claude-deep-work

This skill should be used at Phase 3 of deep-work to execute slice-based TDD (RED → GREEN → REFACTOR) for each approved plan slice. Dispatches in solo (inline) or team (implement-slice-worker subagent delegation by cluster_id) mode, runs computational sensors (lint/typecheck/coverage), performs 2-stage slice review (delegate reviewer + evaluator_model), and emits M3-envelope-wrapped receipts/SLICE-*.json via hooks/scripts/wrap-receipt-envelope.js. Triggered by 'implement phase', '구현 시작', '/deep-work continue', /deep-implement slash, cross-platform Skill({ skill: "deep-work:deep-implement", args: "..." }), or orchestrator dispatch after plan approval.

This skill should be used at Phase 2 of deep-work, after research.md approval, to design a detailed implementation plan with TDD slices. Decomposes work into SLICE-NNN units with `depends_on` DAG, acceptance contracts (`failing_test`, `acceptance_threshold`), and optional `cluster_id` hint for parallel-safe grouping. Emits plan.md with the inline slice DAG (Implement Phase's deep-implement skill parses it for worker fan-out — no external slices.md/slice-graph.json is emitted). Triggered by 'create implementation plan', '계획 수립', /deep-plan slash, cross-platform Skill({ skill: "deep-work:deep-plan", args: "..." }), or orchestrator dispatch after research approval.

Author the executable spec contract for deep-work research sessions. Invoked through /deep-spec, $deep-work:deep-spec, or orchestrator dispatch.

This skill should be used when the user invokes /deep-work "task", uses cross-platform Skill({ skill: "deep-work:deep-work-orchestrator", args: "task" }), asks to start a new deep-work session, or requests evidence-driven workflow auto-flow across Brainstorm → Research → Plan → Implement → Test phases. Handles session initialization (profile v3 load, capability detection, session-recommender sub-agent, AskUserQuestion 4-key ask, flag parsing) and dispatches the 5-phase pipeline with Exit Gates between each phase.

Use when the user finishes a deep-work session — merging, opening a PR, keeping the branch as-is, or discarding. Triggers on `/deep-finish`, "finish session", "wrap up session", "세션 마무리", "세션 종료", "PR 만들어줘", or orchestrator auto-call after Phase 5 (Integrate) completes. Emits an M3-envelope-wrapped `session-receipt.json` (and optionally a cross-plugin `handoff.json`). Supports `--skip-integrate` (Phase 5 bypass), `--handoff-to=<plugin>` / `--no-handoff` (post-finish handoff control).

Use when the user manually requests a deep-work phase document review. Parses --phase, --structural, and --adversarial compatibility arguments, then enters the canonical adaptive review protocol.

Use when the user wants to generate or view the comprehensive deep-work session report — session lifecycle, slice summary, evidence trail, TDD compliance, sensor pass-rate, model usage, evaluation outcomes. Triggers on `/deep-report`, `/deep-status --report`, "session report", "deep-work report", "세션 리포트", "세션 보고서", "리포트 생성". Writes to `$WORK_DIR/report.md` and prints inline summary. Sub-page of the deep-status hub.

This skill should be used during deep-work Phase 1 to research an existing codebase (architecture/patterns/risks) via research-codebase-worker, or to investigate a zero-base project's tech stack/conventions/data-model via research-zerobase-worker (using WebSearch + Context7 for up-to-date library docs). Consumes evolve-insights and harnessability M3 envelopes as context. Triggered by 'start research phase', 'analyze codebase', /deep-research slash, cross-platform Skill({ skill: "deep-work:deep-research", args: "..." }), or orchestrator dispatch after brainstorm approval. Solo/Team mode automatic based on project size.

Use when the user wants to resume an active deep-work session — restoring context from previous artifacts and continuing from the current phase. Triggers on `/deep-resume`, "resume session", "deep-work 이어서", "세션 재개", "이전 작업 계속", or when interrupted mid-phase. Detects active session via env var → pointer file → legacy fallback, restores worktree context, migrates state if version-bumped, clears phase cache, and dispatches to the per-phase resume skill. Supports `--session=<id>`, `--resume-from=<phase>`, `--worktree=<path>`.

Use when the user wants to see the current deep-work session status and progress — dashboard, badge, tree, or routing to sub-pages (receipts / history / report / assumptions). Triggers on `/deep-status`, "session status", "deep-work status", "세션 상태", "세션 현황", "상태 확인". Hub command that dispatches to deep-receipt (§6), deep-history (§7), deep-report (§8), deep-assumptions (§9) sub-skills via inline body Read. Supports `--compare` (fork comparison), `--receipts`, `--history`, `--report`, `--assumptions`, `--all`, `--tree`, `--badge`, `--risk`.

This skill should be used at Phase 4 of deep-work, after all slices in plan.md are marked [x], to run 9 verification gates: receipt completeness, plan-alignment drift, cross-slice spec/quality review, sensor clean, mutation, fitness delta, and health required/advisory. On failure, triggers the implement-test retry loop (cleanup implement_completed_at + receipt invalidation). Sets test_passed: true marker that gates deep-finish's session-receipt M3 envelope emit (with parent_run_id chain to consumed evolve-insights run). Triggered by /deep-test slash, cross-platform Skill({ skill: "deep-work:deep-test", args: "..." }), or orchestrator dispatch after implement approval.

Phase 5 — Integrate: 설치된 deep-suite 플러그인 상태를 수집하여 AI가 다음 단계를 추천하고 사용자가 선택·실행하는 대화형 루프. Phase 4(Test) 완료 후 호출되거나 `/deep-integrate` slash 또는 cross-platform Skill({ skill: "deep-work:deep-integrate", args: "..." })로 명시적 재진입 가능. `--skip-integrate`로 건너뛸 수 있다.

Use when the user invokes /deep-work "task", uses cross-platform Skill({ skill: "deep-work:deep-work", args: "task" }) or $deep-work:deep-work "task", asks to start a new deep-work session, or requests the primary Evidence-Driven Development auto-flow entry point. This is a compatibility alias for deep-work-orchestrator and preserves the historical /deep-work entry name.

Use when the user wants to inspect deep-work's assumption health report — Wilson Score-based per-assumption confidence, verdict (justified / loosen / drop), model-aware split, and decay history. Triggers on `/deep-assumptions`, `/deep-status --assumptions`, "assumption health", "rule justification", "loosen enforcement", "어썸션 헬스", "규칙 정당화", "가정 검증". Reads session evidence to decide whether deep-work's hook denylist + receipt validation rules are still justified. Sub-page of the deep-status hub.

This skill should be used at Phase 0 of a deep-work session to explore the problem space before implementation — defining goals, success criteria, boundaries, and comparing 2-3 candidate approaches. Triggered when the /deep-work orchestrator enters brainstorm phase, when /deep-brainstorm is invoked directly, or via cross-platform Skill({ skill: "deep-work:deep-brainstorm", args: "..." }) to ideate on a task without writing code yet. Skippable via --skip-brainstorm.

Use when the user wants to scan and clean up stale deep-work git worktrees (`dw/*` branches) plus fork worktrees. Triggers on `/deep-cleanup`, "clean worktrees", "stale worktree", "worktree 정리", "fork 정리", "deep-work 정리". Scans `git worktree list`, classifies stale/active, offers per-worktree or batch deletion via AskUserQuestion, handles dirty trees with extra confirmation, and prunes `.claude/deep-work-sessions.json` registry for idle fork sessions.

Use when the user enters debug sub-mode during deep-work Phase 3 (Implement) to investigate root causes before applying fixes. Triggers on `/deep-debug`, "systematic debug", "root cause investigation", "디버그 모드", "근본 원인 분석", `verification_cmd` 실패 시 자동 진입, or "fix this" once iron-rule blocks guessing. Four phases: Investigate → Analyze → Hypothesize → Implement. Iron rule: NO fixes without root cause investigation first. Escalates to user after 3 invalidated hypotheses.

Use when the user wants to fork an existing deep-work session to explore an alternative approach without losing the original. Triggers on `/deep-fork`, "fork session", "alternative approach", "세션 fork", "다른 접근", "분기 탐색". Resolves target session-id (explicit arg or auto-detect active), parses `--from-phase=<phase>` (brainstorm|research|plan|implement) to choose restart point, creates a new worktree+session pair, copies brainstorm.md/research.md/plan.md up to the from-phase boundary, and registers `fork_parent` link in `.claude/deep-work-sessions.json`.

Use when the user wants to view cross-session deep-work history — completed session list, TDD compliance rates, model usage (haiku/sonnet/opus/main), evaluator model usage, and aggregate trends. Triggers on `/deep-history`, `/deep-status --history`, "session history", "deep-work history", "세션 이력", "이전 세션", "TDD 준수율 트렌드". Scans `.deep-work/*/session-receipt.json` envelopes recursively, aggregates per-session stats, and displays trend indicators (recent-3 vs prior-3). Sub-page of the deep-status hub.

Use when the user wants a code-metrics + complexity + dependency analysis report on a target file/directory/glob. Triggers on `/deep-insight`, "code insight", "complexity report", "code metrics", "메트릭 분석", "복잡도 리포트", "의존성 분석", or auto-invocation by `/deep-test` as the Insight Tier of the 3-tier Quality Gate (never blocks). Analysis-only — does NOT modify code. Saves results to `$WORK_DIR/insight-report.md` when run inside a deep-work session, prints inline otherwise.

Use when the user wants mutation testing on changed files to verify AI-generated test quality, with optional auto-fix loop for surviving mutants. Triggers on `/deep-mutation-test`, "mutation test", "test quality", "mutation score", "뮤테이션 테스트", "테스트 품질 검증". Primary callsite is Phase 4 (Test) but supports standalone manual use. Computes `killed / (killed + survived)` score (excludes NoCoverage), auto-fix loop (max 3 rounds) transitions back to Implement phase to add tests for survived mutants then re-runs.

Use when the user wants to view, dashboard, or export deep-work slice receipts (`receipts/SLICE-*.json`). Triggers on `/deep-receipt`, `/deep-status --receipts`, "receipt dashboard", "slice receipt", "리시트 보기", "리시트 대시보드", "에비던스 리시트". Default subcommand is `dashboard` (ASCII visual of all slice receipts). Also: `view SLICE-NNN`, `export`, `validate`. Sub-page of the deep-status hub.

Use when the user wants to manually run computational sensors (linter, type checker, coverage) — works inside or outside a deep-work session. Triggers on `/deep-sensor-scan`, "run sensors", "lint check", "type check", "coverage", "센서 실행", "린트 검사", "타입 검사". Detects ecosystems via `sensors/detect.js`, runs per-tool with timeout, returns errors/warnings + FIX suggestions per file:line. Supports `--detect` (detection only), `--lint`, `--typecheck`, `--coverage` flags for selective runs.

Use when the user wants slice-level management inside a deep-work Phase 3 (Implement) session — status dashboard, manual activation, spike mode entry, reset, or per-slice model override. Triggers on `/deep-slice`, "slice status", "spike mode", "slice reset", "슬라이스 상태", "스파이크 모드". Primary escape hatch for TDD-blocked scenarios (phase-guard surfaces `/deep-slice spike|reset` in block messages). Subcommands: `activate SLICE-NNN`, `spike SLICE-NNN`, `reset SLICE-NNN`, `model SLICE-NNN [model]`.

Use when the user wants plan-alignment verification — comparing the approved plan against actual implementation to detect unimplemented items, out-of-scope changes, and design drift. Triggers on `/drift-check`, "plan drift", "drift detection", "plan alignment", "플랜 드리프트", "계획 정합성", "드리프트 검증", or auto-invocation by `/deep-test` as the Required Gate (blocks on drift). Read-only — does NOT modify code. Saves results to `$WORK_DIR/drift-report.md` when in workflow mode.

Use when the user wants SOLID design-principles code review — evaluating SRP / OCP / LSP / ISP / DIP compliance on a target file/directory/glob. Triggers on `/solid-review`, "SOLID review", "design review", "design principles", "SOLID 검증", "디자인 리뷰", "원칙 검증", or auto-invocation by `/deep-test` as the Advisory Quality Gate (does not block). Review-only — does NOT modify code. Saves results to `$WORK_DIR/solid-review.md` when in workflow mode.

This skill provides a high-level overview of the deep-work workflow (Brainstorm → Research → Plan → Implement → Test → Integrate, with M3 envelope receipt emit and Exit Gates between phases). Use when the user asks how deep-work works, requests a workflow overview, asks 'which phase should I start from', or needs to understand the phase-to-phase contracts. For executing an actual phase, prefer the phase-specific skills (deep-brainstorm, deep-research, deep-plan, deep-implement, deep-test, deep-integrate). Triggers: 'deep-work overview', 'workflow 개요', 'how does deep-work work', 'phase 구조 설명'.