SuperAPI AI Directory
Partner中文AI newsFavorites
Skillscode-review-team

code-review-team

Coding & development

View repository

Updated 10-15 20:06

About this skill

Code Review & Team Refactor

First understand the project's rules/skills/agent settings, and review from a corresponding expert perspective. If the user specifies the scope of improvements, spawn an Agent Team and make edits in parallel.

Overall Flow

Phase 1: 컨텍스트 탐색 → 전문가 패널 구성 → 리뷰 → 사용자 전달
  ── 사용자 검토 (개선 범위 지정) ──
Phase 2: 범위 확정 → TeamCreate → 병렬 개선 → 완료 보고

Phase 1: Code Review

Step 1. Explore Project Context

View changes before understand the project settings.

Exploration targets (in order):

Priority Path What to identify
1 CLAUDE.md, .claude/CLAUDE.md Project overview, core rules, conventions
2 .claude/settings.json Allowed tools, permission settings
3 .claude/skills/*/SKILL.md (frontmatter만) List of skills used by the project
4 .claude/agents/*/AGENT.md (frontmatter만) Registered agent roles
5 Project root configuration files package.json, pyproject.toml, go.mod etc. → determine language/framework
6 lint/format settings .eslintrc, ruff.toml, .prettierrc etc. → code style rules

Deliverable: Project profile summary

프로젝트 프로파일:
- 언어/프레임워크: {감지 결과}
- 핵심 규칙: {CLAUDE.md에서 추출한 규칙들}
- 코드 스타일: {lint/format 설정 요약}
- 등록된 스킬: {스킬 목록}
- 등록된 에이전트: {에이전트 목록}

Step 2. Configure Expert Panel

Based on the project profile, dynamically determine the expert perspectivesto participate in the review.

Expert panel configuration guide reference.

Default perspectives (always included):

  • Project rule compliance — whether project-specific rules such as CLAUDE.md, lint settings, etc. are violated
  • Correctness — logic errors, error handling, edge cases

Conditional perspectives (added depending on the profile):

  • Web framework detected → add security perspective
  • DB/ORM detected → add data layer perspective
  • Test framework detected → add test coverage perspective
  • etc. (see references for details)

Step 3. Analyze Changes

Determine the diff range according to the following priorities.

3-1. Determine Diff Range

Priority Condition diff range
1 User specifies commit/tag git diff <ref>..HEAD
2 Feature branch (non-default branch) git diff $(git merge-base <base> HEAD)..HEAD
3 Fallback Ask the user

Automatic base branch detection (used in Priority 2):

# 존재하는 첫 번째 브랜치를 베이스로 사용
for base in origin/dev origin/develop origin/main; do
  git rev-parse --verify "$base" &>/dev/null && break
done

Determine default branch: dev, develop, main, master If it is one of these, consider it the default branch.

3-2. Extract TICKET_ID

git branch --show-current

Extract TICKET_ID pattern from branch name:

  • feature/ABC-123-desc → ABC-123
  • fix/PROJ-456 → PROJ-456
  • ABC-123-some-desc → ABC-123
  • Pattern: [A-Z]+-[0-9]+ (first match)
  • On failure → ask the user

3-3. Collect Changes

git diff {범위} --name-only    # 변경 파일 목록
git diff {범위}                 # 전체 diff
git log {범위} --oneline        # 커밋 히스토리 (컨텍스트 파악용)

Step 4. Perform Review

Review the changed code from each expert perspective.

Review principles:

  • Project context first — Judge based on this project's rules/conventions rather than generalities
  • Specify the issue location precisely (file:line)
  • Provide the rationale for why it is a problem
  • Suggest an improvement direction, but specific fix code is handled in Phase 2

Issue severity:

Severity Criteria
CRITICAL Causes security vulnerabilities, data loss, or system failures
MAJOR Functional errors, violations of core project rules, or severe performance degradation
MEDIUM Code quality, convention inconsistencies, or maintenance difficulties
LOW Minor improvements, style

Step 5. Save and deliver the review document

Path: .ai/tasks/{TICKET_ID}/review.md Format: Review document template

If an existing file exists, review-01.md, review-02.mdincrement the version.

Deliver to the user:

📋 코드 리뷰 완료
📄 문서: .ai/tasks/{TICKET_ID}/review.md

프로젝트 컨텍스트: {언어}, {프레임워크}, {핵심 규칙 요약}
리뷰 관점: {적용된 전문가 목록}

| 등급 | 건수 |
|------|------|
| 🔴 CRITICAL | N |
| 🟠 MAJOR | N |
| 🟡 MEDIUM | N |
| 🟢 LOW | N |

주요 이슈:
1. src/auth.ts:45 — [보안] 사용자 입력 검증 누락
2. src/service.ts:120 — [프로젝트 규칙] CLAUDE.md의 에러 처리 규칙 미준수
...

개선을 진행하려면 범위를 지정해주세요.
예: "CRITICAL 전부", "1, 3번", "전부"

Phase 2: Team improvement

When the user specifies a scope, spawn an Agent Team.

Step 1. Confirm scope and group tasks

Group the target issues by file:

  • Same file / dependent files → one Task
  • Independent files → separate Tasks
  • Maximum 5 Workers

Step 2. Team SPAWN

Team spawn detailed guide refer to.

Perform the following procedure in order .

2-1. Create team

TeamCreate Call the tool.

Parameters 값
team_name "review-{TICKET_ID}"
description "{TICKET_ID} 코드 리뷰 이슈 개선"

2-2. Create tasks

For each issue group, TaskCreate Call the tool. Create all tasks simultaneously .

Parameters 값
subject "개선: {파일명} ({이슈 수}건)"
description Project context + assigned files + list of issues to fix
activeForm "{파일명} 개선 중"

2-3. Spawn Workers

As many times as the number of issue groups, Task Call the tool to spawn Workers. Spawn all Workers simultaneously .

Parameters 값
subagent_type "general-purpose"
team_name "review-{TICKET_ID}"
name "refactor-worker-{N}" (numbered starting from 1)
description "코드 개선 Worker {N}"
mode "bypassPermissions"
prompt Worker prompt (see below)

Content that must be included in the Worker prompt:

  • Project rules/conventions (core rules extracted from CLAUDE.md)
  • Assigned files and issue list (all issues in the relevant group from the review document)
  • Improvement direction for each issue
  • Work procedure instructions (TaskList → TaskUpdate in_progress → edit files → report to leader via SendMessage → TaskUpdate completed)
  • Rule to modify only assigned files

For the detailed Worker prompt template, Team spawn detailed guiderefer to.

2-4. Assign tasks

When a Worker is spawned, TaskUpdate assign a Task to each Worker with the tool.

Parameters 값
taskId The corresponding task ID
owner "refactor-worker-{N}" (same as Worker name)

2-5. Waiting for Completion

When Workers complete their tasks, SendMessagereport the results to. They are received automatically. Wait until reports from all Workers have arrived.

Step 3. Cleanup and Reporting

3-1. Terminating Workers

Request termination from all Workers using the SendMessage tool.

Parameters 값
type "shutdown_request"
recipient "refactor-worker-{N}"
content "작업 완료. 종료합니다."

3-2. Deleting the Team

After all Workers have terminated, TeamDelete call the tool.

3-3. Completion Report

✅ 개선 완료

수정된 이슈: N/M건
수정된 파일:
- src/auth.ts: 2건 (입력 검증 추가, 에러 처리 규칙 적용)
- ...

팀 정리 완료. 커밋하시려면 말씀해주세요.

Detailed Guide