Files

JiWoong Sul 916a50992c docs: CLAUDE.md 작업 프로토콜 추가 및 수정 계획 작성

- CLAUDE.md: Claude-Gemini 교차 토론 프로토콜 추가
- CLAUDE.md: 존재하지 않는 디렉토리 3개 제거
- analysis/fix-plan: 4 Phase 수정 계획 (Claude-Gemini 합의)
- .claude/agents/: dev 리뷰 에이전트 9개 복사
- .claude/skills/: 프로젝트 스킬 4개 복사

2026-03-27 16:52:52 +09:00

2.4 KiB

Raw Blame History

Project Audit Skill

Trigger

/project-audit [absolute_path] or "프로젝트 감사", "코드 리뷰"

Description

Runs all 8 dev review agents on a given directory path. Produces a unified audit report.

Input

Absolute directory path (e.g., /Users/user/projects/my-app)
The path MUST exist and contain a development project

Execution Workflow

Step 1: Reconnaissance

Scan directory structure (Glob)
Identify project type (language, framework)
Find entry points, configs, package manifests

Step 2: Parallel Analysis (4 agents)

dev-architecture: Structure and design
dev-code-quality: Code smells and readability
dev-security: Vulnerabilities and secrets
dev-supply-chain: Dependencies and licenses

Step 3: Parallel Analysis (4 agents, may use Step 2 context)

dev-performance: Bottlenecks
dev-docs-sync: Documentation accuracy
dev-devops: CI/CD and deployment
dev-test-coverage: Test quality

Step 4: Unified Report

Merge all 8 agent results into a single audit document.

Context Window Management (Layered Analysis)

For large projects, each agent follows this scan strategy:

L1 (Always): Entry points, configs, package manifests, README
L2 (Core): Core business logic, domain layer, API routes
L3 (On demand): Utilities, helpers, generated code — only if L1/L2 findings indicate issues

Output Format

Final deliverable in Korean (한국어).

# [Project Name] 종합 감사 리포트

## 종합 건강 점수: [0-100]

## 요약 대시보드
| 영역 | 점수(/10) | 상태 | 핵심 이슈 |
|------|----------|------|----------|
| Architecture | | 🟢/🟡/🔴 | |
| Code Quality | | | |
| Security | | | |
| Supply Chain | | | |
| Performance | | | |
| Documentation | | | |
| DevOps | | | |
| Testing | | | |

## Critical Findings (즉시 조치)
1. ...

## 상세 보고서 링크
- [Architecture](./dev-architecture-report.md)
- [Code Quality](./dev-code-quality-report.md)
- ...

Brutal Analysis Principles

Scores must reflect reality. A project with no tests and hardcoded secrets cannot score above 30
Cross-reference findings between agents (e.g., security finding + missing test = compounded risk)

Claude-Gemini Cross-Debate Protocol

Each agent step includes Claude-Gemini debate. The unified report is also Gemini-reviewed.

Save Path

[project_path]/audit/project_audit_[date].md or user-specified location

2.4 KiB Raw Blame History