in tooling, not CLAUDE.md
5. Validate with quality checklist before finalizing

Core Principle

"The only thing that the model knows about your codebase is the tokens you put into it."

CLAUDE.md is the highest-leverage point for context transfer. Every instruction competes for limited attention. Include only what Claude cannot infer and must apply universally.

Scope Decision

User-Level (~/.claude/CLAUDE.md)

Applies to ALL projects. Include only:

• Personal working style preferences
• Relationship dynamics (pushback, honesty, communication)
• Universal philosophy (YAGNI, root-cause debugging)
• Proactiveness boundaries

Never include:

• Project-specific commands (build, test, lint)
• Tech stack details
• Style rules that vary by project
• Tool/framework references that don't exist everywhere

Project-Level (.claude/CLAUDE.md)

Applies to ONE project. Use the WHAT/WHY/HOW framework:

WHAT (2-5 lines)

• Project purpose in one sentence
• Primary language/framework
• Key directories and their purposes

WHY (optional, 2-3 lines)

• Non-obvious architectural decisions
• Constraints that affect implementation choices

HOW (3-7 lines)

• Essential commands: build, test, lint, run
• Any project-specific workflows
• Where to find more documentation

Instruction Budget

LLMs reliably follow ~150-200 instructions. Claude Code's system prompt uses ~50, leaving limited budget.

Counting instructions: Each actionable directive counts as one instruction.

• "NEVER skip tests" = 1 instruction
• "Use TypeScript for new files" = 1 instruction
• A 5-item list = 5 instructions

Anti-Pattern: Claude as Linter

❌ Never rely on Claude for style enforcement:

• Naming conventions with specific examples
• Comment formatting rules
• Whitespace/indentation requirements
• File header requirements (ABOUTME, copyright)

✓ Use deterministic tools instead:

• ESLint/Biome for JavaScript/TypeScript style
• Black/Ruff for Python formatting
• Pre-commit hooks for file headers
• Claude Code Hooks to run formatters

Why this matters: LLMs follow style rules inconsistently. A linter fails deterministically; Claude fails silently and unpredictably.

Exception: High-level philosophy is acceptable:

• ✓ "Names should describe purpose, not implementation"
• ❌ "Never use names like ZodValidator or MCPWrapper"

Progressive Disclosure

For complex guidance, don't embed everything in CLAUDE.md.

Pattern: Reference external docs, let Claude decide relevance.

## Reference Documentation
- Code conventions: See `docs/conventions.md`
- API patterns: See `docs/api-design.md`
- Testing philosophy: See `docs/testing.md`

Read these when the task involves the relevant area.

Benefits:

• Keeps CLAUDE.md focused
• Claude reads docs only when relevant
• Easier to maintain separate concerns

Structure Template

User-Level

[Opening persona/philosophy - 1-2 lines]

## Working Together
[Relationship dynamics, communication preferences]

## Proactiveness
[When to proceed vs. pause for confirmation]

## Development Philosophy
[Universal principles: YAGNI, debugging approach, etc.]

## Guardrails
[Version control, testing principles - kept brief]

Project-Level

## Project Overview
[WHAT: Purpose, stack, structure]

## Development
[HOW: Build, test, lint commands]

## Architecture
[WHY: Key decisions, constraints - if non-obvious]

## Conventions
[Brief pointers to detailed docs if needed]

Quality Checklist

Before finalizing, verify:

Universal Applicability

• Every instruction applies to every session (user-level) or every task in project (project-level)
• No project-specific commands in user-level config
• No tool references that may not exist

Instruction Economy

• Under instruction budget (count them)
• No redundancy with Claude Code's built-in behaviors
• Examples trimmed or moved to reference docs

No LLM-as-Linter

• No specific naming convention examples
• No comment formatting rules
• No file header requirements
• Style guidance limited to philosophy, not specifics

Progressive Disclosure

• Complex topics reference external docs
• Main file under 100 lines (user) or 150 lines (project)
• Examples in reference files, not inline

Clarity

• Consistent heading hierarchy
• No conflicting instructions
• Actionable directives, not vague guidance

Common Pitfalls

1. Over-specification: Including every preference instead of universally critical ones
2. Style enforcement: Detailed naming/comment rules that Claude follows inconsistently
3. Stale commands: Build/test commands that drift from actual tooling
4. Auto-generation: Using /init instead of hand-crafting
5. Tool assumptions: Referencing MCP servers or tools that aren't always available
6. Conflicting rules: "Match surrounding style" + "Never use X pattern"
7. Instruction inflation: Adding rules after each frustration instead of fixing root cause

When to Skip This Skill

• Quick one-off projects (just use defaults)
• Projects with existing, well-maintained CLAUDE.md
• When the team already has established patterns

Example: Before/After

Before (problematic user-level):

## Naming
NEVER use names like ZodValidator, MCPWrapper, JSONParser...
[20 more lines of specific examples]

## Comments
All files MUST start with ABOUTME: header...

After (improved user-level):

## Code Style
Names should describe purpose, not implementation details.
Use project linters for formatting; don't rely on manual enforcement.

*Why better*: Philosophy over specifics. Linter handles enforcement deterministically.

Integration

After authoring CLAUDE.md:

• User-level: Symlink to ~/.claude/CLAUDE.md or place directly
• Project-level: Commit to .claude/CLAUDE.md in repository root
• Validation: Start a new Claude Code session and observe if instructions are followed