AI agents are quick to say "Done! All tests pass." Sometimes they ran the tests. Sometimes they ran them before the last change, or ran the linter and assumed the rest. This skill from Jesse Vincent's Superpowers collection has one rule: no claim that work is finished, fixed or passing until the agent has run the command that proves it, in the same message, and read what it printed.
What it does
Before any claim about the state of the work, the agent goes through five steps: work out which command would prove the claim, run all of it, read the full output and exit code, check that the output actually says what it's about to claim, and only then say it, with the evidence.
The skill spells out what counts as proof:
| Claim | Proof | Not proof |
|---|---|---|
| Tests pass | Test output showing 0 failures | An earlier run, "should pass" |
| Build succeeds | The build command exiting 0 | The linter passing |
| Bug fixed | The original symptom, now gone | "Code changed, assumed fixed" |
| Sub-agent finished | The diff showing its changes | The sub-agent saying so |
It also lists warning signs in the agent's own wording, like "should", "seems to", or "Great!" before anything has been checked. The core of it is one line:
If you haven't run the verification command in this message, you cannot claim it passes.
From SKILL.md by Jesse Vincent, MIT.
An example
You ask the agent to fix a failing date test. It edits the code. Without the
skill, the next message is often "Fixed! The test should pass now." With it,
the next message runs the test, shows 12 passed, 0 failed, then runs the
whole suite because the change touched a shared helper, and only then says it's
fixed, quoting both results.
When to use it, and when not to
Use it everywhere you let an agent change code, and especially before commits and pull requests. It costs a few seconds per task and catches the most common way agents mislead people.
There's little reason to leave it out, but if your project has no tests, no build and nothing to run, it has nothing to check with.
What we checked
We read every file in the skill's folder at the commit linked above: just
SKILL.md. There are no scripts, no network access and no allowed-tools, and
nothing asks the agent to skip a confirmation or read anything outside your
project. The Superpowers repo is MIT licensed and updated most weeks. Our full
checklist is in the
guide to checking a skill before you install it.
It works well with the Karpathy guidelines skill, which asks the agent to define what "done" means before it starts.