init-deep（初始化深度）

Init Deep Skill

Create a minimal root CLAUDE.md with nested CLAUDE.md files throughout the repository for true progressive disclosure.

Key Insight

Claude automatically loads CLAUDE.md from any directory it reads. Instead of centralizing documentation in .claude/docs/, place CLAUDE.md files directly in relevant directories. When Claude reads src/api/users.ts, it also loads src/api/CLAUDE.md - contextual guidance exactly when needed.

Closest Wins: When multiple CLAUDE.md files apply, the one closest to the file being edited takes precedence. A rule in src/api/CLAUDE.md overrides the same topic in root CLAUDE.md.

When This Skill Activates

CLI Variants

/init-deep              # Update: modify existing + create new where needed
/init-deep --fresh      # Delete all CLAUDE.md files, regenerate from scratch
/init-deep --max-depth=2 # Limit nesting depth (default: unlimited)

Philosophy: Instruction Budget

Claude has ~150-200 instruction capacity before diminishing returns. A bloated root CLAUDE.md means wasted context on every interaction.

Solution: Minimal root + nested CLAUDE.md files that load only when relevant directories are accessed.

Size Limits (Strict)

Telegraphic style required. No verbose explanations. Each line must carry weight.

Deduplication Rules

Child CLAUDE.md files NEVER repeat parent content:

• If root says "use TypeScript" - children don't repeat it
• If root defines import order - children inherit it
• Child files add domain-specific rules only

Review Phase Requirement: Before finalizing, scan all generated CLAUDE.md files for duplicated guidance. Remove duplicates from children, keep in nearest common ancestor.

Directory Scoring System

Not every directory needs a CLAUDE.md. Use weighted scoring to decide.

Scoring Formula

Score = (file_count * 3) + (reference_centrality * 3) + (symbol_density * 1) + (export_count * 1)

Score Thresholds

Quick Heuristics

When scoring isn't practical, use these rules:

Two Operating Modes

Mode 1: Fresh Project (No CLAUDE.md)

Create minimal root + nested structure from scratch.

Mode 2: Existing CLAUDE.md

Analyze, categorize, and migrate content to nested CLAUDE.md files.

Workflow

Phase 1: Discovery

Goal: Understand current state and scope.

Actions:

1. Check if root CLAUDE.md exists
2. Scan for common directories that should have nested CLAUDE.md
3. Check for existing nested CLAUDE.md files
4. Detect if .claude/docs/ exists (old pattern to migrate)
5. Calculate directory scores for prioritization

Dynamic Agent Scaling:

Concurrent Execution: Launch all exploration agents in parallel (one message, multiple Task calls). Do not wait for one to finish before starting others.

Directory Detection Patterns:

src/           # Source code root
tests/         # Test files
test/          # Alternative test directory
lib/           # Library code
api/           # API layer (may be under src/)
components/    # UI components
hooks/         # React/custom hooks
utils/         # Utility functions
services/      # Service layer
scripts/       # Build/dev scripts
packages/      # Monorepo packages
apps/          # Monorepo apps

Nested directory patterns (check inside src/):

src/api/
src/components/
src/hooks/
src/utils/
src/services/
src/lib/
src/routes/
src/pages/
src/stores/
src/models/

Detection Script:

# Find directories that should have CLAUDE.md
for dir in tests test src lib api components hooks utils services scripts packages apps; do
  if [ -d "$dir" ]; then
    echo "FOUND: $dir/"
  fi
done

# Check src/ subdirectories
if [ -d "src" ]; then
  for subdir in api components hooks utils services lib routes pages stores models; do
    if [ -d "src/$subdir" ]; then
      echo "FOUND: src/$subdir/"
    fi
  done
fi

Phase 2: Interview

Goal: Gather essential project info.

For Fresh Projects, Ask:

Question 1: Project Description

What does this project do? (One sentence)

Example: "A CLI tool for managing Kubernetes deployments"

Question 2: Package Manager

Which package manager does this project use?

Options:
- npm (default, will not be documented)
- pnpm
- yarn
- bun
- none (not a JS/TS project)

Question 3: Non-Standard Commands

Any build, test, or lint commands that differ from standard?

Leave blank if using standard commands (npm test, npm run build).
Only specify if different, e.g., "pytest -v" or "make release"

Question 4: Nested CLAUDE.md Selection

I detected these directories. Which should have their own CLAUDE.md?

Detected:
- [x] tests/
- [x] src/components/
- [x] src/api/
- [ ] src/utils/
- [ ] scripts/

Select all that would benefit from contextual guidance.
(Default: tests/ and any domain-specific src/ directories)

For Existing CLAUDE.md Migration, Ask:

Found existing CLAUDE.md ({N} lines). Options:

1. Analyze and migrate - Extract content to nested CLAUDE.md files
2. Replace - Create fresh minimal structure (backup original)
3. Cancel - Keep existing structure

Phase 3: Analysis (Existing CLAUDE.md Only)

Goal: Categorize existing content for migration.

Metrics to Calculate:

Content Categorization:

Map existing content to potential nested locations:

Analysis Output Template:

## CLAUDE.md Analysis

### Metrics
| Metric | Value | Status |
|--------|-------|--------|
| Total Lines | 245 | MIGRATE |
| Instructions | 52 | HIGH |
| Hardcoded Paths | 18 | WARNING |

### Content Migration Plan
| Lines | Content | Destination |
|-------|---------|-------------|
| 1-15 | Project description | Root CLAUDE.md |
| 20-45 | Testing patterns | tests/CLAUDE.md |
| 50-95 | Component conventions | src/components/CLAUDE.md |
| 100-140 | API design patterns | src/api/CLAUDE.md |
| 145-180 | General coding style | src/CLAUDE.md |
| 185-245 | Outdated/stale content | REMOVE (with approval) |

Phase 4: Generation

Goal: Create minimal root + nested CLAUDE.md files.

Parallel Generation: When creating multiple nested CLAUDE.md files, generate them in parallel (multiple Write calls in one message).

Root CLAUDE.md Template (50-100 lines target)

# {Project Name}

{One-sentence description}

## Development

{Package manager line - ONLY if not npm}

{Non-standard commands - ONLY if provided}

## Where to Look

| Task | Location |
|------|----------|
| Add new API endpoint | src/api/ |
| Add UI component | src/components/ |
| Add utility function | src/utils/ |
| Write tests | tests/ |
| Add service integration | src/services/ |
| Modify build/deploy | scripts/ |

## Agent Workflow

Explore finds -> Librarian reads -> You plan -> Worker implements -> Validator checks

When delegating to agents:
- Use positive constraints ("ensure X") not negative ("don't do Y")
- Include context, expected output, acceptance criteria
- Launch independent tasks in parallel

## Guidance

Context-specific guidance lives in nested CLAUDE.md files throughout the repo.
These load automatically when you work in those directories.
Closest CLAUDE.md to the file being edited takes precedence.

Conditional Rules:

• Omit "Package manager" if npm
• Omit "Development" section if no non-standard info
• Customize "Where to Look" table based on actual project structure
• Never include coding conventions (put in src/CLAUDE.md or nested)

Nested CLAUDE.md Templates

Each nested file: 30-80 lines max. Telegraphic style. Domain-specific anti-patterns required.

tests/CLAUDE.md Template (30-50 lines):

# Testing

## Framework

{e.g., Jest, Vitest, pytest, Go testing}

## Patterns

- Test files: `*.test.ts` or `*.spec.ts`
- Describe blocks mirror module structure
- One assertion per test when practical

## Running Tests

{test command}

## Mocking

- Prefer real implementations over mocks
- Mock only: network, time, randomness, file system

## Coverage

Focus on behavior, not percentage. Critical paths need thorough coverage.

## Anti-Patterns

- Don't mock what you can test directly
- Don't test implementation details
- Don't share mutable state between tests
- Don't write tests that pass when code is broken

src/components/CLAUDE.md Template (30-50 lines):

# Components

## Structure

ComponentName/ index.ts # Re-export ComponentName.tsx # Implementation ComponentName.test.tsx

## Naming

- Components: PascalCase
- Props interface: `{ComponentName}Props`
- Files match component name

## Patterns

- Prefer composition over inheritance
- Extract hooks for reusable logic
- Props over context for explicit dependencies

## Exports

{Named exports / default exports preference}

## Anti-Patterns

- Don't use inline styles when CSS modules exist
- Don't prop-drill beyond 2 levels - use context
- Don't mix data fetching with presentation
- Don't create god components (>200 lines)

src/api/CLAUDE.md Template (30-50 lines):

# API Layer

## Design

- RESTful endpoints / GraphQL / tRPC
- Consistent error response format
- Request validation at boundary

## Patterns

- Controllers handle HTTP concerns only
- Business logic lives in services
- Validate input, sanitize output

## Error Handling

{error response format example}

## Authentication

{auth pattern if applicable}

## Anti-Patterns

- Don't expose internal errors to clients
- Don't skip input validation
- Don't leak sensitive data in responses
- Don't mix auth logic with business logic

src/hooks/CLAUDE.md Template (30-50 lines):

# Hooks

## Naming

- Prefix with `use`: `useAuth`, `useForm`, `useFetch`
- Name describes the capability, not implementation

## Patterns

- Single responsibility per hook
- Return stable references (useMemo, useCallback)
- Handle cleanup in useEffect

## Testing

- Test hooks via @testing-library/react-hooks
- Test state transitions, not implementation

## Common Hooks

{list project-specific hooks if established}

## Anti-Patterns

- Don't call hooks conditionally
- Don't create hooks with side effects on mount without cleanup
- Don't return unstable references
- Don't nest hooks more than 2 levels deep

src/utils/CLAUDE.md Template (30-50 lines):

# Utilities

## When to Create

Create a utility when:
- Logic is used in 3+ places
- Function is pure (no side effects)
- Behavior is well-defined and testable

## Patterns

- Pure functions preferred
- No dependencies on external state
- Comprehensive input validation
- Clear, descriptive names

## Naming

- Functions: camelCase, verb-first (`formatDate`, `parseInput`)
- Files: kebab-case or camelCase (match project convention)

## Testing

Every utility function needs tests. They're isolated and easy to test.

## Anti-Patterns

- Don't create utils for single-use logic
- Don't add side effects to pure functions
- Don't create "misc" or "helpers" grab-bags
- Don't depend on global state

src/services/CLAUDE.md Template (30-50 lines):

# Services

## Purpose

Services encapsulate external integrations and complex business logic.

## Patterns

- One service per external dependency
- Services are injectable/mockable
- Handle retries, timeouts, circuit breaking

## Structure

services/ userService.ts authService.ts emailService.ts

## Error Handling

- Wrap external errors in domain-specific errors
- Log at service boundary
- Never expose internal details to callers

## Anti-Patterns

- Don't call services from other services directly - use orchestrator
- Don't mix HTTP/DB concerns in same service
- Don't retry without backoff
- Don't swallow errors silently

src/CLAUDE.md Template (30-50 lines):

# Source Code

## Style

- {indent style}
- {semicolons yes/no}
- {quote style}

## Imports

// 1. External packages // 2. Internal aliases (@/) // 3. Relative imports // 4. Types (import type)

## Naming

- Files: {convention}
- Functions: camelCase
- Classes: PascalCase
- Constants: SCREAMING_SNAKE

## Patterns

- Prefer early returns
- Explicit over implicit
- Composition over inheritance

## Anti-Patterns

- Don't use `any` - use `unknown` and narrow
- Don't nest callbacks >2 levels
- Don't mutate function arguments
- Don't use magic numbers without constants

scripts/CLAUDE.md Template (30-50 lines):

# Scripts

## Purpose

Build, deployment, and development automation scripts.

## Conventions

- Scripts are executable: `chmod +x script.sh`
- Include usage comment at top
- Exit codes: 0 success, 1 error
- Use `set -euo pipefail` for bash scripts

## Running

./scripts/{script-name}.sh

or

npm run {script-name}

## Adding New Scripts

1. Create in scripts/ directory
2. Add npm script alias if frequently used
3. Document purpose in script header

## Anti-Patterns

- Don't hardcode paths - use variables
- Don't skip error handling
- Don't assume dependencies exist - check first
- Don't write scripts without idempotency

Phase 5: Execution

For Fresh Projects:

1. Create root CLAUDE.md
2. Create nested CLAUDE.md files in detected/selected directories (parallel writes)
3. Report created files

For Migration:

1. Create backup: cp CLAUDE.md CLAUDE.md.backup
2. Create nested CLAUDE.md files with migrated content
3. Rewrite root CLAUDE.md to minimal version
4. If .claude/docs/ exists, offer to migrate or remove
5. Present before/after comparison

For --fresh Flag:

1. Find all existing CLAUDE.md files: find . -name "CLAUDE.md" -type f
2. Delete all found CLAUDE.md files
3. Run fresh project workflow

Deduplication Check (Required):

Before finalizing, review all generated files:

1. List all rules in root CLAUDE.md
2. Check each nested file for duplicates
3. Remove any duplicated guidance from children
4. Only domain-specific additions remain in nested files

Before/After Report:

## Migration Complete

### Before
- Root CLAUDE.md: 245 lines
- Nested CLAUDE.md: 0 files
- .claude/docs/: 3 files

### After
- Root CLAUDE.md: 85 lines (65% reduction)
- Nested CLAUDE.md: 4 files
  - tests/CLAUDE.md (35 lines)
  - src/components/CLAUDE.md (42 lines)
  - src/api/CLAUDE.md (38 lines)
  - src/CLAUDE.md (30 lines)

### Backup
Original preserved at CLAUDE.md.backup

### Benefit
Root context cost reduced from 245 to 85 lines.
Domain guidance loads only when working in those directories.

Behavior Rules

MUST DO

• Detect existing CLAUDE.md before creating
• Ask before overwriting any existing CLAUDE.md files
• Create backup before modifying existing root CLAUDE.md
• Keep root CLAUDE.md 50-150 lines
• Keep nested CLAUDE.md files 30-80 lines each
• Include domain-specific anti-patterns in each nested file
• Run deduplication check before finalizing
• Use parallel Task calls for exploration
• Use parallel Write calls for generation

MUST NOT

• Document hardcoded file paths (they go stale)
• Include standard commands (npm test, npm run build)
• Reference .claude/docs/ (old pattern)
• Create overly comprehensive documentation
• Create nested CLAUDE.md in directories that don't benefit
• Duplicate parent guidance in child files
• Execute agents sequentially when parallel is possible

SHOULD DO

• Explain the auto-loading behavior
• Explain "closest wins" precedence
• Suggest which directories would benefit from nested CLAUDE.md
• Offer to migrate .claude/docs/ content if it exists
• Provide line counts in final report
• Use scoring system for borderline directories

Content Location Decision

Error Handling

Directory Doesn't Exist

If selected directory doesn't exist:

1. Ask: "Directory {dir} doesn't exist. Create it with CLAUDE.md, or skip?"
2. Options: Create directory, Skip

Existing Nested CLAUDE.md

If nested CLAUDE.md already exists:

1. Read current content
2. Ask: "Found existing {dir}/CLAUDE.md. Merge, replace, or skip?"
3. Options: Merge (append new patterns), Replace (backup first), Skip

Permission Denied

If cannot write files:

1. Output generated content to chat
2. Say: "Could not write files. Here's the content to add manually:"

Post-Generation Message

After creating files:

Created nested CLAUDE.md structure:

Root:
  - CLAUDE.md ({N} lines)

Nested:
  - tests/CLAUDE.md ({N} lines) - Testing patterns
  - src/components/CLAUDE.md ({N} lines) - Component patterns
  - src/api/CLAUDE.md ({N} lines) - API patterns

How it works:
Claude automatically loads CLAUDE.md from any directory it reads.
When you work in src/api/, the API patterns load automatically.
Closest file to your edit takes precedence ("closest wins").

Next steps:
1. Review each CLAUDE.md and customize for your project
2. Add patterns as they emerge during development
3. Keep files focused and minimal (30-80 lines nested, 50-150 root)