菜单
Creates, updates, reviews, and tests project agent skills under .agents/skills with .claude symlink compatibility. Use when the user asks to write or improve SKILL.md, create a reusable skill, validate skill structure, restructure skill-manager, or prepare workflow skills such as git-stash-pull. 스킬 생성, 수정, 리뷰, 검증, SKILL.md 작성 요청에 사용.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Safely delete a document from an agent-authored content collection (notes, inbox, sources, wiki) by first mapping every inbound reference to it across the whole repo, surfacing structural dependencies (hub↔child, category TOC), and only then removing it with git rm so the deletion is reviewable and reversible. Never auto-rewrites inbound links — it reports them and proposes edits. Handles hubs with children explicitly (delete subtree / reparent / abort), never orphaning children. Use when the user asks to delete/remove a note, wiki card, spec, or inbox capture, or to clean up an obsolete doc without leaving dangling links. 노트·위키· 스펙·inbox 문서 안전 삭제, 참조 파악 후 삭제, dangling 링크 없이 정리 요청에 사용.
Take a raw idea, research BOTH how to best develop it (idea-handling methodologies) AND its subject-domain evidence (official/primary sources first), apply the chosen methodology, and persist the result as an OKF note under src/content/idea, served on the unlisted /idea route. /idea behaves like the notes route (URL-only, excluded from search index, sitemap, and robots) but is additionally NOT linked in the footer — reachable only by a direct URL. Content is committed plaintext, so it is public-if-found; do not put secrets in it. Use when the user hands over an idea to develop, expand, pressure-test, or capture. 아이디어를 방법론+도메인 근거로 조사·발전시켜 unlisted /idea 노트로 남길 때 사용. 아이디어 정리·발전·검증 요청에 사용.
Lightweight idea capture. Given a raw idea (or a few), do a light cleanup, list a handful of questions worth exploring next, and run a quick scan of related references and existing services — then save it as a short note under src/content/idea/inbox, served at /idea/inbox/. This is the fast, low-ceremony sibling of the `idea` skill: no full methodology, no exhaustive primary-source verification, no business analysis — just enough structure to make the idea reviewable and promotable later. Use when the user runs /ideabox or wants to jot and lightly organize an idea to BUILD. NOT for research/analysis captures — those go to the separate top-level `src/content/inbox/` collection (`YYYY-MM-DD-{slug}.md`, no owning skill), so a bare "inbox에 추가해줘" / "조사해서 inbox에 넣어줘" is NOT this skill. 아이디어를 가볍게 정리·질문·레퍼런스 스캔해 idea 인박스에 남길 때 사용.
Batch-polish agent-authored content collections (notes, inbox, wiki, idea) — NOT user-authored posts/read-and-write. Two body edits: (1) highlight the load-bearing spans with `==마커==`, sparingly; (2) align structure to each collection's own charter (hub 통합 + 하위 목차 재구성), deferring to the owning skill's model. A deterministic checker stamps a body-only `polishHash` and flags files whose body drifted since their last polish. Warning-only in pre-push. Runs as an independent pass (manual batch, or scoped to named files) — not coupled to /lint. Use before commit/push, or when the polish check flags drift.
Research references for a topic — official/primary sources first — and persist them into the OKF-conformant wiki collection (src/content/wiki), categorized in a nested tree. Each doc gets a `type` frontmatter, a readable prose core, and primary/secondary reference links. Use when the user wants to gather the sources/evidence behind a topic, build up a reference category, or asks to "조사해서 위키에 정리" / "레퍼런스 카드 만들어줘". Distinct from the one-off deep-research report: /research produces a persistent, browsable, searchable library entry. 레퍼런스 조사·근거 자료 수집· 위키 카테고리 정리·공식 문서 우선 조사 요청에 사용.
Write and revise CV/résumé project entries in a consistent Korean résumé voice — flowing narrative + separated impact, with an optional leading Why. Use when adding a project or experience to the CV, rewriting an entry, tightening wording, removing redundant dates, or converting segmented What/How/Impact bullets into narrative flow. CV 항목 작성·수정, 이력서 문체(개조식) 정리, 프로젝트 추가 시 사용.
| name | skill-manager |
| description | Creates, updates, reviews, and tests project agent skills under .agents/skills with .claude symlink compatibility. Use when the user asks to write or improve SKILL.md, create a reusable skill, validate skill structure, restructure skill-manager, or prepare workflow skills such as git-stash-pull. 스킬 생성, 수정, 리뷰, 검증, SKILL.md 작성 요청에 사용. |
| compatibility | Project source .agents/skills; Claude Code compatibility through .claude/skills relative symlink. |
| repo-operating-targets | .agents/skills, .claude/skills, .codex/skills, AGENTS.md |
| argument-hint | [create | update | review | test <skill>] |
Use this skill when the user asks to create, update, review, test, delete, or restructure a project agent skill.
Typical triggers:
git-stash-pull.agents/skills/.claude symlink.agents/skills/{skill-slug}/에 둔다..claude/skills/{skill-slug}는 ../../.agents/skills/{skill-slug} relative symlink only..codex/skills, .claude/skills, or another harness/domain-specific path unless the user explicitly asks for that external domain.AGENTS.md, .agents/skills/AGENTS.md, and target subtree rules before writing.agent_skill.project_skill and confirm the target slug..agents/skills/AGENTS.md, .agents/skills/index.md, existing target skill files, and only the needed references.SKILL.md lean. Put local policy in local-policy.md, provider principles in provider-principles.md, and detailed lifecycle notes in lifecycle.md. Alongside the success path, write a failure spec (false-success / out-of-scope / insufficient-evidence signals plus the action for each) and observable termination conditions (success / abstain / escalate) — LLMs default to no abstention output, so the wrong-answer space must be written into the skill's context explicitly.SKILL.md, relevant references, and the target skill's eval in the same work.scripts/validate-skill on the target skill and run the target skill's scripts/eval when present..agents/skills/index.md and create or refresh the .claude/skills/{skill-slug} symlink.Treat these as hard blockers before declaring a skill complete:
SKILL.md has valid YAML frontmatter with name and description.name is lowercase kebab-case, at most 64 chars, and matches the directory basename.description says when to load the skill, not only what it is.SKILL.md stays under 500 lines and links to references instead of embedding long policy.SKILL.md or references/. Repo paths may appear only as (a) optional provenance the skill works without, or (b) the skill's explicit operating target. Generic conventions (e.g. reading a target directory's local rules before writing) stay conditional: "if present". scripts/validate-skill lints unconditional repo-path tokens in the SKILL.md body; declare intentional case (b) targets in frontmatter repo-operating-targets (comma-separated prefixes), and phrase case (a) lines as conditional/optional/boundary.SKILL.md resolve from the skill directory.SKILL.md, references, and eval coverage before closeout..claude/skills/{skill-slug} points to the .agents/skills/{skill-slug} source.SKILL.md declares a failure spec (including false-success signals) and observable termination conditions (success / abstain / escalate). A skill with only a success path is incomplete.Run from this skill directory:
./scripts/validate-skill ../skill-manager
./scripts/eval
Validate another project skill:
./scripts/validate-skill ../threads-fetch
When the environment supports subagents, use reviewer.md for an independent quality pass and smoke-tester.md for prompt-level smoke tests. These checks are advisory; the main agent keeps final write-safety and repository routing responsibility.