subagent-creation（子 Agent 创建）

Claude Code Subagent Creation

When to use this skill

• Defining specialized AI experts for specific tasks
• Creating reusable agent configurations for team workflows
• Implementing task delegation patterns
• Setting up automated code review, debugging, or analysis workflows
• Creating agents with custom prompts and tool permissions

Instructions

Step 1: Understanding Subagents

Claude Code subagents are pre-configured AI experts that the main Claude delegates work to.

Core benefits:

• Context isolation: Each agent has separate 200K token context
• Specialized expertise: Focused prompts for specific domains
• Reusability: Share agents via Git across projects
• Flexible permissions: Control which tools each agent can use
• No nesting: Prevents infinite loops - subagents cannot spawn more subagents

Agent types:

• Built-in agents: Explore (read-only, Haiku), Plan (research), General-purpose (Sonnet)
• Custom agents: User-defined with custom prompts and permissions

Step 2: Agent Configuration File

Subagents are defined as Markdown files with YAML frontmatter.

File location (priority order):

1. Project-level: .claude/agents/{agent-name}.md
2. User-level: ~/.claude/agents/{agent-name}.md

File format:

---
name: code-reviewer
description: Review code changes for quality, security, and best practices. Use immediately after code changes.
tools: [Read, Grep, Glob, Bash, LSP]
model: inherit
---

# Code Reviewer

You are a senior code reviewer with expertise in:
- Code quality and maintainability
- Security vulnerabilities
- Performance optimization
- Best practices and design patterns

## Review Checklist

- [ ] Does the code follow project conventions?
- [ ] Are there any security vulnerabilities?
- [ ] Is the code readable and maintainable?
- [ ] Are there performance concerns?
- [ ] Are tests adequate?
- [ ] Is documentation complete?

## Review Guidelines

1. **Prioritize critical issues**: Security bugs, data races, memory leaks
2. **Be constructive**: Provide clear explanations and suggestions
3. **Consider trade-offs**: Don't optimize prematurely
4. **Reference standards**: Link to relevant docs/style guides

## Output Format

Critical Issues (Must Fix)

• Issue: Description File: path/to/file.ts:123 Suggestion: How to fix

Suggestions (Should Fix)

...

Nice to Have

...

Step 3: YAML Frontmatter Configuration

Required fields:

Optional fields:

Model options:

• inherit: Use same model as main Claude
• sonnet: Claude 3.5 Sonnet
• haiku: Claude 3 Haiku (faster, cheaper)
• opus: Claude 3 Opus

Tool options: Read, Write, Edit, Grep, Glob, Bash, LSP tools

Step 4: System Prompt Writing

Best practices:

1. Define role clearly: "You are a [role] with expertise in [domains]"
2. Include checklist: Concrete steps for evaluation/execution
3. Provide examples: Show input-output pairs
4. Specify format: Define exactly how results should be presented
5. Set boundaries: What the agent should NOT do

Example prompt structure:

You are a [ROLE] specializing in [DOMAIN].

## Responsibilities
- Task 1
- Task 2

## Process
1. Step 1: ...
2. Step 2: ...

## Checklist
- [ ] Item 1
- [ ] Item 2

## Output Format
[Template]

## Constraints
- Constraint 1
- Constraint 2

Step 5: Creating Common Agent Types

Agent 1: Code Reviewer

.claude/agents/code-reviewer.md:

---
name: code-reviewer
description: Review code changes for quality, security, and best practices.
tools: [Read, Grep, Glob, LSP]
model: inherit
---

# Code Reviewer

Review code changes focusing on:
1. **Security**: Authentication, authorization, injection risks
2. **Quality**: Clean code principles, maintainability
3. **Performance**: Time/space complexity, bottlenecks
4. **Tests**: Coverage, edge cases

## Priority Levels
- **Critical**: Security vulnerabilities, data corruption
- **High**: Performance issues, logic errors
- **Medium**: Code smell, missing tests
- **Low**: Style, minor improvements

Agent 2: Debugger

.claude/agents/debugger.md:

---
name: debugger
description: Analyze errors and implement fixes. Use when encountering bugs or failures.
tools: [Read, Write, Edit, Bash, Grep, Glob, LSP]
model: inherit
---

# Debugger

Systematically debug issues:
1. **Understand the problem**: Read error messages, logs
2. **Reproduce**: Try to recreate the issue
3. **Analyze**: Identify root cause
4. **Fix**: Implement minimal fix
5. **Verify**: Confirm the fix works
6. **Test**: Check for regressions

## Debugging Strategy
- Use `grep` to search for error-related code
- Check recent changes with `git log`
- Add logging if needed to trace execution
- Fix one issue at a time
- Verify fix doesn't break existing functionality

Agent 3: Test Writer

.claude/agents/test-writer.md:

---
name: test-writer
description: Write comprehensive unit and integration tests for new code.
tools: [Read, Write, Edit, Grep, Glob]
model: inherit
---

# Test Writer

Write tests following these principles:
1. **AAA pattern**: Arrange, Act, Assert
2. **Descriptive names**: Test names explain what they verify
3. **One assertion per test**: Clear failure reasons
4. **Test happy path**: Main functionality
5. **Test edge cases**: Boundary conditions, nulls, errors
6. **Mock external dependencies**: Isolate code under test

## Test Coverage Goals
- Unit tests: 80%+ coverage
- Integration tests: Critical user flows
- E2E tests: Key user journeys

## Framework-Specific Guidelines
**Jest**: Use `describe`, `test`, `expect`, `beforeEach`, `afterEach`
**Pytest**: Use `def test_`, `assert`, `@pytest.fixture`
**Go testing**: Use `TestXxx`, `t.Run`, `assert.Equal`

Agent 4: Performance Analyzer

.claude/agents/performance-analyzer.md:

---
name: performance-analyzer
description: Analyze code performance and identify optimization opportunities.
tools: [Read, Grep, Glob, LSP]
model: sonnet
---

# Performance Analyzer

Focus on:
1. **Time complexity**: Algorithm efficiency
2. **Space complexity**: Memory usage
3. **Database queries**: N+1 queries, missing indexes
4. **I/O operations**: File system, network calls
5. **Caching**: Missed caching opportunities

## Analysis Steps
1. Profile code to find hotspots
2. Review algorithm choices
3. Check database query patterns
4. Look for redundant computations
5. Identify parallelization opportunities

## Report Format

Performance Issues Found

Critical (High Impact)

• Impact: [description] Suggestion: [optimization]

Moderate (Medium Impact)

...

Low (Minor)

...

Agent 5: Documentation Writer

.claude/agents/doc-writer.md:

---
name: doc-writer
description: Write clear, comprehensive documentation for code, APIs, and features.
tools: [Read, Write, Edit, Grep, Glob]
model: inherit
---

# Documentation Writer

Write documentation that is:
- **Clear**: Simple language, avoid jargon
- **Complete**: Cover all use cases
- **Accurate**: Keep docs in sync with code
- **Actionable**: Include examples
- **Searchable**: Use consistent terminology

## Documentation Types
- **README**: Project overview, setup, usage
- **API docs**: Endpoints, parameters, examples
- **Code comments**: Why, not what
- **Changelog**: Version history, breaking changes

## Writing Guidelines
1. Start with user goals
2. Provide examples for each feature
3. Link related documentation
4. Update docs when code changes
5. Use active voice and present tense

Step 6: CLI Configuration

Create agents via CLI for automation:

Single agent:

claude --agents '{
  "code-reviewer": {
    "description": "Review code changes for quality and security",
    "tools": ["Read", "Grep", "Glob", "LSP"],
    "model": "inherit"
  }
}'

Multiple agents:

claude --agents '{
  "code-reviewer": {"description": "...", "tools": ["Read"]},
  "debugger": {"description": "...", "tools": ["Read", "Write", "Edit"]},
  "test-writer": {"description": "...", "tools": ["Read", "Write"]}
}'

Add hooks for automation:

# PostToolUse hook - automatically invoke debugger on errors
claude --hooks '{
  "PostToolUse": {
    "onError": "debugger"
  }
}'

Step 7: Using Subagents

Explicit Invocation

Directly call an agent:

"Use code-reviewer to review the recent authentication changes."

"Invoke debugger agent to fix the failing test."

Automatic Delegation

Claude automatically delegates based on agent descriptions:

"Refactor the authentication logic for better security."
→ Claude delegates to: code-reviewer (security expert)

"Fix the database connection timeout error."
→ Claude delegates to: debugger (error fixing)

Agent Chaining

Chain multiple agents for complex tasks:

"Use performance-analyzer to identify bottlenecks, then debugger to fix them."

"Let code-reviewer check the changes, then doc-writer update the documentation."

Resume Previous Context

Resume a previous agent session:

"Resume the code-reviewer session with agentId abc123 to continue where we left off."

Step 8: Version Control

Share agents via Git:

Commit agents:

cd /path/to/project
git add .claude/agents/
git commit -m "feat: add code-reviewer and debugger agents"
git push

Clone project with agents:

git clone https://github.com/myorg/project.git
# Agents are automatically available in .claude/agents/

Examples

Example 1: Complete Agent Creation Workflow

# 1. Create project-level agents directory
mkdir -p .claude/agents

# 2. Create code-reviewer agent
cat > .claude/agents/code-reviewer.md << 'EOF'
---
name: code-reviewer
description: Review code changes for quality, security, and best practices.
tools: [Read, Grep, Glob, LSP]
model: inherit
---

# Code Reviewer

Review code focusing on security, quality, performance, and tests.

## Priority
- Critical: Security vulnerabilities, data corruption
- High: Performance issues, logic errors
- Medium: Code smell, missing tests
- Low: Style improvements
EOF

# 3. Commit to Git
git add .claude/agents/code-reviewer.md
git commit -m "feat: add code-reviewer subagent"
git push

Example 2: Automated Code Review Workflow

Scenario: After completing a feature, automatically review code.

Setup:

# Create post-commit hook
cat > .git/hooks/post-commit << 'EOF'
#!/bin/bash
claude --agents '{
  "code-reviewer": {
    "description": "Review HEAD commit for quality and security",
    "tools": ["Read", "Grep", "Glob", "LSP"],
    "model": "inherit"
  }
}' << INPUT
Review the changes in the most recent commit.
INPUT
EOF
chmod +x .git/hooks/post-commit

Usage:

git commit -m "feat: add user authentication"
# Post-commit hook automatically invokes code-reviewer

Example 3: Multi-Agent Pipeline

Scenario: Code → Review → Test → Document

In Claude Code:

"Here's the new payment processing code I just wrote.

Use the following agents in sequence:
1. code-reviewer - Check for security issues and quality
2. test-writer - Write unit tests for the payment flow
3. doc-writer - Update API documentation

Return a summary of all findings."

Example 4: Specialized Domain Agent

.claude/agents/database-expert.md:

---
name: database-expert
description: Design and optimize database schemas, queries, and migrations.
tools: [Read, Write, Edit, Grep, Glob, Bash]
model: sonnet
---

# Database Expert

Expertise in:
- SQL design (PostgreSQL, MySQL, SQLite)
- NoSQL design (MongoDB, Redis, DynamoDB)
- Query optimization and indexing
- Migration strategies
- Data modeling and normalization

## Review Checklist
- [ ] Schema normalized? (3NF for relational DBs)
- [ ] Appropriate indexes?
- [ ] Query performance acceptable?
- [ ] Foreign keys/constraints defined?
- [ ] Migration reversible?
- [ ] Backup strategy documented?

## Optimization Tips
1. Use EXPLAIN ANALYZE to analyze query plans
2. Create composite indexes for multi-column WHERE clauses
3. Avoid SELECT * in production
4. Use connection pooling
5. Implement read replicas for read-heavy workloads

Design Principles

1. Single responsibility: Each agent has one focused role
2. Minimal permissions: Only grant tools actually needed
3. Detailed prompts: Include checklists, steps, examples
4. Model selection: Choose model based on task complexity
5. Version control: Track agent configurations in Git
6. Reusability: Design agents to work across projects
7. Clear descriptions: Help Claude understand when to delegate

Common Patterns

Best Practices

1. Start with tools: Define what the agent needs, not what you think it needs
2. Inherit model: Use model: inherit unless you need specific behavior
3. Test prompts: Verify agents work as expected before relying on them
4. Iterate: Improve agent prompts based on real usage
5. Document agent usage: Add README notes on when/how to use each agent
6. Use built-in agents: Explore, Plan, General-purpose cover many common cases

Common Pitfalls

• Too broad scope: Agents should focus on one domain, not do everything
• Missing constraints: Specify what agents should NOT do
• Over-permissive tools: Grant minimal tools needed for the task
• Unclear descriptions: Make descriptions specific so Claude knows when to delegate
• No examples: Show input/output patterns in prompts

References

• Claude Code Documentation
• Agent Configuration Guide
• YouTube: Subagents Tutorial