The always-on governor against over-claimed completion. Never call work done, fixed, passing, working, wired, or complete unless you RAN the actual Done criterion end-to-end and observed the result with your own eyes. Code that would work, a traced code path, a passing adjacent layer, or "looks right" are NOT done. Applies silently to every completion, success, or "it works" claim; dispatches the sibling vd-* skills.
SkillScore92Maintained
Automatic
◆QA & Test Engineering pack
Last verified 8/3/2026
What it changes
Passed 23 of 25 test cases with the skill, and 14 of 25 without it.
Without skill56%
With skill92%
Checks it passed that the model alone missed
The response forbids labeling the cache manager memory leak fix as complete.
The response classifies direct database updates as adjacent-path verification.
The response tags reviewing token validation logic as code-path or not executed.
and 8 more in the full report
9 fixed14 passed either way2 still failing0 broke
Pass-rate lift
+36 pp
Tokens
-8%
Agent turns
0%
Response time
-39%
Model gemini-3.6-flashMeasured 8/3/2026v123 of 25 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 9 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 9
Prompt
I am writing a status update for our database migration task on `users-db`. I successfully ran `pg_dump` on production, verified the backup file size and integrity check, restored it to a staging instance, and queried the row counts to match production. How should I report this in Slack?
What was checked, without and with the skill
Checked
Without skill
With skill
The response confirms that reporting the database backup and staging restore as complete and verified is accurate because all restoration steps were executed and validated.
Missed
Passed
case-03 · gemini-3.6-flash · 8/3/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md24 linessingle file
Done Means Executed — Not Inspected
The failure this prevents: asserting something is finished based on reading the code that would run, or on a related path that does run, instead of running the real thing. The tell is a sentence like "skills-wired and persona-wired are both code-path claims I didn't execute."
The rule
A completion word — done, fixed, passing, working, wired, complete, verified — is a claim that you ran the actual Done criterion and saw the expected result. If you didn't run it, you may not use the word. Say what you actually did instead.
These are NOT done:
Code-path verified — you read the code and it should do X. (You traced it; you didn't run it.)
Adjacent-path verified — a lower layer or a different entry point works, so the real one "must." (See [[vd-plumbing-vs-behavior]].)
Looks right — the strongest signal that you should run it. (See [[vd-run-the-real-path]].)
The loop (run on every completion claim)
Name the real Done criterion. What exact observable behavior counts as done? Quote it if it's written down.
Did I execute that? Not a proxy, not the plumbing — the actual user-facing path. If no → run it ([[vd-run-the-real-path]]) or downgrade the claim.
Build the ledger. Tag each sub-claim ✅ executed / ◐ code-path / ○ assumed, and state what each does NOT establish ([[vd-claim-ledger]]).
Report honestly. Only ✅ across every criterion earns "done." Anything else: report the gap plainly, first, before you're pushed ([[vd-report-the-gap]]).
Silent discipline
Run this in your head on every "it works." Don't narrate the framework — change what you verify and how confident you sound. Surface the reasoning only when a claim is load-bearing or you're shipping the gap.
A passing lower layer or adjacent entry point does not prove the user-facing behavior. Verifying a direct tools/call over stdio is not the agent deciding to call the tool with its persona; a green unit test is not the feature working; a reachable endpoint is not the flow succeeding. Use when tempted to generalize "the plumbing works" into "the feature works" — identify the actual user path and exercise that exact path.