This skill should be used when executing tasks from an approved plan.md. Triggers — ONLY when an approved plan.md already exists (or the user is approving a plan just presented) — on phrases like "구현", "implement", "이대로 진행", "진행해", "계속", "T<N> 진행", "go" (meaning "proceed with the plan"). If NO plan exists yet and the user asks to design or implement a non-trivial change, that is plan-feature first — it writes the plan, this skill runs it. Runs a FULLY AUTONOMOUS loop — processes ALL tasks (T1...Tn) without asking between tasks, stopping only when every task completes or a Halt Condition fires. Resuming mid-plan ("T6부터 계속") means T6 through the LAST task plus Phase F/G, not just T6. For trivial single-line edits without a plan, do NOT use this skill — Claude applies the change directly and lets hooks validate.
Plan a non-trivial code change — decompose into tasks, pre-resolve decisions, define acceptance, output one plan.md. Triggers on Korean (계획/설계/기능 추가/리팩토링/구현/수정/변경/여러 곳/전체/만들어줘/앱·도구 만들어) and English (plan/design/implement/refactor/build an app/create a tool), a multi-task request, or a change spanning multiple files, altering logic/signatures, adding a function/class/method, or refactoring a resource layer (i18n/theming/DI). Routing — approved plan.md + execute-only (구현/이대로 진행/계속/T<N> 진행/go) → implement-task; bug/crash reports (버그/에러/크래시/안 됨/왜 안 돼) → pjc-systematic-debugging first; single-component work (Domain/Application service, ViewModel, wiki page) → add-domain-service/add-viewmodel/llm-wiki. Do NOT trigger for trivial edits (single-line text/label/color/size, typo, comment, config, ≤3-line code adding no new symbol/signature, or multi-element pure value substitution of any count — Claude edits directly). Ambiguous scope — ask "A) edit directly / B) make a plan".
LLM WIKI(Obsidian vault) 운영·관리 스킬. 프로젝트를 위키에 추가/등록·정보 갱신(ingest)·삭제·상태 변경, 특정 기능/가이드/레시피를 위키에 추가, 위키 점검(lint)·질문(query)·복구(잘못 갱신된 페이지를 백업으로 되돌리기) 시 사용한다. "위키에 추가/등록", "{프로젝트} 위키에 등록", "위키 업데이트", "{프로젝트} 변경분 반영", "위키 점검/lint", "{기능} 위키에 추가", "위키 되돌려줘/복구", "큐 정리/pending 정리"(경량 큐 소비) 등의 요청에 발동. 코드 작업 세션에서 기능 구현·버그 수정 전에 관련 feature/recipe를 read-only로 참조할 때(절차 K)도 사용한다. 어느 디렉터리에서든(코드 프로젝트 폴더 포함) 호출 가능 — 위키 vault 밖에서도 동작한다. 단순 코드 수정·위키와 무관한 일반 지식 질문에는 발동하지 않는다(그건 plan-feature/implement-task 영역).
This skill should be used when starting work on a project that has no AGENTS.md file, to generate one. Triggered automatically by plan-feature when AGENTS.md is missing, manually with "/pjc:bootstrap-agents-md", or when the user asks to create a project agent guide - e.g. "AGENTS.md 만들어줘", "프로젝트 가이드 문서 자동으로 만들어줘", "Claude가 이 프로젝트 컨벤션을 알게 해줘". Detects project stack from marker files (.csproj, package.json, pyproject.toml, go.mod, Cargo.toml, etc.) and generates a minimal AGENTS.md from one of 9 stack templates (plus generic/multi-stack). If stack is unknown, asks the user. Do NOT trigger when an AGENTS.md or CLAUDE.md already exists, when only editing/adding a line to an existing AGENTS.md, or for writing a README.
Use whenever the user reports a bug, test/build failure, runtime error, exception, crash, unexpected behavior, performance regression, memory leak, race condition, deadlock, or CI/CD issue. Triggers on Korean (버그/에러/오류/예외/크래시/안 됨/동작 안 함/이상해/왜 이래/왜 안 돼/이상한 현상/테스트 실패/빌드 실패/재현/디버깅) and English (bug/fix/debug/error/exception/crash/fails/broken/regression). Root cause investigation is mandatory before any patch. If the compiler/stack trace pinpoints the cause (file·line·reason) and the fix is a small single-file change, use the lightweight path, not a skip. Not for non-bug "fix" requests (reformatting/renaming — trivial edits, not debugging). Skip ONLY when the user explicitly asks to apply a fix they already diagnosed ("그냥 이 한 줄만 수정해줘, 원인 다 안다"). pjc/DDD-integrated variant (regression-test-first fix, spec-compliance review, cross-project llm-wiki lookup); prefer over generic systematic-debugging in pjc projects.
This skill should be used when the user requests adding business logic, a domain service, an application service, or any logic that should live in the Domain or Application layer of a DDD project. Triggers on phrases like "서비스 추가", "도메인 로직", "비즈니스 로직", "use case 추가", "add service", "add use case". Generates Domain interface + implementation + DI registration + unit test scaffold. Do NOT trigger for logic that belongs inside a single Aggregate (add a method to the Aggregate instead), for pure UI/ViewModel work (use add-viewmodel), for infrastructure/config-only changes, or for projects without a Domain/Application layer split (single-project apps, scripts, or utilities with no DDD layering — forcing a nonexistent layer is wrong; implement directly per the project's structure).
This skill should be used when the user requests adding a new screen, dialog, page, window, or any UI component that needs a ViewModel in a WinUI 3 / WPF / MAUI project using CommunityToolkit.Mvvm. Triggers on phrases like "ViewModel 추가", "새 화면", "다이얼로그 추가", "페이지 만들기", "add screen/page/dialog/window". Generates ViewModel + View skeleton with proper MVVM bindings and DI registration. Do NOT trigger for non-XAML stacks (React/web, ASP.NET WebAPI controllers), simple UI text/label/style tweaks on an existing view, or debugging an existing ViewModel (use pjc-systematic-debugging). Android Jetpack ViewModel is out of scope.
This skill should be used when recording a CONFIRMED project fact (build/run command, DB access method, file/artifact location, test/verify command) into an EXISTING AGENTS.md — either right after the suggest-agents-record hook emits "[AGENTS 기록 제안]" and the user accepts, or when the user explicitly asks to record such a fact. Triggers on "AGENTS.md에 기록", "프로젝트 사실 기록", "빌드 명령 기록해줘", "이 명령 AGENTS에 추가", "DB 접근법 적어둬", "테스트 명령 기록", "AGENTS.md에서 이 항목 빼줘", "더 이상 안 쓰는 명령 지워줘", or accepting the hook's suggestion. The fact may be added, updated, or removed. Do NOT trigger for — creating a new AGENTS.md from scratch (use bootstrap-agents-md), planning or writing code (plan-feature/implement-task). Records into AGENTS.md ONLY — never the global/project CLAUDE.md — and ONLY after showing the exact change and getting user approval (no silent writes). Real secrets/connection strings/credentials are forbidden — environment variable names only.