Create Skill Repo

create-skill-repo

Stamps out a new skills repository: gh repo create provisions it on GitHub (private by default, MIT license, technology .gitignore, README) and clones it into the current directory; the bundled template (assets/template/ inside this skill) is then overlaid — OSS hygiene files, CI with a skills.sh consumer job, validation hooks, in-repo add-skill/publish-repo dev skills, plugin/marketplace manifests, skills.sh.json, and a gitignored .local/PROMPT.md that drives the follow-up authoring session. The scaffold overlay is left uncommitted (GitHub's repo creation makes only its own initial commit): the owner reviews, commits, pushes. Reference: . If invoked as /create-skill-repo , $ARGUMENTS is [idea].

When NOT to use

• Writing or fixing a skill inside an existing repo → add-skill (or that repo's

bundled copy).

• Turning a research pack into skills → skill-from-research (it calls this

skill when a new repo is actually needed).

• Improving the scaffold itself → edit assets/template/ in a skillskit checkout

(changes affect future repos only).

Workflow

1. Derive the four inputs from the user's request; ask only if genuinely absent:

• name: kebab-case repo name. Single skill → the skill's name

(icon-designer); collection → a plural theme (python-skills).

• description: one professional, benefit-led sentence with plain category

keywords — it becomes the GitHub repo description, README intro, and plugin manifest. ≤ ~250 chars, no hype words.

• idea: the user's high-level intent, 1–4 sentences, as literally as

possible — it seeds .local/PROMPT.md for the in-repo authoring session.

• gitignore: the GitHub .gitignore template matching the skills' dominant

technology (Python default — the scaffold's own tooling is Python).

1. Scaffold (deterministic — always via the script, never hand-copy or raw gh):

``bash python3 "${CLAUDE_SKILL_DIR}/scripts/scaffold.py" \ --description "" --idea "" [--gitignore ] ` Options: --owner (default: the authenticated gh user), --public (default private), --topics extra1,extra2 (added to agent-skills,claude-code,skills,skills-sh), --dest DIR (default: cwd), --local-only (offline: folder + git init, no GitHub; requires --owner), --keep-on-fail. The script validates inputs (including that the seeded .local/PROMPT.md goal fits 4000 characters — shorten the idea if it fails), checks gh auth (needs the workflow` scope), refuses names taken locally or on GitHub, validates the scaffold, and cleans up the local clone on failure. It NEVER runs git commit/push (the owner does) and never deletes remote repos — on failure it prints the exact cleanup command instead.

1. Verify: success ⇔ the script's final line is SCAFFOLD OK: (exit 0).

On failure, read its report, fix the cause (never by weakening the generated repo's scripts/validate_skills.py), and re-run.

1. Hand off — tell the user exactly:

`` cd && claude # then paste the contents of .local/PROMPT.md (starts with /goal) ` Mention: the repo shell is live (private) on GitHub but the scaffold is uncommitted — review, commit, push (exact commands are in the script output); .local/ is gitignored — drop research packs/sources there before starting; publishing to skills.sh later runs via the repo's bundled /publish-repo`.

Output spec

New folder / under the current directory: a clone of the freshly created GitHub repo (private, MIT, tech .gitignore merged with the template's), overlay left uncommitted for owner review, validator passing, repo topics set, .local/PROMPT.md filled with the idea (≤4000 chars, enforced), no leftover {{PLACEHOLDER}} tokens anywhere.

Gotchas

• Requires an authenticated gh (gh auth login) with the workflow scope; the

script prints the exact gh auth refresh command if missing.

• Names must match ^[a-z0-9]+(-[a-z0-9]+)*$ and must not contain

"claude"/"anthropic" (reserved by the skills spec); free locally and on GitHub.

• --gitignore values are case-sensitive GitHub template names

(gh api gitignore/templates); the script pre-checks them.

• The template ships its gitignore as _gitignore (a live one inside skill

assets would be applied by the host repo's git); the script restores the real name during overlay — don't "fix" it in the assets.

• .local/PROMPT.md is personal and machine-local — no committed copy exists in

generated repos.

Files

• scripts/scaffold.py — the deterministic scaffolder (stdlib only).
• assets/template/ — the complete repo scaffold that gets overlaid.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Paldom
• Source: Paldom/skillskit
• License: MIT

Install and usage instructions live in the source repository linked above.