Skip to content

sno-ai/sno-station-skills / skills/J-peer-review-and-audit/peer-review/skill

peer-review

Verified

Adversarial review and fix loop for code, specs, plans, and migration manifests.

assets/adversarial-test-review.prompt.md

View on GitHub
You are reviewing TEST MATERIAL. This is a different review from a production
code review, and doing the production review here is the failure mode this
prompt exists to prevent.

A test exists to prove one thing about something else. It is not a product, it
has no users, and nobody depends on its internals. **Review it against its own
purpose and nothing else.**

## The only question

**Does this test actually prove the thing it claims to prove?**

Everything below is a way of asking that question, not a separate checklist.

1. **Does it exercise the behaviour it names?** Read what the test says it is
   for — its name, its description, its assertions — then read what it actually
   calls. A test whose name promises one thing and whose body touches another is
   the single most valuable finding here, because it reads green forever while
   proving nothing.

2. **Can it fail for the right reason?** Construct the defect this test is
   supposed to catch and walk it through the test as written. If the test still
   passes with that defect present, say so and name the defect. A test that
   cannot go red is not a weak test; it is not a test.

3. **Can it fail for a WRONG reason?** Does it pin something the change under
   review does not control — sampled model output, wall-clock time, network
   order, a path that only exists on one machine? Those produce red rows that
   cost hours and teach the team to ignore the suite.

4. **Is the proof boundary the lowest one that works?** A behaviour provable by
   a unit assertion, proven by a live end-to-end run, is expensive coverage
   bought at the wrong layer. Say which layer would be sufficient, and only when
   the difference is real.

5. **Is this already proven elsewhere?** Duplicate proof is worse than no proof:
   it costs the same to maintain and it lends false confidence about breadth. If
   you can see the sibling test that already covers it, name it.

6. **Does the setup guard fail loudly?** A fixture that returns nothing when its
   precondition is missing turns every downstream assertion into a vacuous pass.
   If a missing key, an absent file, or an empty input can make this test green,
   that is a finding.

## Explicitly OUT OF SCOPE — do not report these

- The test's own robustness, error handling, or edge cases. It is scaffolding.
- Style, naming, formatting, comment density, helper hygiene, DRY-ness.
- Refactoring suggestions, abstractions, parameterisation, "consider extracting".
- Performance of the test itself, unless it cannot finish inside its own timeout.
- Anything about the production code the test exercises. That is a separate
  review with a separate prompt; mention it in one line at most and move on.
- Speculative additional cases the test "could also cover" — unless their absence
  means the stated purpose is not actually proven.

If your findings list is mostly items from this section, you have run the wrong
review. Delete them and answer the only question instead.

## Output

Same structure the other reviews use: a verdict, then findings, each carrying a
**Trigger** (the realistic sequence that reaches the problem), a **Consequence**
(what a reader of this suite would wrongly believe, and what it costs), and a
**Refutation** (an honest attempt to kill your own finding).

**This is one pass.** There is no second round on a test file. Say everything
worth saying now, and nothing that is not.

If the test is sound, say so plainly in one or two sentences and stop. A short
clean report is the expected outcome for most test files and is not a sign the
review was shallow.
receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementationStars: 323kUpdated Claude CodeCodexVerifiedInstall
requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirementsStars: 323kUpdated Claude CodeCodexVerifiedInstall
using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallbackStars: 323kUpdated Claude CodeCodexVerifiedInstall
verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions alwaysStars: 323kUpdated Claude CodeCodexVerifiedInstall