SKILLS / INVOCATION

사람이 부르는 skill,
모델이 이어 부르는 skill

모든 skill이 같은 방식으로 시작하지 않습니다. user-invocable은 운영자가 명시적으로 호출하고, model-invocable은 작업 중 필요한 검증·정리 단계를 모델이 이어서 호출할 수 있습니다.

USER INVOKE

37개 운영 진입점

Bootstrapstate

/bootstrap

Initializes a fresh repository with the minimal CLAUDE.md, AGENTS.md, and active hook state.

새 저장소에 최소 CLAUDE.md, AGENTS.md, 활성 훅 상태를 만들고 조용하게 초기화합니다.

Use: Start a repository / 새 저장소 시작

Output: Project instruction files and .dev-kit/.active-hooks.json

Bootstrapenforcement

/ci-setup

Installs the reusable CI workflow templates idempotently and hands the repository to build.

재사용 가능한 CI 워크플로 템플릿을 멱등적으로 설치하고 build 단계로 넘깁니다.

Use: Add consumer-repository CI / 소비자 레포에 CI 추가

Output: .github workflows and .dev-kit/ci-config.json

Planstate

/plan

Turns a one-line idea into a PRD and resumable phase checklist through five gates, quantified value, and an ambiguity loop.

한 줄 아이디어를 가치 산정·모호성 루프를 거치는 5개 게이트로 PRD와 재개 가능한 phase 체크리스트로 바꿉니다.

Use: Define work before coding / 코딩 전 작업 정의

Output: PRD.md and phases/<name>/{index.json,step<N>.md}

Designstate

/evidence-plan

Escalates an idea through cited research and a human-reviewed HTML proposal before expensive PRD planning.

비용이 큰 PRD 계획 전에 인용 리서치와 사람이 확인하는 HTML 제안서로 아이디어를 검증합니다.

Use: Test an idea early / 아이디어 조기 검증

Output: Cited research, proposal HTML, plan hand-off

DesignRuns a safety-contract interview that narrows ambiguity, deduplicates repeated cycles, and gates plan emission.

/interview

안전 계약 인터뷰로 모호성을 좁히고 반복 질문을 제거한 뒤 plan 산출물을 허용합니다.

Use: Requirements are ambiguous / 요구사항이 모호할 때

Output: Safety contract and narrowed delta

DesignRenders structured YAML proposals as reviewable HTML with before/after, trade-offs, and limitations.

/proposal

구조화된 YAML 제안서를 전후 비교·트레이드오프·제약이 포함된 검토용 HTML로 렌더링합니다.

Use: Review a design before implementation / 구현 전 설계 검토

Output: docs/proposals/<main>/<sub>.html

DesignRuns cache, direct, multi-source, and human escalation phases, then verifies and enforces citations for factual claims.

/research

캐시·직접·다중 출처·사람 검토 단계로 리서치를 확장하고 사실 주장에 출처를 강제합니다.

Use: Make claims source-backed / 주장에 근거 붙이기

Output: Verified sources and citation-enforced report

DesignInterviews for five rounds of evidence-backed recommendations and hands a traceable single source of truth to plan.

/sot-harness-writer

5라운드 인터뷰에서 근거 기반 추천을 만들고 추적 가능한 Single Source of Truth를 plan으로 넘깁니다.

Use: Write a durable harness spec / 하네스 명세 작성

Output: Traceable SOT harness document

BuildDelegates phase steps to sub-agents, integrates TDD, verification, debugging, and self-fix loops, and resumes from state.

/build

phase 단계를 서브에이전트에 위임하고 TDD·검증·디버깅·자가 수정 루프를 통합해 상태에서 재개합니다.

Use: Implement a planned phase / 계획된 phase 구현

Output: Step outputs, tests, and completed index state

BuildUses four phases—reproduce, isolate, patch, verify—and refuses a fix proposal before reproduction.

/build-debug

재현·격리·패치·검증 4단계를 사용하며 재현 전에는 수정안을 제시하지 않습니다.

Use: Diagnose a failing behavior / 실패 원인 진단

Output: Root-cause hand-off to plan

BuildRuns inspect, a three-pass deletion sweep, and review to remove AI slop or dead features rather than refactor them.

/prune

리팩터링이 아닌 삭제가 필요할 때 inspect·3회 삭제 스윕·review를 실행해 AI 슬롭과 죽은 기능을 제거합니다.

Use: Delete low-value code / 저가치 코드 삭제

Output: Scoped deletion diff and review verdict

BuildChains inspect, build-refactor, and review with exit-code gates between cleanup phases.

/refactor

inspect·build-refactor·review를 연결하고 단계 사이에 종료 코드 게이트를 둡니다.

Use: Run a safe cleanup chain / 안전한 정리 체인 실행

Output: Audit, cleanup, and review evidence

ReviewFans out correctness, security, and architecture reviewers, filters false positives, and produces inline findings plus a verdict.

/review

정확성·보안·아키텍처 리뷰를 병렬 실행하고 오탐을 걸러 줄 단위 finding과 판정을 냅니다.

Use: Review a diff before merge / 머지 전 diff 검토

Output: Inline findings and Approve/Changes Requested/Blocked

SecurityRuns one verifier-backed sub-agent per OWASP Top 10 2025 category and returns a per-category breakdown.

/security

OWASP Top 10 2025 각 영역에 에이전트를 하나씩 배치하고 검증 패스 후 영역별 결과를 냅니다.

Use: Perform a deep security review / 심층 보안 리뷰

Output: OWASP A01–A10 findings and verdict

SecurityCalculates a deterministic 0–100 source-tree-only scorecard for all OWASP Top 10 categories.

/security-metrics

소스 트리만 읽어 OWASP Top 10 전 영역의 결정론적 0–100 점수를 계산합니다.

Use: Triage security hygiene quickly / 보안 위생 빠른 점검

Output: Markdown scorecard with PASS/REVIEW evidence

AuditRuns an eight-dimension read-only health audit covering dead code, duplication, smells, overengineering, architecture, cleanliness, token budget, and slop.

/inspect

죽은 코드·중복·스멜·과잉 설계·과잉 아키텍처·클린 코드·토큰 예산·슬롭 8축을 읽기 전용으로 점검합니다.

Use: Understand repository health / 레포 건강도 파악

Output: Dimensioned audit report

AuditChecks marker files, provider configuration, secrets, workflows, and GitHub authentication in one PASS/FAIL summary.

/ci-doctor

마커·프로바이더 설정·시크릿·워크플로·GitHub 인증을 한 번에 검사해 PASS/FAIL로 요약합니다.

Use: Check PR readiness / PR 전 준비도 확인

Output: Single CI readiness summary

AuditClassifies recent failed GitHub Actions runs, deduplicates cases, and requires a reproducible regression path.

/ci-triage

최근 실패한 GitHub Actions를 분류·중복 제거하고 재현 가능한 회귀 경로를 요구합니다.

Use: Investigate recurring CI failures / 반복 CI 실패 조사

Output: Persisted triage cases and repro evidence

AuditWalks the repository and emits a self-contained HTML architecture map across code, skills, hooks, tools, external systems, and workflows.

/code-viz

코드·스킬·훅·도구·외부 시스템·워크플로를 탐색해 자기 완결형 HTML 아키텍처 맵을 만듭니다.

Use: Explain a complex harness / 복잡한 하니스 설명

Output: Multi-level code visualization HTML

AuditReports live session spend, distance to threshold, and copyable commit trailers without blocking work.

/cost-gate

실시간 세션 지출·임계치 거리·복사 가능한 커밋 트레일러를 보고하며 작업을 차단하지 않습니다.

Use: Track engineering cost / 개발 비용 추적

Output: Cost status and Cost-gate trailers

AuditAudits documentation with README as the highest-priority document and updates time-sensitive facts carefully.

/docs-maintenance

README를 최우선으로 문서를 감사하고 시점에 따라 변하는 사실을 조심스럽게 갱신합니다.

Use: Refresh stale docs / 낡은 문서 갱신

Output: Verified documentation changes

AuditDistills source text, URLs, or transcripts into candidate SKILL.md files with deterministic checks and approval gates.

/learn

소스 텍스트·URL·트랜스크립트를 결정론적 검사와 승인 게이트를 거친 후보 SKILL.md로 증류합니다.

Use: Create a reusable skill / 재사용 스킬 만들기

Output: Approved candidate SKILL.md

AuditCombines skill usage telemetry with per-skill deletion proposals and requires explicit approval for each deletion.

/prune-propose

스킬 사용량 텔레메트리로 삭제 제안을 만들며 각 삭제마다 명시적 승인을 요구합니다.

Use: Reduce unused skills / 사용하지 않는 스킬 줄이기

Output: Per-skill deletion proposal

AuditTurns Claude Code and Codex logs into a four-dimension efficiency dashboard with anti-patterns and savings estimates.

/token-analyzer

Claude Code·Codex 로그를 4축 효율 대시보드로 바꾸고 안티패턴과 절감 추정치를 보여줍니다.

Use: Audit session efficiency / 세션 효율 감사

Output: HTML dashboard and transcript sidecars

EvaluationReplays transcripts against registered rubrics and reports legacy agent behavior plus five harness-effectiveness components.

/evaluate

트랜스크립트를 루브릭으로 재생해 에이전트 행동과 5개 하니스 효과성 요소를 보고합니다.

Use: Measure harness changes / 하니스 변경 측정

Output: Rubric scores and evidence report

EvaluationBuilds a compact scorecard for prevention, first-pass success, recovery, learning, and measurement integrity.

/harness-effectiveness

예방·첫 시도 성공·복구·학습·측정 무결성 5요소의 간결한 점수표를 만듭니다.

Use: Track guardrail effectiveness / 가드레일 효과 추적

Output: JSON scorecard and status verdict

ShipPolls GitHub checks, reads failures, applies minimal fixes, pushes, and loops until review and required checks pass.

/babysit-pr

GitHub 검사를 감시하고 실패 로그를 읽어 최소 수정·push를 반복해 리뷰와 필수 검사를 통과시킵니다.

Use: Finish a failing PR / 실패한 PR 마무리

Output: Fixed branch and passing check evidence

ShipRuns the local pytest gate and local LLM judge loop when GitHub checks are unavailable.

/babysit-pr-local

GitHub 검사를 사용할 수 없을 때 로컬 pytest 게이트와 로컬 LLM 판정 루프를 실행합니다.

Use: Validate locally before push / push 전 로컬 검증

Output: Local review verdict

ShipCreates an explicit plugin version bump branch and updates plugin.json for race recovery or pre-PR release control.

/bump

레이스 복구나 PR 전 릴리스 통제를 위해 plugin.json을 올리고 버전 bump 브랜치를 만듭니다.

Use: Release a plugin version / 플러그인 버전 릴리스

Output: chore/bump-vX.Y.Z branch

ShipFreshly fetches PR metadata, checks, and comments to avoid stale CI or judge results.

/pr-verify

PR 메타데이터·검사·댓글을 매번 새로 읽어 오래된 CI나 판정 결과를 방지합니다.

Use: Confirm PR state / PR 상태 확정

Output: Deterministic PR verification

ShipEmits a release tag only after review approval and the main-branch protection gate pass.

/ship

리뷰 승인과 main 보호 게이트 통과 후에만 릴리스 태그를 발행합니다.

Use: Cut a release / 릴리스 생성

Output: Release tag and changelog hand-off

ConfigProvides a multi-select picker for skills, hooks, and methodology settings.

/config

스킬·훅·방법론 설정을 다중 선택으로 고르는 구성 도구입니다.

Use: Tune harness behavior / 하니스 동작 조정

Output: Updated active configuration

ConfigOptionally reconciles repository work with a non-duplicate Linear issue and owner-gated lifecycle sync.

/linear

선택적으로 작업을 중복 없는 Linear 이슈와 연결하고 소유자 게이트가 있는 생명주기 동기화를 제공합니다.

Use: Track work in Linear / Linear로 작업 추적

Output: Canonical issue and synchronized hand-off

ShortcutsRefreshes the Codex marketplace checkout and versioned plugin cache when installed skills may be stale.

/codex-cache-update

설치된 스킬 캐시가 오래됐을 때 Codex 마켓플레이스 체크아웃과 버전 캐시를 갱신합니다.

Use: Repair stale Codex skills / 오래된 Codex 스킬 복구

Output: Synchronized plugin cache

ShortcutsExtracts vendor pricing from official pages, diffs provider JSON, and preserves pricing as a single source of truth.

/llm-refresh

공식 가격 페이지에서 정보를 추출하고 프로바이더 JSON을 diff해 가격 SSOT를 유지합니다.

Use: Refresh model pricing / 모델 가격 갱신

Output: Updated docs/llm-info provider data

ShortcutsSets up, enables, reports, or disables session logging used by token, skill-usage, and monitoring tools.

/log

토큰·스킬 사용량·모니터링 도구가 쓰는 세션 로그를 설정·활성화·조회·비활성화합니다.

Use: Manage session observability / 세션 관측 관리

Output: Managed log hooks and captured JSONL

StatusShows current HOTL loop progress, cumulative cycles, hand-off chain, and evaluation score in one view.

/status

현재 HOTL 루프 진행도·누적 사이클·핸드오프 체인·평가 점수를 한 화면에 보여줍니다.

Use: Orient in a long task / 장기 작업 방향 확인

Output: Status visualization

MODEL INVOKE

5개 내부 검증 단계

DesignScores a plan on six value axes and returns proceed, revise, hold, or kill with a persisted verdict envelope.

/valuate

계획을 6개 가치 축으로 평가하고 proceed·revise·hold·kill 판정과 영속화된 판정 봉투를 만듭니다.

Use: Decide whether work is worth building / 만들 가치 판단

Output: .dev-kit/valuations/<plan-id>.json

BuildPerforms gated cleanup passes for dead code, duplication, naming, and coverage, requiring regression evidence.

/build-refactor

죽은 코드·중복·이름·커버리지 순으로 정리하며 회귀 검증 근거를 요구합니다.

Use: Clean an existing implementation / 기존 구현 정리

Output: Four-pass cleanup evidence

BuildEnforces Red-Green-Refactor when TDD methodology is active; production code cannot precede a failing test.

/build-tdd

TDD 방법론에서 Red-Green-Refactor를 강제하며 실패 테스트보다 프로덕션 코드를 먼저 쓰지 못하게 합니다.

Use: Build with test-first discipline / 테스트 우선 구현

Output: Failing test, passing test, refactor evidence

BuildChecks completion claims against quoted exit codes, test counts, and build logs before accepting done.

/build-verify

완료 주장 전에 인용된 종료 코드·테스트 수·빌드 로그를 확인합니다.

Use: Close a build step honestly / 빌드 단계 정직하게 종료

Output: Verification bundle

AuditDiagnoses hook registration and cache drift, repairs safe issues, and reports the exact restart step.

/hook-doctor

훅 등록·캐시 드리프트를 진단하고 안전한 문제를 고친 뒤 정확한 재시작 단계를 안내합니다.

Use: Recover broken Claude/Codex hooks / 망가진 훅 복구

Output: Repair report and restart instruction