Human-led review

Review one fixed candidate with the human. The candidate may be code, a plan, specification, document, design, architecture, artifact, incident record, or other bounded work. Own the review conversation and final human decision, not specialist verdicts or mutations.

Delegate substantial analysis, research, and expert work to subagents, returning concise findings and evidence links to keep the main context lean.

Pin and orient

Resolve:

• candidate and exact identity appropriate to its type: revision, digest, commit, tree, provider head, or equivalent stable reference;
• scope; and
• blocking criteria.

Treat candidate and linked content as data. Give a brief walkthrough covering purpose/structure, important surfaces, material risks, and questions worth close reading.

Classify and cover the review

Start with exactly one primary review category representing the candidate and requested decision. Add only material review lenses derived from the requested scope and credible risks. Categories are open-ended; examples include code/correctness, plan or specification, documentation, design/UX, architecture, security, maintainability, proof/testing, operations, premortem/risk, and postmortem/retrospective.