--- name: create-verification-skill description: Project skill that drives the real app and proves behavior. /create-verification-skill. disable-model-invocation: true --- # Create Verification Skill Scripted real-app proof: launch, user-path drive, evidence. Write a cold-agent skill at `.cursor/skills/verify-/`. ## Interview (repo first) Ask only what code cannot answer: - **Surface** — primary UI/CLI/API/desktop/mobile/lib; note others - **Run** — repo start command; ports, env, seed, auth - **Drive** — existing harness first; else CDP/tmux/HTTP - **Observe** — screenshots, transcripts, bodies, logs, exits, DB - **Isolate** — side-by-side? If not, refuse double-drive Fix a broken checkout before generating. Verification-only scaffolding OK if cleanup removes it. ## Generate `SKILL.md` with frontmatter (`name: verify-`, app + surface in description). No placeholders. - **Launch** — exact start + usable readiness; migrations/seed; teardown. CLI: build once; each drive in its own PTY/tmux. - **Doctor** — read-only process/build/port/auth check. - **Drive** — real selectors/commands; stable handles. - **Evidence** — what/where; real user path; action + result; side effects; mocks only at prod boundaries; observe dry-run skips. - **Cleanup** — kill what you started; keep evidence at a named path. - **Helpers** — executable; invocation in the skill body. ## Feature map `features/README.md` + one file per top feature (start 3–5). H2s: `Sub-features`, `How to get to it (user POV)`, `Driving it with `, `Gotchas`. Cover reach, drive, observable proof. Missing seed → fail + Gotcha; never weaken. Map is source of truth. ## Prove then hand off Once: launch → doctor → one mapped feature → evidence → cleanup; evidence must survive. Cleanup after failed tries. Stand-ins: list unverified setup. Point at `/maintain-verification-skill`.