Créer un nouveau skill Claude

---

description: Scaffold a new Claude command/skill in .claude/commands/ following this repo's house format (frontmatter + steps + verify + final report). argument-hint: <skill name + one-line purpose> allowed-tools: Read, Write, Glob

Scaffold a new skill: $ARGUMENTS. Skills in this repo live as .claude/commands/<name>.md and are the project's slash commands. Keep the new one token-light — it should point agents at the right subsystem doc rather than re-explaining architecture. (Scaffolding an AGENT definition in .claude/agents/ instead? That's /add-agent.)

Read first

1. An existing sibling that resembles the new skill's shape (add-building.md for a "add a game thing" workflow; smalltweak.md for a git-flow workflow) — match its structure, don't invent a new one.

Steps

1. Pick a kebab-case name; the file is .claude/commands/<name>.md.
2. Write the frontmatter (three keys, same order as siblings):
   • description: one sentence, imperative — this is what shows in the skill list.
   • argument-hint: <what the user passes>.
   • allowed-tools: the MINIMAL set. Read/Edit/Write/Grep/Glob for edit workflows; add narrowly-scoped Bash(...) entries only for the exact commands the skill runs (e.g. Bash(py tools/smoke.py*), Bash(py -m pytest*), or specific Bash(git ...) verbs). Do NOT grant a blanket Bash, and never Bash(py -m unittest*) — the pre-pytest incantation runs everything, is not the gate, and the test_guard.py hook denies it.
3. Write the body in the house shape:
   • a one-line restatement using **$ARGUMENTS** + any lock/scope caveat;
   • Read first (token-light) — name the ONE subsystem doc to load (engine|game|editor/<sub>/CLAUDE.md) instead of pasting architecture;
   • Steps — numbered, smallest-change-first;
   • Avoid — the repo-specific foot-guns for this task;
   • Verify — the narrowest check (py tools/smoke.py / the specific py -m pytest tools/tests/test_<area>.py -q files / live run), and "state what you verified". Route to §"Test Suite Policy" in the root CLAUDE.md; never state a different rule in the new skill. That section is the ONLY authority on when tests run — a skill that says "run the full suite", writes a per-step testgate check, or tells a subagent to use --affected is a bug, and the mechanical parts are hook-denied anyway. If the skill is a closing/handoff move, say the single full check belongs to the MAIN SESSION, once;
   • Final report — changed files + verification + whether a subsystem doc needed a durable update.
4. Keep it short. If it's growing past ~60 lines, the detail probably belongs in a subsystem CLAUDE.md that the skill links to.

Verify

• The file parses as valid front-matter markdown (three ----fenced keys) and reads cleanly. It should appear as /<name> after the session reloads skills.

Final report

• The new file path; its allowed-tools; which subsystem doc it routes to.
• Tag every claim measured / verified / inferred (see /report).