Structure
Create your company
Skill

review

when a diff, draft, or proposal needs a verdict before a signature — and when findings come back on work of yours

Gives a verdict on a diff, draft or proposal, or helps you answer findings on your own work. Every finding has to name the concrete way the work fails; without that it is an opinion.

structure generate skill review

Skill — review

Invoke in any agent conversation with /review, from either seat: giving (the verdict is yours to write) or receiving (the findings are about your work). One law governs both, and the steps are a starting opinion; edit them until they read the way this company reviews.

When to use. Before anyone signs a gated proposal. When an agent hands work back and "looks good" is about to be typed. When a review of your own work lands and the first instinct is to agree with all of it.

The Iron Law. A finding without a failure scenario is an opinion. A finding earns its rank only by naming the concrete way the work fails — these inputs, this state, this wrong outcome — and anything that cannot say that rides at the bottom as a note, unranked, blocking nothing.

Giving: read the evidence before the work. On a gated proposal the approve card already carries the bundle, in order, and the review follows that order rather than inventing its own:

  • Intent — the author's own purpose. First question, always: does the diff do what this line says it is for, and nothing it does not say?
  • Checks — what verification introduced. Silence means still checking; "could not be read" is not clean; and introduced errors are never approved past without the decision saying why.
  • Revisions — the plan moved. Read the chain before trusting a memory of an earlier version, because whoever read this card last week read a different proposal under the same title.
  • Blast radius — how far the signature reaches: the documents and their rungs, the files that moved beneath them, the open proposals this lands on top of. Wide radius, slower yes.

Then the diff itself, against the standards the company wrote down — the owning department's steering docs first (under engineering, departments/engineering/coding-standards.md), never taste.

The findings. Ranked by consequence, worst first, each carrying four parts: the claim, the failure scenario, the evidence (file and line, or the offending span quoted verbatim), and the smallest fix that removes the failure. A finding that cannot cite its span gets investigated until it can, or dropped. Style notes and preferences go in a final unranked section, clearly not findings.

Receiving. Reproduce the failure scenario before writing the fix — a finding you cannot reproduce goes back as a question carrying what you ran, not as a silent fix and not as a rebuttal. Agree by fixing, disagree with evidence; reflexive agreement is how a wrong finding becomes a wrong change.

Where the verdict lands. A review that concludes a gated document should change is a proposal: structure propose <path> -m "title", never a direct write past the gate the review exists to feed. The approve stays a human's click, on the same card, with the same evidence.

The refusal. Asked to promote style nitpicks to blockers, or to pad the list so the review looks thorough, this skill refuses to rank them as findings: a nitpick names no failure scenario, so it rides as a note and never blocks an approve. And from the other seat it refuses the rubber stamp — no verdict is given without the evidence read, because a signature over an unread bundle is the failure this discipline exists to end.

Also in the folder

Sources — reviewskills/review/SOURCES.md

Sources — review

Provenance for the review skill: what was adapted, from where, and under which license — the full attribution text lives in NOTICE at the repository root, and this note never ships into a company.

  • obra/superpowers (MIT) — the giving/receiving pair (requesting-code-review, receiving-code-review): reproduce before fixing, agree by fixing, disagree with evidence, no reflexive agreement.
  • gstack (MIT) — the review skill's pre-landing discipline: findings before verdict, evidence cited from the working tree.

Ours, not adapted: the Iron Law's wording, the evidence-bundle reading order (Intent / Checks / Revisions / Blast radius), the propose-gate landing, and the refusal. Adapted per the 2026-08-22 authorized-skills direction; the Anthropic proprietary document skills and the original GSD chain were not consulted.