Verify done

Skill brain-marchine/verify-done-skills/skills/verify-done

Fail-closed verification before accepting Done/Fixed/做完了. Use when the user or agent claims a task is finished, fixed, complete, verified, or ready to merge — including UI fixes, bug fixes, overflow, layout, tests passing, or "I'm done". Also trigger for 做完了、修好了、验收、验证完成、可以合并. Runs bundled check scripts and refuses verbal Done until PASS.From its SKILL.md

Install

npx -y skills add brain-marchine/verify-done-skills --skill verify-done

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.1 KB, 719 tokens by cl100k_base, as published. Nobody here has run it

verify-done

Agent said Done. Reality must agree.

You MUST run verification before saying the work is finished. Verbal "Done / Fixed / 做完了" without a PASS from the scripts below is Fake Done — forbidden.

When this skill applies

User or you claim: Done, Fixed, complete, finished, verified, LGTM, ready to merge
Chinese: 做完了、修好了、弄好了、验收通过、可以提交
Any UI / layout / overflow / bugfix claim

Hard rules

Do not say Done / Fixed / 做完了 until the check script exits 0 (PASS).
On FAIL (exit 1): show evidence, keep fixing, re-run the script.
Prefer running the bundled script over inventing your own checklist.
If the script path is unclear, locate SKILL.md for this skill and run scripts/check.mjs next to it.

Procedure

1. Locate this skill

Common install paths:

Project Cursor: .cursor/skills/verify-done/
Project Claude: .claude/skills/verify-done/
User Cursor: ~/.cursor/skills/verify-done/
User Claude: ~/.claude/skills/verify-done/

2. Run the check (from the user's project root)

node "<path-to-verify-done>/scripts/check.mjs" --cwd . --claim "<exact claim text>"

Windows PowerShell example:

node "$PWD\.cursor\skills\verify-done\scripts\check.mjs" --cwd . --claim "Fixed the inventory overflow"

Optional demo URL check (when a local page should prove a UI fix):

node "<path-to-verify-done>/scripts/check.mjs" --cwd . --claim "Fixed overflow" --url http://127.0.0.1:4173

3. Report in chat (required template)

## verify-done: FAIL | PASS
- Claim: ...
- Checks: git_diff | fake_tests | npm_test | url_probe (list which ran)
- Evidence: (paste key lines from script stdout)
- Next: (if FAIL, concrete fix steps; if PASS, you may say Done)

4. Only after PASS

You may say the work is done. Summarize what passed.

What the script checks

Check	Behavior
`git_diff`	Claimed a fix but no file changes → FAIL
`fake_tests`	Vacuous tests (`expect(true).toBe(true)`, empty bodies, `it.todo`) → FAIL
`npm_test`	If `package.json` has `test` script, runs it (skip if missing)
`url_probe`	If `--url` given, FAIL when page still contains `overflow-bug` / `LAYOUT BROKEN`

Details: references/fake-done.md

Anti-patterns (do not do)

"Tests are green in my head" without running the script
Editing the script to force PASS
Saying Done because the user sounded impatient
Skipping verify on "tiny" UI claims — those are Fake Done hotspots

What ships with it: 2 files

7.4 KB alongside SKILL.md, 1 of them executable

references/

fake-done.md608 B

scripts/

check.mjsruns6.8 KB