authoring-a-playbooklisted

# Authoring a playbook A bento playbook is a triggered skill that opens with a pasteable checklist and proves its work on a real artifact. Keep them few and earned. ## Gate first — should it exist? Add a playbook only when the same task shape has recurred (≥2–3 times) and you can name a concrete trigger. One occurrence → do it inline via `figure-it-out`, don't skill it. Speculative playbooks are cost paid by every future reader. Default to not adding. ## Steps (paste as todos) 1. **Name the trigger.** When does this fire? Write it as the `description` with concrete trigger phrases — that's what makes the skill auto-invoke. 2. **Write the checklist.** Numbered, ordered, checkable steps the agent pastes as todos. Lead with them; prose is secondary. 3. **Demand proof.** Include a step that verifies on a real artifact (transcript, test, file, before/after) — never a self-report. 4. **Add guardrails.** The 2–3 failure modes this playbook prevents. 5. **Keep it generic.** bento playbooks must help in any repo. Repo/org-specific detail belongs in that repo's L2, not here. 6. **Prove it lifts.** Add a `claude plugin eval` case (with/without arm) and run `claude plugin eval` — keep the playbook only if it changes behavior. No lift → don't ship. ## Guardrails - Fix the principle, not the example. Minimal GOOD/BAD pairs; add a BAD example only when the instruction alone isn't enough. - No verbatim production data, no dates/history in the body — state current scop