code-review_skill

Overview

This skill performs focused, PR-level code reviews that prioritize structural and cross-file issues over minor style nitpicks. It filters large diffs, preserves cross-file context, and surfaces high-impact problems like duplication, god functions, and weak typing. The goal is concise, actionable findings with numbered suggested actions and a required approval step before making changes.

How this skill works

It inspects the full PR diff (after filtering lock files and large noise) to detect cross-file patterns, duplicated logic, overly large functions, and type-deficient code. For large diffs it progressively excludes docs/tests and uses large-context LLM workflows when needed. The output is a structured review report with findings, prioritized actions, and an explicit ask-for-approval step before any automated edits.

When to use it

• When asked to review a PR, diff, or full change set ("review code", "review PR", "review diff").
• When you need cross-file pattern detection rather than file-by-file comments.
• When diffs are large and you need guidance on chunking or using large-context models.
• When you want high-impact, structural feedback (duplication, god functions, weak types).
• Before merging substantial changes or refactors that may introduce architectural issues.

Best practices

• Review at the PR level; focus on structural problems over minor style issues.
• Always filter out lock files first, then docs, then tests if needed to reduce noise.
• Flag weak types, hasattr/getattr patterns, duplicated logic, and functions >100 lines.
• Preserve large chunks (e.g., ~15k lines) to retain cross-file context; avoid tiny chunks.
• Provide a numbered list of suggested actions and request explicit user approval before applying fixes.

Example use cases

• Run a full PR review to find duplicated validation logic across services and propose a shared utility.
• Detect a 200-line god function and suggest extracting validation, transformation, and persistence responsibilities.
• Spot widespread use of raw dicts/Any and recommend introducing typed models or protocols.
• Use LLM-assisted large-context review for a 100k+ token diff to surface cross-file architectural smells.
• Produce a concise report with prioritized fixes and ask the author which items to implement.