Verification Before Completion

Verification Before Completion

A gate. You don't claim "done" without fresh, executed proof. Adapted from obra/superpowers.

Core principle

No completion claims without fresh verification evidence.

Confidence is not evidence. A previous green test run is not evidence. A successful agent report is not evidence. The diff "looking right" is not evidence.

Evidence is: a command run in the current state of the code, its output read, and the output supporting the claim.

When to activate

Before saying or writing any of:

• "Done."
• "Fixed."
• "Works."
• "Passing."
• "Complete."
• "Ready to commit / merge / ship."
• "Should be good."
• Any positive close that asserts the work succeeded.

Also before:

• Running git commit after a non-trivial change.
• Opening a PR.
• Marking a task as completed.
• Handing off to the user as finished.
• Switching to the next task in the same turn.

The gate

1. Identify the proof command. What single command (or short sequence) would demonstrate the claim is true?
   • Tests: pnpm test path/to/file.test.ts, pytest tests/test_x.py -v
   • Type check: pnpm tsc --noEmit, mypy .
   • Build: pnpm build, cargo build
   • Behavior: run the app and exercise the path (use qa-manual skill)
   • Lint / format: pnpm lint, ruff check
2. Run it fresh. Not "I ran it earlier." Run it now, in the current tree.
3. Read the full output. All of it — including warnings and partial failures. Don't grep for "PASS" and call it done.
4. State the conclusion with evidence. Quote the relevant lines when claiming success. If output is large, summarize and reference the path to the log.

Hard rules

• No "should" without proof. "This should fix it" → run it.
• No "probably" / "seems to" / "looks right." Those are confidence signals, not evidence. Run something.
• Linter passing ≠ code working. It means the lint rule didn't trigger.
• Type check passing ≠ logic correct. It means types align.
• A subagent reporting success is not verification. You must independently run the proof command.
• Tiredness does not justify skipping. If you're tired, stop. Don't ship unverified.

Common failure patterns to flag and refuse

Output format

After the gate clears, produce:

• The command that was run.
• The relevant output (quoted or referenced).
• An explicit Verified: or Not verified: statement.

Example:

Verified.

Ran: pnpm test src/auth/oauth.test.ts

Output: Tests: 12 passed, 12 total. The case redirects on missing state param (the original bug) now passes.

If the gate fails:

Not verified.

Ran: pnpm test src/auth/oauth.test.ts

Output: 1 failing — redirects on missing state param: expected 302, got 500. Need to handle the null branch in callback.ts:42 before claiming done.

Pair with

• qa-manual — for behavior verification of web features (the proof command is "drive the app in Chrome").
• verify-pr — heavier verification scoped to a whole PR.
• miguel-review — review the diff for shape; this skill verifies behavior.