Skill · Coding

verify-before-done

Proves a task is finished before claiming it: restates the acceptance criteria, runs tests, typecheck, lint and build, exercises the changed path for real, checks for leftover to-do markers, debug logs and skipped tests, and reports evidence per criterion plus what wasn't verified. Use before saying a task is done, fixed or working, before committing or opening a PR, or when asked to 'double-check', 'verify' or 'make sure it works'.

The skill itself. Agents read the name and description first and load the rest only when a task matches.

verify-before-done/SKILL.md
---
name: verify-before-done
description: "Proves a task is finished before claiming it: restates the acceptance criteria, runs tests, typecheck, lint and build, exercises the changed path for real, checks for leftover to-do markers, debug logs and skipped tests, and reports evidence per criterion plus what wasn't verified. Use before saying a task is done, fixed or working, before committing or opening a PR, or when asked to 'double-check', 'verify' or 'make sure it works'."
---

# Verify before done

Before you say a task is done, fixed or working, prove it.

## Steps

1. Restate the acceptance criteria as a numbered list. If none were given, write the ones the request implies and say that you inferred them.
2. Run the project's checks with the commands the repo already defines (package scripts, Makefile, CI config), and paste the key lines of output (counts, failures, exit status), not the whole log:
   - tests: the affected ones, then the full suite if it's reasonable to run;
   - typecheck;
   - lint;
   - build.
3. Exercise the changed path the way a user would: run the CLI with real arguments, call the endpoint and show the response, open the page and check the changed state, run the migration on a local database. Record what you saw.
4. Map each criterion to the evidence that proves it. A criterion without evidence is not met.
5. Search the diff (`git diff` against the base branch) for leftovers: to-do or FIXME markers you added, debug logging, commented-out code, skipped or focused tests (.skip, .only, xit, fit), hard-coded test values, temporary files.
6. If a check fails and you think it was already failing, prove it: check out the base branch in a separate worktree (`git worktree add ../base-check <base-branch>`), run the same command there, then remove the worktree with `git worktree remove ../base-check`.
7. Report: each criterion as met or not met with its evidence; the checks run and their results; what you did not verify and why (no credentials, no browser, slow end-to-end suite); anything the user should check themselves.

## Rules

- Never write "should work", "should be fixed" or "I believe this works". Report the evidence or name the gap.
- If a check fails because of your change, the task is not done. Fix it or report the failure plainly.
- Don't weaken a test, skip a check or loosen a lint rule to get green unless the user agrees.
- Only run commands that are safe here. No deploys and no writes to shared data.

## Example output

```
1. Expired coupons rejected: met. "rejects expired" passes; POST /cart/coupon with code OLD10 returned 422 {"error":"coupon_expired"}.
2. Valid coupons still apply: met. 13/13 existing coupon tests pass.
Checks: npm test 212 passed; tsc 0 errors; eslint 0 errors; build ok.
Leftovers: none found.
Not verified: receipt email wording (no mail sandbox in this environment).
```