SuperAPI AI Directory

setup-repo

Deployment & operations

View repository

Updated 10-15 22:14

About this skill

Set up this repo in one run

One-run entry that makes a repo usable by the engineering skills. A thin orchestrator: it drives the two setup primitives in order and contains only order and routing facts — never restate their steps or copy their content.

Standard answers (editable — this user's repos always pick these)

  • Issue tracker: GitHub
  • Empty repo (no remote, may not even be a git repo): create a GitHub private repo from the current directory first, so the GitHub default still holds (step 2)
  • Triage labels: the default vocabulary (needs-triage, needs-info, ready-for-agent, ready-for-human, wontfix)
  • Agent file: edit the one that exists — CLAUDE.md if present, else AGENTS.md; never create the other
  • Domain docs: single-context (GLOSSARY.md + docs/adr/ at the root) unless monorepo signals say otherwise

Process

1. Explore

git init when the directory has no .git — git remote -v errors outside a repo. Then run the exploration step: git remote -v, AGENTS.md/CLAUDE.md at the root, GLOSSARY.md/GLOSSARY-MAP.md, docs/adr/, docs/agents/, .scratch/, monorepo signals, and whether the triage skill is installed.

2. Ensure a remote when the repo has none

When exploration found no remote at all, create one before the tracker setup:

  1. Directory is already a git repo (step 1).
  2. No commit yet? Create a minimal initial commit (README.md or .gitignore) — empty push fails (src refspec HEAD does not match any).
  3. gh repo create <dir-name> --private --source=. --push (Issues on by default; remote = dir name).
  4. Failure (gh unauthenticated / network / name clash) → fall back to asking (GitHub / GitLab / local markdown / other).

The remote must exist before the primitive runs — with a GitHub remote its Section A proposes GitHub, and the local-markdown fallback never fires.

3. Run setup-matt-pocock-skills with the defaults pre-filled

Follow its process, but apply the standard answers above instead of asking each section. Ask only when exploration evidence contradicts a default or evidence is missing, and lead each asked section with the recommended answer:

  • Remote is GitLab → propose GitLab; no remote at all → step 2 already created a GitHub private repo — the GitHub default holds.
  • Only CLAUDE.md exists, or neither agent file exists → the agent-file choice needs the user.
  • Monorepo signals present → offer the multi-context layout.
  • triage not installed → skip its section silently.
  • Prior output already exists in docs/agents/ → confirm before re-running (a re-run is only needed to switch trackers).

Its own rule stands: never create AGENTS.md when CLAUDE.md already exists (or vice versa).

4. Then run setup-coding-standards

It writes CODING_STANDARDS.md at the repo root (the baseline verbatim, plus verified repo conventions). Its guardrail is inherited unchanged: if the repo already has a standards file, it stops and reports — write nothing.

5. One combined confirm, then write

Present a single draft covering everything: the ## Agent skills block, the docs/agents/* files, and the CODING_STANDARDS.md content. Let the user edit before writing. Then write it all.

6. Done

Report what was written and where. The primitives remain available on their own when only part of the setup is wanted (e.g. setup-matt-pocock-skills alone for a throwaway repo).

Out of scope

  • Extending the ## Agent skills block shape or adding sections the primitives don't write: the primitives are the single home for their outputs.
  • Anything an existing standards file already covers: the human decides what changes.