| 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>] |
Skill Manager
When to use
Use this skill when the user asks to create, update, review, test, delete, or restructure a project agent skill.
Typical triggers:
- "스킬 만들어줘", "스킬 수정", "SKILL.md 작성", "skill review"
- creating a workflow skill such as
git-stash-pull
- checking whether a skill belongs under
.agents/skills/
- validating frontmatter, references, scripts, index entry, or
.claude symlink
Boundary
- project skill 원본은
.agents/skills/{skill-slug}/에 둔다.
.claude/skills/{skill-slug}는 ../../.agents/skills/{skill-slug} relative symlink only.- Do not write project skill originals under
.codex/skills, .claude/skills, or another harness/domain-specific path unless the user explicitly asks for that external domain.
- Follow root
AGENTS.md, .agents/skills/AGENTS.md, and target subtree rules before writing.
- Do not store secrets, browser cookies, tokens, or private API keys in a skill.
Workflow
- Route - Classify as
agent_skill.project_skill and confirm the target slug.
- Inspect - Read
.agents/skills/AGENTS.md, .agents/skills/index.md, existing target skill files, and only the needed references.
- Research - For provider or standard claims, use existing raw sources first. For current provider behavior, verify official docs before changing policy.
- Author - Keep
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.
- Align behavior - When changing executable behavior, update the CLI,
SKILL.md, relevant references, and the target skill's eval in the same work.
- Validate - Run
scripts/validate-skill on the target skill and run the target skill's scripts/eval when present.
- Register - Update
.agents/skills/index.md and create or refresh the .claude/skills/{skill-slug} symlink.
- Close out - Report changed files, validation result, and any follow-up proposal.
Review Gates
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.- Portability: the skill is self-contained. Required behavior must not depend on reading repository documents or repository paths outside the skill directory — internalize needed policy into
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.
- Relative links in
SKILL.md resolve from the skill directory.
- CLI behavior defaults to stdout; file writes require explicit options.
- Behavior-changing CLI edits update matching
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.
Commands
Run from this skill directory:
./scripts/validate-skill ../skill-manager
./scripts/eval
Validate another project skill:
./scripts/validate-skill ../threads-fetch
Optional Independent Checks
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.
