Repo Skill Creator

Repo Skill Creator

Create and maintain repository skills with intent-specific workflows. Validation is read-only with respect to Git state; creation, finalization, staging, and runtime installation are separate actions.

Trigger Rules

1. Classify the request as create, finalize-path, validate, or finalize-and-stage before running a command.
2. If the user provides a skills/* path and indicates the skill is already generated/copied, run path-finalize mode directly without additional confirmation.
3. A request to check, audit, or validate existing skills runs validation only. It must not create a skill or modify the Git index.
4. Run auto-discover finalize-and-stage only when the user explicitly asks to prepare newly added skills for commit or to stage them.
5. A bare invocation with no path or action defaults to validation. If exactly one new skills/<slug>/SKILL.md is discoverable, report it as the likely finalize target, but do not stage or create anything implicitly.

Mode A: Create New Skill

Primary command (recommended):

pnpm skills:new

Manual creation commands:

pnpm skills:init <skill-name> --path skills --resources scripts,references
pnpm skills:finalize -- skills/<skill-name>

Mode B: Finalize Existing Skill by Path

For the selected finalize-path action, run from the current repository root. Relative and absolute targets must resolve to a direct child of this repository's skills/ directory:

pnpm skills:finalize -- <skill-path>

Examples:

pnpm skills:finalize -- skills/code-inspector-init
REPO_ROOT="$(git rev-parse --show-toplevel)"
pnpm skills:finalize -- "$REPO_ROOT/skills/code-inspector-init"

Finalize pipeline is fixed and must run in order:

1. pnpm skills:quick-validate skills/<skill-slug>
2. pnpm skills:validate
3. pnpm skills:index

This mode does not stage files.

Mode C: Validate Existing Skills

Use for “check”, “audit”, “validate”, or a bare invocation without a requested creation/finalize action:

pnpm skills:validate

When source changes already exist and generated index freshness is part of the check, also run:

pnpm skills:index
git diff --check

Do not call skills:new, skills:finalize:new, or git add in this mode.

Mode D: Auto Discover + Finalize + Stage

Run only when the user explicitly asks to finalize and stage newly added skills:

pnpm skills:finalize:new

Behavior contract:

1. Discover newly added skill candidates from git status --short --untracked-files=all:
   • include A (index-added) and ?? (untracked)
   • only accept slugs where skills/<slug>/SKILL.md itself is in A or ??
2. Finalize each discovered slug in stable sorted order:
   • pnpm skills:finalize -- skills/<slug>
   • stop immediately on first failure
3. Stage only related files after all finalize runs succeed:
   • git add skills/<slug> for each processed slug
   • git add apps/web/src/generated/skills-index-lite.json apps/web/src/generated/skills-detail-index.json for whichever generated index files changed
4. If no new skill is discovered:
   • report that no new skill was found
   • do not create a skill unless the user also requested creation

Scope Boundaries

• Do not auto-run pnpm skills:openai-yaml unless explicitly requested.
• Do not auto-run local install/sync (skills:install:local, skills:test:local) unless explicitly requested.
• Do not stage in create, finalize-path, or validate mode. Staging requires an explicit finalize-and-stage/commit-preparation request.
• Do not use git add -A; only stage skill-related files from this workflow.

Output Contract

When executing this skill, always return:

1. Commands executed (or planned commands in dry-run).
2. Success/failure status.
3. Detected/processed skill slugs and staged file paths, with none when the selected mode does not stage.
4. Next-step suggestion only when useful (for example, local agent testing via pnpm skills:test:local).

Skill Rules

1. Use lowercase hyphen-case skill names.
2. Keep SKILL.md frontmatter valid with non-empty name and description.
3. frontmatter.description must be English-only (ASCII characters only), and should clearly include when to use the skill.
4. Add optional directories (scripts/, references/, assets/) only when needed.
5. Re-run validation and index generation after updates and before commit.