prove-it-works
Verify fixes, features, and refactors in the running system using verify-<app>, otherwise agent-browser or tuistory, plus tests. Use before finishing; use /review or adversarial-review for diffs, or simplify for reuse, quality, and efficiency.
Cynnamoroll · AI review on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add Cynnamoroll/factory-stack --skill prove-it-worksAbout this skill
name: prove-it-works description: "Verify fixes, features, and refactors in the running system using verify-, otherwise agent-browser or tuistory, plus tests. Use before finishing; use /review or adversarial-review for diffs, or simplify for reuse, quality, and efficiency." license: MIT compatibility: droid
Prove it works
Check the real artifact directly. Compilation, self-reports, proxies, and unit tests alone do not prove the user-facing result.
1. Name the claim
State one user-checkable claim, such as "saving a note with a 500-character title keeps the full title after reload." For bugs, use the original repro.
2. Pick the surface
Drive the change on the surface a user touches:
- the project's verify skill, if
.factory/skills/verify-*/exists. Follow its Launch, Doctor, Drive, Evidence, and Cleanup sections. - otherwise Droid's built-in
agent-browserskill for web and Electron apps,tuistoryfor terminal apps, orcurlfor HTTP services - for a library, a small script that imports the built package and calls the changed API the way a consumer would
Wrong surface or "inconclusive" is not a pass. Say which it was.
3. Run it
- Run the project's tests that cover the change (from
AGENTS.mdor the package scripts). Rerun only what failed plus what the fix touched. - Drive the claim on the chosen surface. Capture the action and the resulting state: a screenshot, a terminal transcript, a response body, a row read back.
- Check side effects as well as what is visible: files written, rows stored, messages sent.
- When a check fails, suspect the observation method first, then the system.
Prefer a script that reruns the same comparison over a one-time look. Keep it under /tmp/ unless it is worth committing as a regression test.
4. When it fails
Fix the cause and rerun step 3 without weakening assertions or skipping steps. If credentials, production-only state, or hardware block the run, name the blocker and the closest check's result.
Done when
The reply states the claim, the surface used, the commands run, and the evidence (pasted output or artifact paths). A bug fix shows the repro failing before and passing after. Otherwise the reply says why the change could not be exercised and what was checked instead.