| name | reviewer-brief |
| description | Frame review output by severity, name the violated invariant, cite evidence, list missing tests, and state residual risk. Use when reviewing, auditing, critiquing, or producing a verdict. |
Reviewer Brief
A review is most useful when it is structured. Severity, evidence, and an explicit verdict are the three things a generator needs to act on a review without re-reading it.
When to use
- Producing a review, audit, critique, or sanity-check verdict.
- Asked to evaluate work against a stated contract.
- Reading a generator's output and deciding what to do with it.
Procedure
- Read the work end-to-end before commenting. Form a model of what the change is trying to do; do not respond to the first line before understanding the last.
- Compare to the stated contract. Does the work do what it said it would? Note unrelated edits as separate items, not blockers.
- Group findings by severity. Must-fix (blocking), should-fix (defect, non-blocking), nit (taste). Sort within each group by impact.
- Name the violated invariant. "This is wrong" is unhelpful. "This re-enters the cache during concurrent writes" is actionable.
- Cite evidence. file:line or quoted passage. The author should not have to guess what you meant.
- List missing tests. New behavior without a test, bug fix without a regression test — call it out by name.
- State residual risk. What might still break that this review did not check.
- End with a verdict. Ready / changes requested / discussion needed. No verdict is the worst review.
Output
Findings by severity, each with invariant · evidence · suggested action. Final verdict line. Do not quote the work back to the author.
Anti-patterns to avoid
- Mixing taste, defects, and blockers in one undifferentiated list.
- Modifying the work yourself — Verifier-class personas describe what should change; the Generator does the change.
- Verdict-less reviews; the author cannot ship without one.