AgentSkillsCN

Code Review Orchestrator

将专项评审委托给子代理并综合其结果。适用于审查代码时。

SKILL.md
--- frontmatter
name: Code Review Orchestrator
description: Delegate specialized reviews to subagents and synthesize their findings. Use when reviewing code.

Code Review Orchestrator

Overview

This skill orchestrates specialized review subagents to analyze code and documentation. Each subagent applies a specific review skill and returns focused findings. The orchestrator collects and presents results.

When to Use

  • •Reviewing skills in .claude/skills/
  • •Evaluating CLAUDE.md files
  • •Assessing slash command prompts in .claude/commands/
  • •Any AI prompt or LLM instruction content review

Process

1. Identify Files to Review

Determine scope:

  • •Single or multiple files
  • •File type (skill, CLAUDE.md, command)
  • •Relevant review dimensions

2. Select Review Skills

Map content type to review skills:

For AI Prompt Content (skills, CLAUDE.md, commands):

  • •prompt-brevity/SKILL.md - Token efficiency and clarity

For Python Files (skills, CLAUDE.md, commands):

  • •python-code-review-types/SKILL.md - Ensuring python types are used correctly.

For all Pull Request Files (changed files in PRs):

  • •pr-file-review/SKILL.md - Identify unnecessary test artifacts and temporary files
    • •Give this agent a list of files in the PR or a command to find the list of files.

IMPORTANT: If no specific review skill exists for the file type being reviewed, STOP and inform the user:

code
No specific review skill exists for [file type/category]. Available review skills:
- prompt-brevity (for AI prompts, skills, CLAUDE.md, slash commands)

To review [file type], create a new review skill at:
.claude/skills/collaboration/reviews/[skill-name].md

Use /create-skill to generate a new review skill template.

DO NOT perform any direct code review. Only delegate to existing review subskills.

3. Launch Subagents in Parallel

For each selected review skill, launch a general-purpose subagent:

code
You are a specialized code reviewer. Apply the following skill to review the specified files.

**Skill to Apply**: [Read and follow .claude/skills/collaboration/reviews/[skill-name].md]

**Files to Review**: [list of file paths]

**Output Format**:

For each issue found, output:

---
**File**: [file path]
**Lines**: [line range, e.g., "45-52" inclusive or "78"]
**Issue**: [brief description]
**Suggestion**: [exact multiline replacement; omit if no specific change]
**Reason**: [why this improves the code]
---

Return ONLY the issues found. No commentary or summary.

4. Collect Subagent Results

Wait for all subagents to complete.

5. Present Findings

Output subagent results directly:

markdown
## Review Results

### Prompt Brevity Review

[paste subagent output]

### [Future Review Type]

[paste subagent output]

Do not add: Commentary, summaries, severity rankings, or additional suggestions.

Example

Example 1: Single Skill File Review

Input: Review .claude/skills/testing/test-driven-development.md

Subagent Task:

code
You are a specialized code reviewer. Apply the following skill to review the specified files.

**Skill to Apply**: Read and follow .claude/skills/collaboration/reviews/prompt-brevity.md

**Files to Review**: .claude/skills/testing/test-driven-development.md

**Output Format**:

For each issue found, output:

---
**File**: [file path]
**Lines**: [line range inclusive]
**Issue**: [brief description]
**Suggestion**: [exact replacement; omit if no specific change]
**Reason**: [why this improves the code]
---

Return ONLY the issues found.

Orchestrator Output:

markdown
## Review Results

### Prompt Brevity Review

---
**File**: .claude/skills/testing/test-driven-development.md
**Lines**: 12-15
**Issue**: Verbose explanation with redundant phrasing
**Suggestion**: Test-driven development: write tests before implementation code
**Reason**: 68% token reduction (23 → 7 tokens) while preserving core message
---

---
**File**: .claude/skills/testing/test-driven-development.md
**Lines**: 45
**Issue**: Fluff phrase "you should make sure to"
**Suggestion**: run the tests
**Reason**: Direct imperative is clearer and saves 5 tokens
---

Example 2: Multiple File Review

Input: Review all skills in .claude/skills/collaboration/

Action: Launch one subagent with all file paths

Orchestrator Output:

markdown
## Review Results

### Prompt Brevity Review

---
**File**: .claude/skills/collaboration/writing-commit-messages.md
**Lines**: 8-10
**Issue**: Ceremonial phrase "in order to"
**Suggestion**: to write good commit messages
**Reason**: Standard brevity pattern saves 2 tokens
---

---
**File**: .claude/skills/collaboration/writing-commit-messages.md
**Lines**: 67-72
**Issue**: Redundant examples demonstrating same concept
**Reason**: Reduces redundancy by 40 tokens without information loss (no specific replacement provided - requires context-aware consolidation)
---

---
**File**: .claude/skills/collaboration/code-review.md
**Lines**: 34
**Issue**: Hedge words "you might want to consider"
**Suggestion**: consider
**Reason**: Removes unnecessary hedging, saves 4 tokens
---

Anti-patterns

  • •

    ❌ Don't: Perform direct code reviews when no subskill exists

    • •✅ Do: Inform user and suggest creating appropriate review skill
  • •

    ❌ Don't: Add your own commentary or summary

    • •✅ Do: Present subagent findings exactly as returned
  • •

    ❌ Don't: Filter or editorialize subagent results

    • •✅ Do: Trust subagent expertise and show all findings
  • •

    ❌ Don't: Combine findings from different review types

    • •✅ Do: Keep each review skill's findings separate
  • •

    ❌ Don't: Launch subagents sequentially

    • •✅ Do: Launch all subagents in parallel for efficiency
  • •

    ❌ Don't: Add severity ratings or prioritization

    • •✅ Do: Let findings speak for themselves

Testing This Skill

Test Scenario 1: Single File Review

  1. •Select one skill file to review
  2. •Launch prompt-brevity subagent
  3. •Present results
  4. •Success: Only subagent findings displayed

Test Scenario 2: Parallel Multi-File Review

  1. •Select 3+ files for review
  2. •Launch subagent with all files in single task
  3. •Verify parallel execution (not sequential)
  4. •Success: All findings collected and presented cleanly

Test Scenario 3: No Issues Found

  1. •Review a well-written, concise skill file
  2. •Launch subagent
  3. •Verify proper handling of "No issues found"
  4. •Success: Clean output without inventing issues

Extensibility

To add new review dimensions:

  1. •Create new review skill in .claude/skills/collaboration/reviews/[new-skill].md
  2. •Update "Select Review Skills" section to include new skill
  3. •Launch additional subagent with new skill
  4. •Present findings in separate section

Example future skills:

  • •reviews/security-practices - Check for security anti-patterns
  • •reviews/accessibility - Verify inclusive language and examples
  • •reviews/consistency - Ensure style consistency across skills

Related Skills

  • •collaboration/reviews/prompt-brevity - Token efficiency review skill
  • •collaboration/pr-file-review - Pull request file artifact review
  • •meta/creating-skills - Skill creation guidelines

Remember: Orchestrate, don't review. Delegate to experts and present findings.