Claude Code Workflow Studio

Use Case: AI Code Review Workflow Design

公開日: 2024年2月1日 著者: breaking-brake

Learn how to design an effective code review workflow with Claude Code Workflow Studio. Achieve quality improvement and review efficiency.

Introduction

Code review is a critical quality assurance process, but it’s time-consuming and prone to oversights. With an AI review workflow designed in Claude Code Workflow Studio, you can conduct consistent, high-quality reviews efficiently.

Challenges

Typical review challenges development teams face:

  • Review Delays: Hard to secure reviewer time, delaying merges
  • Quality Variance: Check items and strictness vary by reviewer
  • Missed Checks: Overlooking security and performance issues
  • Knowledge Gaps: No reviewers expert in specific domains

Solution: Visual Review Workflow

Design review flow in Claude Code Workflow Studio to achieve:

  • Consistency: Review from same perspectives every time
  • Speed: AI provides immediate feedback
  • Comprehensiveness: Covers security, performance, and quality
  • Flexibility: Selectable review levels and targets

Workflow Design Example

Overall Structure

[Start]
   ↓
[Code Scanner] ← Get changed files
   ↓
[Choose Review Level] ← Select review strictness
   ↓
[Security Reviewer] ← Security check
   ↓
[Performance Reviewer] ← Performance check
   ↓
[Quality Reviewer] ← Code quality check
   ↓
[Summary Generator] ← Summarize review results
   ↓
[End]

Step 1: Code Scanner

Node Configuration

Node Type: Sub-Agent

Settings:

  • Name: Code Scanner
  • Prompt:
Identify files changed in the recent commit:

1. Get changed files using git diff
2. Read the changes in each file
3. Classify change types:
   - New files
   - Modified existing files
   - Deleted files
4. Count changed lines of code

Clarify the scope of what to review.
  • Tools: Bash, Read, Grep
  • Model: Haiku (fast scanning)

Design Points

  • Use Haiku: Simple file retrieval works fast with Haiku
  • Bash Permission: Required to execute git commands
  • Scope Clarification: Specify what to review in next step

Step 2: Choose Review Level

Node Configuration

Node Type: AskUserQuestion

Settings:

  • Question: What review level?
  • Header: Review Level
  • Options:
    • Option 1:
      • Label: Quick
      • Description: Check only obvious issues (within 5 minutes)
    • Option 2:
      • Label: Standard
      • Description: Comprehensive security, performance, quality (about 15 minutes)
    • Option 3:
      • Label: Thorough
      • Description: Detailed analysis and improvement suggestions (30+ minutes)

Design Points

  • Time Estimates: Specify duration for each level
  • Use Cases: Quick for emergencies, Thorough before production release
  • Gradual: Adjust prompts according to level

Step 3: Security Reviewer

Node Configuration

Node Type: Sub-Agent

Settings:

  • Name: Security Reviewer
  • Prompt:
Review code from a security perspective:

## Check Items

### 1. Injection Attacks
- SQL Injection
- Command Injection
- XSS (Cross-Site Scripting)

### 2. Authentication & Authorization
- Proper authentication checks
- Permission validation implementation
- Session management security

### 3. Sensitive Information
- Hardcoded API keys or passwords
- Proper handling of personal information
- Sensitive information in logs

### 4. Input Validation
- User input validation
- File upload restrictions
- Data type checking

## Output Format
For each issue:
- Severity (Critical/High/Medium/Low)
- Location (filename and line number)
- Issue description
- Suggested fix
  • Tools: Read, Grep
  • Model: Opus (advanced security analysis)

Design Points

  • Use Opus: Security is critical, analyze carefully with Opus
  • Structured Output: Specify severity and fix method
  • OWASP Compliant: Cover common vulnerability patterns

Step 4: Performance Reviewer

Node Configuration

Node Type: Sub-Agent

Settings:

  • Name: Performance Reviewer
  • Prompt:
Review code from a performance perspective:

## Check Items

### 1. Algorithm Efficiency
- Computational complexity (evaluate with Big-O notation)
- Unnecessary loops or recursion
- Appropriateness of data structures

### 2. Database Queries
- N+1 problem
- Index utilization
- Unnecessary data retrieval

### 3. Memory Usage
- Potential memory leaks
- Unnecessary object retention
- Bulk loading of large data

### 4. Asynchronous Processing
- Blocking operations
- Parallelization opportunities
- Cache utilization

## Output Format
- Issue location
- Current computational complexity
- Expected complexity after improvement
- Specific improvement code examples
  • Tools: Read, Grep
  • Model: Sonnet

Design Points

  • Quantitative Evaluation: Specify efficiency with Big-O notation
  • Improvement Code: Present specific Before/After
  • Practical: Suggest implementable improvements

Step 5: Quality Reviewer

Node Configuration

Node Type: Sub-Agent

Settings:

  • Name: Quality Reviewer
  • Prompt:
Review code from a quality perspective:

## Check Items

### 1. Readability
- Consistent naming conventions
- Appropriate comments
- Eliminate magic numbers

### 2. Maintainability
- Function length (recommended within 20 lines)
- Cyclomatic complexity
- Duplicate code detection

### 3. Testability
- Dependency injection usage
- Pure function usage
- Test-friendly structure

### 4. Best Practices
- Language/framework conventions
- Appropriate design pattern application
- Error handling

## Output Format
- Priority (Must/Should/Nice-to-have)
- Improvement suggestions
- References (if any)
  • Tools: Read, Grep
  • Model: Sonnet

Design Points

  • Prioritization: Classify as Must/Should/Nice-to-have
  • Educational: Explain why improvements are needed
  • Practical: Conform to team coding standards

Step 6: Summary Generator

Node Configuration

Node Type: Sub-Agent

Settings:

  • Name: Summary Generator
  • Prompt:
Integrate review results and create a summary:

## Review Summary

### 📊 Overall Rating
- Score: X/100
- Review Level: {{review_level}}
- Changed Files: X
- Changed Lines: X

### 🔴 Critical (Must Fix Immediately)
- Issue 1
- Issue 2

### 🟡 Important (Fix Recommended)
- Issue 1
- Issue 2

### 🟢 Minor (Address If Time Allows)
- Suggestion 1
- Suggestion 2

### ✅ Positives
- Good implementations
- Improved areas

### 📝 Next Actions
1. Fix critical issues
2. Consider addressing important issues
3. Check test coverage

Save review results in reviews/ directory.
  • Tools: Read, Write
  • Model: Sonnet

Design Points

  • Visual: Express priorities with emojis
  • Actionable: Specify what to do next
  • Positive: Point out good aspects to maintain motivation

Node Connections

All review nodes execute in sequence:

Code Scanner
   ↓
Choose Review Level
   ↓
Security Reviewer
   ↓
Performance Reviewer
   ↓
Quality Reviewer
   ↓
Summary Generator

To adjust prompt detail level by review level, you can add Branch nodes for branching.

Save and Export

Save

  1. Workflow name: ai-code-review
  2. Click Save button
  3. Saved to .vscode/workflows/ai-code-review.json

Export

Generated files:

  • .claude/agents/Code_Scanner.md
  • .claude/agents/Security_Reviewer.md
  • .claude/agents/Performance_Reviewer.md
  • .claude/agents/Quality_Reviewer.md
  • .claude/agents/Summary_Generator.md
  • .claude/commands/ai-code-review.md

How to Run

Normal Execution

/ai-code-review

Integration with Git Hook

Incorporate into pre-push hook for automatic review before push:

#!/bin/bash
# .git/hooks/pre-push

echo "Running AI code review..."
claude-code /ai-code-review

Best Practices

1. Using Review Levels

  • Quick: Hotfixes, urgent response
  • Standard: Regular feature additions/fixes
  • Thorough: Before release, important refactoring

2. Customizing Prompts

Add team-specific rules:

## Our Coding Standards
- 2-space indentation
- Comments must be in English
- Functions have single responsibility

3. Handling False Positives

When AI misdetects, specify exceptions in prompt:

These are not issues:
- Hardcoded values in test code
- Sample API keys in config files ("your-api-key-here")

4. Continuous Improvement

Save review results to files and review periodically:

# Analyze past review results
/analyze-review-trends

Team Adoption Tips

Step 1: Trial Run

First try individually to verify accuracy and usefulness

Step 2: Adjust Rules

Adjust prompts to match team coding standards

Step 3: Gradual Introduction

Initially use as “reference opinion”, make mandatory when reliability increases

Step 4: Combine with Human Review

AI review catches basic issues, humans focus on design and business logic

Measuring Effectiveness

Measure changes after workflow introduction:

  • Reduced review time
  • Improved bug detection rate
  • Post-release issue count
  • Team code quality score

Summary

AI review workflows designed in Claude Code Workflow Studio achieve both code quality improvement and efficiency.

Key Takeaways:

  • ✅ Covers security, performance, and quality
  • ✅ Flexibly respond with review levels
  • ✅ Can incorporate team-specific rules
  • ✅ Maximum effectiveness combined with human review

Next, resolve questions in Frequently Asked Questions.