Help us improve
Share bugs, ideas, or general feedback.
From deep-work
Explains the deep-work workflow phases (Brainstorm → Research → Plan → Implement → Test → Integrate) and their contracts. Use for overview, phase selection guidance, or understanding phase-to-phase gates.
npx claudepluginhub sungmin-cho/claude-deep-suite --plugin deep-workHow this skill is triggered — by the user, by Claude, or both
Slash command
/deep-work:deep-work-workflowThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
**v5.6.0 신규 기능:**
Orchestrates evidence-driven workflow across Brainstorm, Research, Plan, Implement, and Test phases. Handles session init, resume, migration, and model routing.
Auto-loop execution workflow with quality gates. Use when starting any non-trivial implementation task. Provides automatic task decomposition, code implementation, testing (L1-L4), and iterative quality gates until completion. Invoke with /autoworker.
Orchestrates all code-modifying development tasks like bug fixes, enhancements, and new features using adaptive phases for analysis, specs, TDD, implementation, and verification.
Share bugs, ideas, or general feedback.
v5.6.0 신규 기능:
/deep-fork [session-id] [--from-phase=PHASE]현재 세션을 fork하여 다른 접근법을 탐색합니다. 원래 세션은 보존됩니다.
FORK_PROJECT_ROOT)fork_info: fork된 세션의 상태 파일에 부모 관계 기록fork_children: 부모 세션의 상태 파일에 자식 목록 기록fork-snapshot.yaml: fork 시점 상태 스냅샷 (비교 기준점)/deep-status --tree: fork 관계 트리 시각화/deep-status --compare: fork 관계 자동 감지 비교v5.5.2 신규 기능:
.claude/deep-work-guard-errors.log 파일 기록v6.0.2 신규 기능:
deep-work/ → .deep-work/ (숨김 폴더). 마이그레이션 자동 처리phase_review 필드 추가 (기존 review_results 하위 호환)v5.5 신규 기능:
/deep-work "task" 하나로 전체 워크플로우가 자동 진행됩니다.
Plan 승인이 유일한 필수 인터랙션입니다.
v5.3 신규 기능:
/deep-status --history에서 세션 간 품질 추이 시각화/deep-status --badge로 shields.io 뱃지 생성Primary workflow (7): /deep-work, /deep-research, /deep-plan, /deep-implement, /deep-test, /deep-status, /deep-debug
Special utility (4): /deep-fork, /deep-mutation-test, /deep-phase-review, /deep-sensor-scan
Quality Gate (3): /drift-check, /solid-review, /deep-insight — /deep-test가 자동 실행; standalone 호출 가능.
Internal (6): /deep-brainstorm, /deep-finish, /deep-report, /deep-receipt, /deep-history, /deep-assumptions — orchestrator 또는 /deep-status가 내부 참조. 수동 호출도 공식 경로.
Escape hatch (1): /deep-slice — phase-guard가 TDD 블록 시 안내 (spike, reset).
Utility (2): /deep-cleanup, /deep-resume — standalone 기능. 향후 이관 후 삭제 예정.
Core mechanisms:
When AI coding tools work on complex tasks without structure, common failure modes emerge:
The Deep Work workflow prevents these by strictly separating brainstorming, analysis, planning, coding, testing, and integration into six distinct phases — the first five with enforced gates, plus Phase 5 Integrate as an optional post-test recommendation loop.
/deep-brainstorm) — OptionalGoal: Explore "why before how" — define the problem, compare approaches, establish success criteria.
What happens:
$WORK_DIR/brainstorm.mdWhat's blocked: All code file modifications (enforced by hook)
Skip: Use --skip-brainstorm to start directly at Research.
/deep-research)Goal: Build a complete mental model of the relevant codebase before making any decisions.
What happens:
$WORK_DIR/research.mdWhat's blocked: All code file modifications (enforced by hook)
Key principle: "You cannot plan what you don't understand, and you cannot understand what you haven't read."
Features:
/deep-research --scope=api,data re-analyzes specific areas onlyFor detailed guidance, see Research Guide or Zero-Base Guide.
/deep-plan)Goal: Create a detailed, reviewable, approvable implementation plan.
What happens:
$WORK_DIR/plan.mdWhat's blocked: All code file modifications (enforced by hook)
Key principle: "The plan is the contract between human and AI. No implementation without approval."
Features:
plan.v1.md, plan.v2.md with change logs/deep-implement 불필요).Note: Plan phase does not use Team mode — planning requires a single coherent document produced by one agent.
For detailed guidance, see Planning Guide.
/deep-implement)Goal: Mechanically execute the approved plan, task by task.
What happens:
implement_completed_at만 기록하고 current_phase 전환은 Orchestrator가 담당.sensor_results fields.slice_confidence (done/done_with_concerns) and specific concerns in the receipt.What's allowed: All tools — code modification is now permitted
Key principle: "The best implementation is a boring implementation. No creativity, no surprises, just faithful execution."
Features:
/deep-test 불필요).slice_review.mode: "self"For detailed guidance, see Implementation Guide.
/deep-test)Goal: Verify the implementation through comprehensive automated testing.
What happens:
sensor_results from receipts (no re-execution) to verify all slices passed computational sensors/deep-mutation-test handles this transition internallyWhat's blocked: All code file modifications (enforced by hook)
Key principle: "Trust but verify. The test phase catches what implementation missed."
Features:
$WORK_DIR/test-results.mdFor detailed guidance, see Testing Guide.
Phase 4 Test 완료 후 옵션으로 호출되는 "다음 단계 추천 루프". 설치된 deep-review/deep-docs/deep-wiki/deep-dashboard/deep-evolve 플러그인의 아티팩트를 읽어 AI가 최대 3개의 다음 단계를 추천하면, 사용자가 선택·실행하거나 skip·finish한다. --skip-integrate로 건너뛸 수 있고, /deep-integrate로 명시적 재진입도 가능하다. 자세한 UX/데이터 계약은 docs/superpowers/specs/2026-04-18-phase5-integrate-design.md 참조.
Compares plan.md items with actual git diff. Reports implemented, missing, out-of-scope, and design drift.
Standalone mode available: /drift-check [plan-file].
Evaluates code against the 5 SOLID design principles with a per-principle scorecard.
Standalone mode available: /solid-review [target]. See SOLID Guide.
Measures file metrics, complexity indicators, and dependency graphs. Never blocks workflow.
Standalone mode available: /deep-insight [target]. See Insight Guide.
Generates a comprehensive session report (research, plan, implementation, test outcomes, phase durations).
Auto-generated after all tests pass. Manual: /deep-report or /deep-status --report.
Hooks enforce phase boundaries and track activity:
phase-guard.sh): During Research, Plan, and Test phases — Write/Edit tools are blocked for all files except $WORK_DIR/ documents and the state file. During Implement — all tools available. No session — no restrictions.file-tracker.sh): During Implement phase — automatically logs modified file paths to $WORK_DIR/file-changes.log with timestamps. Used by /deep-report and /deep-insight.session-end.sh): On CLI session end — if a deep-work session is active, outputs a reminder message.This is not a suggestion — it's a hard gate. The AI literally cannot modify code files until the plan is approved, and cannot modify code during testing.
/deep-work "Add user authentication with JWT tokens"
# v6.3.1 Phase Exit Gates — 각 phase 완료 시 사용자 확인 (진행/재실행/일시정지)
# → Brainstorm → [Exit Gate] → Research → [review+approval + Exit Gate]
# → Plan → [review+approval + Exit Gate] → Implement → [Exit Gate]
# → Test → [Exit Gate: Integrate 또는 Finish] → Finish
# 수동 오버라이드가 필요할 때:
/deep-research # 리서치 다시 실행
/deep-plan # 플랜 수정
/deep-implement # 구현 재실행
/deep-test # 테스트 재실행
/deep-status # 상태 확인 (--receipts, --history, --report, --assumptions)
/deep-debug # 디버깅 모드
During /deep-work initialization:
Each session creates a unique task folder under .deep-work/:
.deep-work/
├── 20260307-143022-jwt-기반-인증/
│ ├── research.md
│ ├── plan.md
│ ├── test-results.md
│ └── report.md
├── 20260306-091500-api-리팩토링/
│ ├── research.md
│ ├── plan.md
│ ├── plan.v1.md ← plan version history
│ ├── test-results.md
│ └── report.md
Previous sessions are preserved when starting new ones. Use /deep-status to view history or /deep-status --compare to compare sessions.
First run saves setup answers to .claude/deep-work-profile.yaml as default preset. Subsequent runs skip all questions. Multi-preset support: dev, quick, review etc.
Flags: --profile=quick, --team, --zero-base, --skip-research, --no-branch, --setup
/deep-work 진입 시 stale 세션은 자동 감지되지만, active 세션 선택·worktree 컨텍스트 복원·phase별 resume dispatch는 /deep-resume을 통해서만 가능합니다.
Session state is stored in .claude/deep-work.{SESSION_ID}.md (e.g., .claude/deep-work.s-a3f7b2c1.md) with YAML frontmatter tracking. Legacy single-session path .claude/deep-work.local.md is auto-migrated on first use.
Use /deep-status at any time to see the current state, progress, phase durations, and next recommended action.
Use it when:
Consider Team mode when:
Skip it when:
Lightweight mode (skip to /deep-plan directly):
/deep-work then select "Plan부터" to skip researchUse built-in plan mode for quick task decomposition, Deep Work for complex subtasks needing thorough research and strict phase gates. They combine well: plan mode for initial design, Deep Work for implementation.
All commands auto-detect the user's language and output in that language. Korean is the reference format; Claude translates naturally while preserving structure.