Code Review

Structured code review with severity-labeled feedback. Discovers scope from git state, loads language-specific rules on demand.

When to Use

• Reviewing uncommitted or committed code changes
• Reviewing a GitHub PR
• MR review or quality check
• Establishing review standards

Project Conventions

Before reviewing, check if the project has a docs/CodeStyle.md file (or similar convention document). If present, load it and treat project-specific rules as overrides to the general rules below.

Review Process

Phase 1: Scope Discovery

Determine what to review. The user specifies the mode, or the skill asks.

Mode 1 — Uncommitted changes:

1. Run git status to collect changed, staged, and untracked files
2. Present a compact file list grouped by status
3. Ask the user to confirm or exclude files

Mode 2 — Committed changes:

1. Ask the user if they are on the correct branch
2. Ask for the diff target: master, a branch name, or a commit ID
3. Run git diff <target>...HEAD --name-status to get the file list
4. Present the list, ask the user to confirm or exclude files

Mode 3 — GitHub PR:

1. Ask the user for the PR number
2. Run gh pr diff <number> --name-only to get changed files
3. Present the list, ask the user to confirm or exclude files

After confirmation, record the confirmed file list and the diff mode. Each file's diff will use the matching method:

• Uncommitted → working tree diff (git diff / git diff --cached)
• Committed → git diff <target>...HEAD -- <file>
• GitHub PR → gh pr diff <number> -- <file>

Also check for docs/CodeStyle.md and load if present.

Note the change size. If the confirmed list exceeds ~400 changed lines, suggest splitting.

Phase 2: High-Level Review

1. Apply general code quality rules (see below)
2. Check file organization and architecture fit
3. Group confirmed files by language:
   • .vue → Vue group
   • .ts / .js → TypeScript group
   • .rs → Rust group
4. Determine the review order (one language at a time)

Phase 3: Detailed Review

Process one language group at a time to avoid filling context with all references at once.

For each language group:

1. Load that language's references (see Language References table below)
2. For each file in the group: a. Get the diff using the recorded mode b. Read the full file for context where needed c. Apply language-specific rules from the loaded references d. Label each finding with a severity (see below) e. Use question-based feedback for non-blocking items
3. When the group is done, move to the next language group

If security concerns are spotted in any file, load references/security.md at that point.

Phase 4: Summary & Report

1. Run through checklists (general + language-specific)
2. Write the review report to docs/reviews/<branch-name>.md (create folder if needed)
   • Get branch name: git branch --show-current
   • If uncommitted mode with no branch (detached HEAD), use the date: review-YYYY-MM-DD.md
   • Ask user to confirm or change the save path before writing
3. Report format — see Report Template below

Report Template

# Code Review: <branch-name>

**Date:** YYYY-MM-DD
**Reviewer:** AI (code-review skill)
**Mode:** Uncommitted / Committed (diff target: `<target>`) / GitHub PR #N
**Files reviewed:** N

## Summary

<1-3 sentence overall assessment. State the verdict: looks good / has issues to address / needs significant rework.>

### Stats

| Severity | Count |
|----------|-------|
| blocking | N |
| important | N |
| nit | N |
| suggestion | N |
| praise | N |

## Findings

### <file-path>

- **[severity]** <finding title>
  <description — what, why it matters, suggestion>

- **[severity]** <finding title>
  <description>

### <file-path>

- **[severity]** ...

## Good Patterns

- <what was done well and why it's worth keeping>

## Checklist

<paste the filled general + language-specific checklists with [x] for passed items>

Severity Labels

Use one label per finding. Every finding must have a label.

Label	Meaning	Action required
[blocking]	Bug, security issue, data corruption risk	Must fix before merge
[important]	Test gap, unclear naming, moderate perf issue	Should fix; discuss if disagree
[nit]	Style, minor naming, readability	Nice to have, not blocking
[suggestion]	Alternative approach worth considering	No action needed
[learning]	Educational — explains why something matters	No action needed
[praise]	Good work, reinforcement of good patterns	No action needed

Feedback Approach

Ask questions instead of making statements:

• "What happens if items is an empty array?" instead of "This will fail on empty arrays"
• "How should this behave if the API call fails?" instead of "You need error handling"

Use collaborative language:

• "Consider..." / "Have you thought about..." / "Would it make sense to..."
• Not: "You must..." / "This is wrong" / "Why didn't you..."

Be specific and actionable:

• Include what the problem is, why it matters, and a concrete suggestion
• Reference the relevant rule or pattern when applicable

Balance criticism with praise:

• Call out good patterns with [praise]
• Acknowledge thoughtful decisions

General Code Quality Rules

These apply to any language. Brief checks — detailed patterns live in language references.

• Single responsibility — each function/component does one thing
• Size limits — functions < 50 lines, files < 300 lines, function parameters < 4
• Naming — descriptive names, no single-letter variables (except loop indices), consistent casing
• Dead code — no commented-out code, no unused imports or variables
• Complexity — no nesting deeper than 3 levels, no long boolean chains (extract to named variables)
• Error handling — no swallowed errors, no empty catch blocks, no ignored promise rejections
• DRY — no copy-paste blocks longer than 5 lines (extract to shared function)
• Change scope — the change does one thing; < 400 changed lines preferred

General Checklist

• Each function/component has a single clear responsibility
• No function exceeds ~50 lines
• No file exceeds ~300 lines
• Names are descriptive and consistent
• No commented-out code or unused imports
• No nesting deeper than 3 levels
• All errors are handled (no empty catch, no swallowed rejections)
• No duplicated logic blocks > 5 lines
• Change is focused — does one thing

Language References

Load the matching references based on file extensions in the confirmed list.

File extension	Reference	Load
.vue	references/vue/	All files in folder
.ts, .js	references/typescript/	All files in folder
.rs	references/rust/	All files in folder
Any (security concern)	references/security.md	When security issues spotted

Each reference folder contains topic files with ❌/✅ patterns and a checklist.md for quick final-pass.