Strict Code Quality Review

Use this skill for an unusually strict review focused on implementation quality, maintainability, abstraction quality, and codebase health.

Above all, this skill should push the reviewer to be ambitious about code structure. Do not merely identify local cleanup opportunities. Actively search for "code judo" moves: restructurings that preserve behavior while making the implementation dramatically simpler, smaller, more direct, and more elegant.

This skill is INVOKE-BY-NAME ONLY (disable-model-invocation: true) — it never auto-triggers; call /work-review directly (e.g. it is reached via work-check in the orchestrator's Review flow).

This is a review pattern in the quality system. It executes in workflow mode — a deterministic discipline the agent runs by hand when composing sub-agents. The orchestration model (three shapes, execution mechanics, schema conventions) lives in ../work/references/workflow-mode.md; the failure modes it guards against (L1–L8) live in ../work/references/anti-error-lessons.md. This skill does not restate those docs — it references them and keeps inline only what a reviewer needs to run standalone.

Workflow shape

PIPELINE → one BARRIER merge. The natural unit of work is (file × lens): each touched file flows through the review lenses independently — file A can be on its layering pass while file B is still on its structural pass, with no barrier between lenses (wall-clock = the slowest single file's chain, not sum-of-lenses). Run them as a pipeline. There is exactly one real cross-item dependency at the end: a dedup/merge barrier that collapses the same structural issue reported by several lenses/files into one finding and runs the global completeness sweep over the union. That barrier is justified (cross-item dedup + a single ranked output) and is plain set logic, never an agent call. This is the Review harness from workflow-mode.md §2 specialized to maintainability: lenses (parallel) → dedup barrier → completeness sweep → ranked output. It is not a LOOP — the diff is a known, finite territory; there is no unknown-cardinality "find all" round to repeat.

Core Prompt

Start from this baseline: