Agent skill
harness
하네스, 테스트 주도 팀, TDD 팀, 병렬 개발, 스펙 구현 - Spec/PRD-driven parallel development using Agent Teams. Generates tests first, then orchestrates teammates to make all tests pass. Use when implementing a feature or project from a spec with agent teams. Do NOT use for single-file changes or simple bug fixes.
Install this agent skill to your Project
npx add-skill https://github.com/aimskr/aims-claude-toolkit/tree/main/skills/harness
Metadata
Additional technical details for this skill
- tags
-
tdd agent-teams parallel-development spec-driven
- author
- jaehashin
- version
- 1.3.0
- category
- development
SKILL.md
Harness - Spec-Driven Parallel Development
You are a team lead running a test-driven harness. Your ONE goal: make all tests pass.
PROGRESS 파일:
.harness/PROGRESS-{feature}.md— 다중 하네스 동시 실행을 지원한다.
Phase Overview
Phase 0: Spec Intake → Analyze spec, extract feature name, create .harness/PROGRESS-{feature}.md
Phase 1: Test Suite → Lead writes ALL tests (only phase lead writes code)
Phase 2: Team Setup → Spawn agent team, assign tasks
Phase 3: Harness Loop → Test → Analyze → Decide → Track → Commit → Repeat
Phase 4: Convergence → 100% pass confirmed, lessons learned, cleanup
Phase 0: Spec Intake
- Ask user for spec location
- Check
~/.claude/skills/harness/LESSONS_LEARNED.mdfor prior lessons - Read spec, analyze project structure (stack, conventions, test runner)
- Extract feature name: spec/PRD 제목 → kebab-case 변환 → 사용자 확인
- 예: "사용자 인증 시스템 MVP" →
auth-system-mvp - "하네스 이름:
auth-system-mvp— 맞나요?" - 영문/숫자/하이픈만 허용
- 예: "사용자 인증 시스템 MVP" →
.harness/디렉토리 준비- 없으면
mkdir .harness .gitignore에.harness/엔트리 없으면 추가
- 없으면
- 충돌 체크:
.harness/PROGRESS-{feature}.md가 이미 존재하면 사용자에게 경고- 덮어쓸지 / 이름 변경할지 확인
- Define module boundaries and file ownership map
- Create
.harness/PROGRESS-{feature}.md(seereferences/progress-template.md) - Confirm module breakdown with user
Phase 1: Test Suite Generation
Lead writes tests directly. For each module:
- Happy path, edge cases, error cases, integration points
- Tests verify general correctness, NOT specific outputs
- Run all tests → confirm ALL FAIL
- Commit:
test: initial test suite (0/N passing)
For test reporting format and execution tiers, see references/phase-details.md.
Phase 2: Team Setup
Inform user: "Phase 2 시작: delegate mode (Shift+Tab)로 전환해주세요."
Spawn team: researcher, impl-1~N (max 4), tester, reviewer.
Each implementer gets file ownership, target tests, and the PROGRESS file path (.harness/PROGRESS-{feature}.md).
For team creation prompts and spawn context, see references/phase-details.md.
Phase 3: Harness Loop
Core loop — repeat until all tests pass or max 10 iterations.
Step 1: Test → Message tester (quick/smoke/full tier)
Step 2: Analyze → Compare with previous .harness/PROGRESS-{feature}.md
Step 3: Decide → Regression? Stalled? Complete? Act accordingly
Step 4: Track → Update .harness/PROGRESS-{feature}.md (EVERY iteration)
Step 5: Commit → Implementers + lead commit
Step 6: Guard → Pause at 10 iterations, ask user
Decision Quick Reference
| Condition | Action |
|---|---|
| Regression | STOP new work → fix immediately |
| Stalled 2+ rounds | Researcher investigates → new tasks |
| Wrong test | Confirm with user → adjust with approval |
| Module complete | Reviewer checks → reassign implementer |
Parallelism by Pass Rate
- 0-80%: Max parallelism by module
- 80-95%: Cluster by dependency graph
- 95-100%: Researcher first, then single implementer
For full loop details, mistake recording protocol, and sub-agent guidelines, see references/phase-details.md.
Phase 4: Convergence
- Full test suite → confirm 100%
- Update
.harness/PROGRESS-{feature}.mdstatus toCOMPLETE - Extract lessons → append to
LESSONS_LEARNED.md - Delete
.harness/PROGRESS-{feature}.md(lessons already extracted) - If
.harness/is empty, remove the directory - Shutdown team, final report
문서화 (Phase 4 완료 후 자동 실행)
작업 완료 시 auto-documenter를 호출하여 프로젝트 문서를 업데이트한다.
Lead Rules Summary
See references/lead-rules.md for complete 19 rules. Key rules:
- NEVER implement code yourself (except Phase 1)
- Regression = highest priority
- File ownership is sacred
- No hard-coding, no over-engineering
- Record every mistake with prevention rule
Completion
Phase 4 완료: 전체 테스트 100% 통과 + .harness/PROGRESS-{feature}.md 삭제 + LESSONS_LEARNED.md 업데이트.
Troubleshooting
Tests pass but implementation is wrong: Tests may be too loose. Tighten assertions to verify correctness, not just absence of errors.
Agent team member is stuck: Reassign the task or spawn a researcher to investigate. Never wait more than 2 iterations on a stalled task.
Regression after a fix: Immediately stop all other work. The implementer who caused the regression owns the fix. Update .harness/PROGRESS-{feature}.md with the regression event.
PROGRESS file conflict: If .harness/PROGRESS-{feature}.md already exists at Phase 0, another harness may be running with the same name. Ask user to rename or confirm overwrite.
Recommended Agent Skills
Expand your agent's capabilities with these related and highly-rated skills.
build-error-resolver
빌드 에러, 빌드 오류, 컴파일 에러, 빌드 실패 해결, 의존성 에러 - Specialized in resolving build errors, compilation failures, and dependency issues. Use when build fails, compilation errors occur, or dependency conflicts arise. Do NOT use for runtime bugs or logic errors (use debug-specialist instead).
brainstorming
브레인스토밍, 아이디어, 기획, 구상, 아이디어회의, 설계, 요구사항 분석, 접근법 탐색 - Use before creating new features or significant changes to explore user intent, requirements, and design options. Collaborative brainstorming through step-by-step questioning. Do NOT use for simple bug fixes, config changes, or tasks with clear requirements already defined.
doc-coauthoring
문서 작성, 문서화, 문서, 스펙 작성, 기술 문서, 제안서, RFC, 설계 문서, 의사결정 문서 - Collaborative document co-authoring through 3 stages: context gathering, iterative refinement, and reader testing. Use when writing docs, proposals, tech specs, decision docs, or RFCs. Do NOT use for PRD/product requirements (use prd-strategist) or implementation plans (use writing-plans).
testing-strategy
테스트 전략, 테스팅 계획, QA 전략, 품질 보증, 테스트 피라미드, 테스트 시나리오, 커버리지 목표 - Designs test strategies including test pyramid ratios, scenario categories, and coverage targets. Use when planning how to test a feature, designing QA approach, or creating test plans. Do NOT use for TDD implementation (use tdd-workflow) or E2E test execution (use e2e-runner).
learning-research
학습 리서치, 학습 자료, 공부 자료, 학습, 공부, 스터디, 개념 정리, 이해 - 특정 주제를 깊이 이해하기 위한 학습 자료를 수집·정리한다. 병렬 전문가 서브에이전트로 개념, 원리, 실습, 심화 자료를 수집하고 이해도별로 체계화된 학습 노트를 생성한다. Obsidian vault에 자동 저장. 의사결정용 자료 조사는 research 스킬을, 시장조사는 market-research 스킬을 사용할 것.
devil-advocate
Didn't find tool you were looking for?