| name | strict-code-review |
| description | Reviews code for maintainability, structure, abstractions, large files, and cleanup opportunities. Use when the user asks for strict, harsh, deep, or ambitious code review. |
Strict Code Review
Use this skill for a strict review focused on implementation quality, maintainability, abstraction quality, and codebase health.
Default posture: findings first and read-only. Edit files or mutate PR state only when the user explicitly asks for that follow-up.
Default scope is the current PR diff or current branch diff, including related uncommitted local changes. If the user gives an explicit scope, use that instead: a PR URL, pasted diff, file, package, module, component, plan, or current conversation direction.
For stacked branches, do not default to main...HEAD when that would include parent-branch work. Resolve the PR base/head, dependency branch, or explicit commit range first. Prefer the PR's actual changed commits, HEAD^..HEAD, or the user-provided range when the task is to review only the top stack item.
Core Standard
Rethink how to structure or implement the change to meaningfully improve code quality without changing behavior.
Look for "code judo" moves: restructurings that preserve behavior while making the implementation dramatically simpler, smaller, more direct, and more inevitable in hindsight.
Do not stop at local cleanup when a clearer framing could delete whole branches, helpers, modes, conditionals, wrappers, layers, or concepts.
Workflow
- Identify the review scope.
- Inspect the explicit scope, PR diff, current branch diff, or current direction. If the branch is stacked, state the base/range used and why it excludes unrelated parent work.
- Include related uncommitted local files. Exclude only files that are clearly unrelated, generated noise, or explicitly scoped out.
- Read
.agents/CONTEXT.md, .agents/CONTEXT-MAP.md, .agents/contexts/*/CONTEXT.md, or .agents/adr/ when domain language, ownership, module seams, or ADR-backed decisions affect the review.
- Classify findings by structural severity, not cosmetic preference.
- Prefer high-conviction findings over a long list of nits.
- For every finding, give the concrete file path, code area, why it hurts maintainability, and the smallest credible remedy.
Review Lenses
- Framing: find code-judo moves that delete concepts, branches, helpers, modes, or special cases instead of relocating complexity.
- Architecture and ownership: check coupling, module seams, shared-path feature leakage, and whether the logic lives in its canonical file, package, or layer.
- Control flow and state: expose incidental sequencing, repeated conditionals, nullable modes, partial updates, and hidden invariants; prefer a direct state model or explicit dispatcher.
- Abstractions: remove thin wrappers and pass-through helpers; introduce a focused helper or owned abstraction only when it removes repeated reasoning or owns real behavior.
- Types and duplication: flag unnecessary casts,
any, unknown, optionality, ad-hoc shapes, and bespoke copies of canonical helpers.
- Cohesion and scale: check whether the change makes a file or module harder to scan, especially when it pushes a file across 1000 lines; split only along a real responsibility boundary.
- Orchestration: separate business logic from coordination, run independent work in parallel when clearer, and keep related state updates atomic.
Skill Boundaries
- Use
simplify for the smallest useful scope, narrower API surface, and accidental-complexity reduction. Use this skill for structural code-quality findings that may block a PR or justify a larger rewrite.
- Use
validate-direction before hardening a rewrite that changes architecture, public API, migration strategy, an ADR, or a durable implementation direction.
- Use
handoff when the useful next move is clear but too large for the current session.
- Create durable project documentation only when the user explicitly requests capture.
Blocker Bar
Correct behavior alone does not clear the review. Treat a high-conviction finding as a presumptive blocker when a credible scoped remedy exists and the change:
- preserves incidental complexity that a plausible code-judo move would delete;
- pushes a file from below 1000 lines to above 1000 without a strong structural reason;
- tangles shared flow with ad-hoc branching, scattered feature checks, or logic in the wrong layer;
- adds unnecessary wrappers, casts, optionality, nullable modes, or duplicate helpers that obscure the design.
Put structural improvements below this bar in Strong improvements, not Blockers.
Output
Strict review scope:
- ...
Blockers:
1. <finding title>. `<path>`: <code area>. <why this structurally hurts>. Remedy: <specific change>.
Strong improvements:
1. ...
Leave as-is:
- ...
Needs validation:
- ... <!-- only if a suggested rewrite needs user/product confirmation -->
Suggested next skill:
- `simplify` | `validate-direction` | `handoff` | none
