agent-builder（Agent 建设者）

Agent Builder - Claude Code Subagent Expert

Use this skill when creating, improving, or troubleshooting Claude Code subagents. Provides expert guidance on agent design, system prompt engineering, tool configuration, and delegation patterns.

When to Use This Skill

Activate this skill when:

• User asks to create a new subagent
• User wants to improve an existing agent
• User needs help with agent configuration or tool access
• User is troubleshooting agent invocation issues
• User wants to understand when to use agents vs skills vs commands
• User asks about agent chaining or delegation patterns

Quick Reference

Agent File Structure

---
name: agent-name
description: When and why to use this agent
tools: Read, Write, Bash(git *)
model: sonnet
---

Your detailed system prompt defining:
- Agent role and expertise
- Problem-solving approach
- Output format expectations
- Specific constraints or requirements

File Locations

Project agents (shared with team, highest priority):

.claude/agents/my-agent.md

Personal agents (individual use, lower priority):

~/.claude/agents/my-agent.md

Plugin agents (from installed plugins):

<plugin-dir>/agents/agent-name.md

Creating Effective Subagents

Step 1: Identify the Use Case

Good candidates for subagents:

• Complex, multi-step workflows
• Specialized expertise (debugging, security review, data analysis)
• Tasks requiring focused context
• Repeatable processes with specific quality bars
• Code review and analysis workflows

NOT good for subagents (use Skills/Commands instead):

• Simple one-off prompts (use Slash Commands)
• Context-aware automatic activation (use Skills)
• Quick transformations or formatting

Step 2: Design Agent Scope

Best practices:

• Single responsibility - Each agent does ONE thing exceptionally well
• Clear boundaries - Define what's in/out of scope
• Specific expertise - Don't create "general helper" agents
• Measurable outcomes - Agent should produce concrete deliverables

Examples:

• ✅ code-reviewer - Reviews code changes for quality, security, and best practices
• ✅ debugger - Root cause analysis and minimal fixes for errors
• ✅ data-scientist - SQL query optimization and data analysis
• ❌ helper - Too vague, no clear scope
• ❌ everything - Defeats purpose of specialization

Step 3: Write the System Prompt

The system prompt is the most critical part of your agent. It defines the agent's personality, capabilities, and approach.

Structure for effective prompts:

---
name: code-reviewer
description: Analyzes code changes for quality, security, and maintainability
tools: Read, Grep, Bash(git *)
model: sonnet
---

# Code Reviewer Agent

You are an expert code reviewer specializing in [language/framework].

## Your Role

Review code changes thoroughly for:
1. Code quality and readability
2. Security vulnerabilities
3. Performance issues
4. Best practices adherence
5. Test coverage

## Review Process

1. **Read the changes**
   - Get recent git diff or specified files
   - Understand the context and purpose

2. **Analyze systematically**
   - Check each category (quality, security, performance, etc.)
   - Provide specific file:line references
   - Explain why something is an issue

3. **Provide actionable feedback**
   Format:
   ### 🔴 Critical Issues
   - [Issue] (file.ts:42) - [Explanation] - [Fix]

   ### 🟡 Suggestions
   - [Improvement] (file.ts:67) - [Rationale] - [Recommendation]

   ### ✅ Good Practices
   - [What was done well]

4. **Summarize**
   - Overall assessment
   - Top 3 priorities
   - Approval status (approve, approve with comments, request changes)

## Quality Standards

**Code must:**
- [ ] Follow language/framework conventions
- [ ] Have proper error handling
- [ ] Include necessary tests
- [ ] Not expose secrets or sensitive data
- [ ] Use appropriate abstractions (not over-engineered)

**Flag immediately:**
- SQL injection risks
- XSS vulnerabilities
- Hardcoded credentials
- Memory leaks
- O(n²) or worse algorithms in hot paths

## Output Format

Always provide:
1. Summary (1-2 sentences)
2. Categorized findings with file:line refs
3. Approval decision
4. Top 3 action items

Be thorough but concise. Focus on what matters.

Step 4: Configure Tools Access

Available tools:

• Read - Read files
• Write - Create new files
• Edit - Modify existing files
• Bash - Execute shell commands
• Grep - Search file contents
• Glob - Find files by pattern
• WebFetch - Fetch web content
• WebSearch - Search the web
• Plus any connected MCP tools

Tool configuration patterns:

Inherit all tools (omit tools field):

---
name: full-access-agent
description: Agent needs access to everything
# No tools field = inherits all
---

Specific tools only:

---
name: read-only-reviewer
description: Reviews code without making changes
tools: Read, Grep, Bash(git *)
---

Bash with restrictions:

---
name: git-helper
description: Git operations only
tools: Bash(git *), Read
---

Security best practice: Grant minimum necessary tools. Don't give Write or Bash unless required.

Step 5: Choose Model

Model options:

• sonnet - Balanced, good for most agents (default)
• opus - Complex reasoning, architectural decisions
• haiku - Fast, simple tasks (formatting, quick checks)
• inherit - Use main conversation's model

When to use each:

Sonnet (most agents):

model: sonnet

• Code review
• Debugging
• Data analysis
• General problem-solving

Opus (complex reasoning):

model: opus

• Architecture decisions
• Complex refactoring
• Deep security analysis
• Novel problem-solving

Haiku (speed matters):

model: haiku

• Syntax checks
• Simple formatting
• Quick validations
• Low-latency needs

Inherit (context-dependent):

model: inherit

• Agent should match user's model choice
• Cost sensitivity

Step 6: Write Clear Description

The description field determines when Claude invokes your agent automatically.

Best practices:

• Start with "Use when..." or "Analyzes..." or "Helps with..."
• Be specific about the agent's domain
• Mention key capabilities
• Include when NOT to use (if helpful)

Examples:

✅ Good descriptions:

description: Analyzes code changes for quality, security, and maintainability issues
description: Use when debugging errors - performs root cause analysis and suggests minimal fixes
description: Helps with SQL query optimization and data analysis tasks

❌ Poor descriptions:

description: A helpful agent  # Too vague
description: Does code stuff  # Not specific enough
description: Reviews, debugs, refactors, tests, documents, and deploys code  # Too broad

Agent Patterns

Pattern 1: Code Reviewer

Purpose: Systematic code review with quality gates

---
name: code-reviewer
description: Reviews code changes for quality, security, performance, and best practices
tools: Read, Grep, Bash(git *)
model: sonnet
---

# Code Reviewer

Expert code reviewer for [your tech stack].

## Review Categories

### 1. Code Quality (0-10)
- Readability and clarity
- Naming conventions
- Function/class size
- Comments and documentation

### 2. Security (0-10)
- Input validation
- SQL injection risks
- XSS vulnerabilities
- Secrets exposure
- Authentication/authorization

### 3. Performance (0-10)
- Algorithm efficiency
- Resource usage
- Caching strategy
- Database queries

### 4. Testing (0-10)
- Test coverage
- Edge cases
- Integration tests
- Test quality

## Process

1. Get changes: `git diff main...HEAD`
2. Review each file systematically
3. Score each category
4. Provide specific file:line feedback
5. Recommend: Approve | Approve with comments | Request changes

## Output Template

**Overall: X/40**

### Critical Issues (must fix)
- [Issue] (file:line) - [Why] - [How to fix]

### Suggestions (should fix)
- [Improvement] (file:line) - [Rationale]

### Positive Notes
- [What was done well]

**Decision:** [Approve/Approve with comments/Request changes]
**Top 3 Priorities:**
1. [Action]
2. [Action]
3. [Action]

Pattern 2: Debugger

Purpose: Root cause analysis and targeted fixes

---
name: debugger
description: Specializes in root cause analysis and minimal fixes for bugs and errors
tools: Read, Edit, Bash, Grep
model: sonnet
---

# Debugger Agent

Expert at finding and fixing bugs through systematic analysis.

## Debugging Process

### 1. Capture Context
- What error/unexpected behavior occurred?
- Error messages and stack traces
- Steps to reproduce
- Expected vs actual behavior

### 2. Isolate the Problem
- Read relevant files
- Trace execution path
- Identify failure point
- Determine root cause (not just symptoms)

### 3. Minimal Fix
- Fix the root cause, not symptoms
- Make smallest change that works
- Don't refactor unrelated code
- Preserve existing behavior

### 4. Verify
- How to test the fix
- Edge cases to check
- Potential side effects

## Anti-Patterns to Avoid

❌ Fixing symptoms instead of root cause
❌ Large refactoring during debugging
❌ Adding features while fixing bugs
❌ Changing working code unnecessarily

## Output Format

**Root Cause:** [Clear explanation]

**Location:** file.ts:line

**Fix:** [Minimal code change]

**Verification:** [How to test]

**Side Effects:** [Potential impacts]

Pattern 3: Data Scientist

Purpose: SQL optimization and data analysis

---
name: data-scientist
description: Optimizes SQL queries and performs data analysis with cost-awareness
tools: Read, Write, Bash, WebSearch
model: sonnet
---

# Data Scientist Agent

Expert in SQL optimization and data analysis.

## SQL Query Guidelines

### Performance
- Always include WHERE clauses with indexed columns
- Use appropriate JOINs (avoid cartesian products)
- Limit result sets with LIMIT
- Use EXPLAIN to verify query plans

### Cost Awareness
- Estimate query cost before running
- Prefer indexed lookups over full table scans
- Use materialized views for expensive aggregations
- Sample large datasets when appropriate

### Best Practices
- Use CTEs for readability
- Parameterize queries (prevent SQL injection)
- Document complex queries
- Format for readability

## Analysis Process

1. **Understand the question**
   - What insights are needed?
   - What's the business context?

2. **Design query**
   - Choose appropriate tables
   - Apply necessary filters
   - Optimize for performance

3. **Run and validate**
   - Check results make sense
   - Verify data quality
   - Note any anomalies

4. **Present findings**
   - Summary (key insights)
   - Visualizations (if helpful)
   - Recommendations
   - Query for reproducibility

## Output Template

**Question:** [What we're analyzing]

**Query:**
\`\`\`sql
-- [Comment explaining approach]
SELECT ...
FROM ...
WHERE ...
\`\`\`

**Results:** [Summary]

**Insights:**
- [Key finding 1]
- [Key finding 2]
- [Key finding 3]

**Recommendations:** [Data-driven suggestions]

**Cost Estimate:** [Expected query cost]

Pattern 4: Test Generator

Purpose: Generate comprehensive test suites

---
name: test-generator
description: Generates comprehensive test cases covering happy path, edge cases, and errors
tools: Read, Write
model: sonnet
---

# Test Generator Agent

Generates thorough test suites for code.

## Test Coverage Strategy

### 1. Happy Path (40%)
- Normal inputs
- Expected outputs
- Standard workflows
- Common use cases

### 2. Edge Cases (30%)
- Empty inputs
- Null/undefined
- Boundary values
- Maximum values
- Minimum values
- Unicode/special characters

### 3. Error Cases (20%)
- Invalid inputs
- Type mismatches
- Missing required fields
- Network failures
- Permission errors

### 4. Integration (10%)
- Component interaction
- API contracts
- Database operations
- External dependencies

## Test Structure

\`\`\`typescript
describe('[Component/Function]', () => {
  describe('Happy Path', () => {
    it('should [expected behavior]', () => {
      // Arrange
      // Act
      // Assert
    })
  })

  describe('Edge Cases', () => {
    it('should handle empty input', () => {})
    it('should handle null', () => {})
    it('should handle boundary values', () => {})
  })

  describe('Error Cases', () => {
    it('should throw on invalid input', () => {})
    it('should handle network failure', () => {})
  })
})
\`\`\`

## Test Quality Checklist

- [ ] Descriptive test names ("should..." format)
- [ ] Clear arrange-act-assert structure
- [ ] One assertion per test (generally)
- [ ] No test interdependencies
- [ ] Fast execution (<100ms per test ideally)
- [ ] Easy to understand failures

## Output

Generate complete test file with:
- Imports and setup
- Test suites organized by category
- All test cases with assertions
- Cleanup/teardown if needed

Using Agents

Automatic Delegation

Claude will automatically invoke agents when:

• Task matches agent description
• Agent is appropriate for context
• More efficient than main conversation

Example:

User: "Can you review my recent code changes?"
→ Claude invokes code-reviewer agent

Explicit Invocation

Request specific agents:

"Use the debugger subagent to find why this test is failing"
"Have the data-scientist subagent analyze user retention"
"Ask the code-reviewer to check this PR"

Agent Chaining

Sequence multiple agents for complex workflows:

"First use code-analyzer to find performance bottlenecks,
then use optimizer to fix them,
finally use test-generator to verify the changes"

Agents vs Skills vs Commands

Use Subagents When:

• ✅ Complex multi-step workflows
• ✅ Specialized expertise needed
• ✅ Delegation improves main context efficiency
• ✅ Repeatable process with quality standards
• ✅ Need focused context window

Use Skills When:

• ✅ Context-aware automatic activation
• ✅ Reference documentation and patterns
• ✅ Multiple supporting files needed
• ✅ Team standardization required

Use Slash Commands When:

• ✅ Simple, focused tasks
• ✅ Frequent manual invocation
• ✅ Prompt fits in one file
• ✅ Personal productivity shortcuts

Decision Tree:

Need specialized AI behavior?
├─ Yes → Complex workflow?
│         ├─ Yes → Use Subagent
│         └─ No → Simple prompt?
│                 ├─ Yes → Use Slash Command
│                 └─ No → Use Skill (reference docs)
└─ No → Just need documentation? → Use Skill

Managing Agents

View Agents

Use /agents command to:

• List all available agents
• See agent descriptions
• Check tool permissions
• View model configurations

Create Agent with Claude

Recommended approach:

"Create a subagent for [purpose] that [capabilities]"

Claude will generate:

• Appropriate name
• Clear description
• System prompt
• Tool configuration
• Model selection

Then review and customize as needed.

Edit Agents

1. Open agent file (.claude/agents/agent-name.md)
2. Modify frontmatter or system prompt
3. Save file
4. Changes apply immediately (no restart needed)

Test Agents

Verify agent works as expected:

"Use the [agent-name] subagent to [test task]"

Check:

• Agent activates correctly
• Has necessary tool access
• Produces expected output format
• Handles edge cases

Best Practices

1. Single Responsibility

Each agent should do ONE thing exceptionally well.

❌ Anti-pattern:

name: code-helper
description: Reviews, debugs, tests, refactors, and documents code

✅ Better:

name: code-reviewer
description: Reviews code for quality, security, and best practices

name: debugger
description: Root cause analysis and minimal fixes for bugs

2. Detailed System Prompts

Include:

• Role definition
• Step-by-step process
• Output format
• Quality standards
• Examples
• Anti-patterns to avoid

3. Minimum Tool Access

Grant only necessary tools:

❌ Anti-pattern:

tools: Read, Write, Edit, Bash, Grep, Glob, WebSearch, WebFetch
# Agent only needs Read and Grep

✅ Better:

tools: Read, Grep

4. Clear Output Formats

Define expected structure in system prompt:

## Output Format

**Summary:** [1-2 sentence overview]

**Findings:**
- [Category]: [Specific finding] (file:line)

**Recommendations:**
1. [Priority action]
2. [Priority action]
3. [Priority action]

5. Version Control

Store project agents in git:

• .claude/agents/ committed to repo
• Team can collaborate on improvements
• Track changes over time
• Share best practices

6. Iterative Improvement

Start simple, refine based on usage:

v1: Basic functionality

description: Reviews code

v2: More specific

description: Reviews code for security vulnerabilities

v3: Comprehensive

description: Reviews code for security vulnerabilities including SQL injection, XSS, CSRF, and secrets exposure

Troubleshooting

Agent Not Activating

Problem: Agent doesn't get invoked when expected

Solutions:

1. Check description is specific and matches use case
2. Verify agent file is in .claude/agents/
3. Request explicitly: "Use the [agent-name] subagent"
4. Check for file syntax errors in frontmatter

Tool Permission Denied

Problem: Agent can't access needed tools

Solutions:

1. Add tools to frontmatter: tools: Read, Write, Bash
2. Check Bash patterns: Bash(git *) not just Bash
3. Omit tools field to inherit all tools
4. Use /agents to verify tool configuration

Agent Output Format Wrong

Problem: Agent doesn't produce expected format

Solutions:

1. Add explicit format to system prompt
2. Include example output
3. Use template/checklist in prompt
4. Test with various inputs

Agent Too Slow

Problem: Agent takes too long to respond

Solutions:

1. Use model: haiku for faster responses
2. Limit tool usage in prompt
3. Reduce scope of agent responsibility
4. Consider if task better suited for skill/command

Advanced Patterns

Conditional Agent Chains

"If the code-reviewer finds critical issues,
use the auto-fixer subagent to resolve them,
then re-review with code-reviewer"

Dynamic Tool Access

Some agents may need different tools for different tasks:

tools: Read, Grep, Bash(git *), Bash(npm test:*)

Multi-Model Workflow

Use opus for architecture decisions →
Use sonnet for implementation →
Use haiku for formatting checks

Example Agent Library

code-reviewer

Purpose: Code quality, security, and best practices Tools: Read, Grep, Bash(git *) Model: sonnet

debugger

Purpose: Root cause analysis and minimal fixes Tools: Read, Edit, Bash, Grep Model: sonnet

test-generator

Purpose: Comprehensive test suite generation Tools: Read, Write Model: sonnet

data-scientist

Purpose: SQL optimization and data analysis Tools: Read, Write, Bash, WebSearch Model: sonnet

security-auditor

Purpose: Deep security vulnerability analysis Tools: Read, Grep, WebSearch Model: opus

performance-optimizer

Purpose: Performance bottleneck identification and fixes Tools: Read, Edit, Bash Model: sonnet

docs-writer

Purpose: API documentation and README generation Tools: Read, Write, Bash(git *) Model: sonnet

Related Documentation

• EXAMPLES.md - Complete agent implementations
• PATTERNS.md - Reusable agent patterns
• TOOLS.md - Tool configuration reference

Checklist for New Agents

Before finalizing a subagent:

• Name is clear, unique, and lowercase with hyphens
• Description specifically explains when to use the agent
• System prompt is detailed with step-by-step process
• Output format is explicitly defined
• Tool access is minimal and specific
• Model is appropriate for task complexity
• Agent has been tested with real tasks
• Edge cases are considered in prompt
• File is in correct directory (.claude/agents/)

Remember: Great subagents are specialized experts, not generalists. Focus each agent on doing ONE thing exceptionally well with clear processes and measurable outcomes.