Review one validated phase for behavior-preserving structural regressions introduced or materially worsened by the change. Use after phase validation and before final delivery.
Installs into .claude/skills of the current project.
Are you the author of Phase Review?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/yarlson-phase-review)
---
name: phase-review
description: Review one validated phase for behavior-preserving structural regressions introduced or materially worsened by the change. Use after phase validation and before final delivery.
---
# Phase Structural Review
Own a narrow maintainability pass after correctness has been established.
## Workflow
1. Read the selected contract, applicable repository rules, final diff, and changed units.
2. Identify only what the selected change introduced or materially worsened: code that deletion, an existing helper, the standard library, or a platform feature would replace; an abstraction, wrapper, flag, or configuration with one consumer; duplication; unclear ownership; loose types; or logic with a shorter readable form.
3. Use `crap-index-assess` when the repository's configured CRAP check covers changed methods or the phase was intended to reduce their CRAP risk. Treat score changes as supporting evidence, not a structural verdict.
4. Fix a finding only when the parent implementation request authorizes scoped edits and the correction is clearly behavior-preserving.
5. Report a worthwhile correction when it requires non-trivial restructuring or broader authority; do not expand scope during the review.
6. Rerun affected checks after every edit.
Report each finding on one line: location, what to cut or simplify, and what replaces it. When nothing qualifies, say so in one line and stop.
Correctness belongs to `phase-validate`; general read-only defect review to `code-review`; test sufficiency to `test-gap-review`; documentation accuracy to `docs-drift-review`; and specialized risks to their owning reviewers.
Do not add behavior, pursue unrelated cleanup, use subjective file-size thresholds, or continue until no imaginable simplification remains.
Finish when structural regressions introduced by the change are fixed or reported with their scope and risk.