Code review

Purpose

Perform structured code review guided by a formal error taxonomy. Instead of open-ended "review this code" analysis, this skill loads specific error categories with concrete review rules, examples, and severity guidance. The reviewer selectively focuses on chosen categories -- improving detection precision by narrowing the search space per the "mental attitude hypothesis" (focused reviewers detect 8x more issues in their focus area).

When to use

• Reviewing a pull request or branch diff
• Reviewing staged changes before commit
• Pre-merge quality gate on a feature branch
• Reviewing specific files for targeted concerns (e.g. "review this for security")
• When asked to "review", "check", "audit", or "critique" code changes

Inputs expected

• Diff source: one of branch (vs trunk), staged changes, specific file paths, or a provided diff
• Category selection: specific categories by name, a review profile, or "all" (defaults to quick if unspecified)
• Severity threshold: minimum severity to report (defaults to Minor -- excludes Nitpick unless requested)
• Project conventions: any project-specific style guides, lint configs, or architectural constraints when known

Review Profiles

Profiles are predefined category bundles. Use a profile name as shorthand, or specify individual categories by name.

Profile	Categories	Use When
quick	Logic & Correctness, Error Handling, Security, Performance	Fast review focusing on highest-signal defect categories
correctness	Logic & Correctness, Data Handling, Error Handling, Concurrency & Timing	Verifying functional behaviour is correct
security	Security, Concurrency & Timing, Resource Management, API & Interface	Security-focused audit
maintainability	Naming & Readability, Code Structure & Organisation, Style & Formatting, Documentation, Design & Architecture	Code quality and long-term health
testing	Testing, Logic & Correctness, Error Handling	Verifying test adequacy and quality
full	All 16 categories	Comprehensive review -- use for critical changes

Process

Step 1: Gather the diff

Determine the diff source from user input:

• Branch review: detect trunk (main or master), find merge-base, run git diff <merge-base>..HEAD
• Staged changes: run git diff --cached
• Specific files: run git diff on the named files, or read them directly
• Provided diff: use the diff text as given

Also gather: git diff --stat for the overview, and git log --oneline for commit context when reviewing a branch.

Step 2: Determine review scope

Resolve the active categories:

1. If the user named a profile, expand it to its category list
2. If the user named specific categories, use those
3. If no selection was given, default to the quick profile
4. Load the corresponding reference files from references/<category>.md

Read references/categories.md for the severity, confidence, and qualifier definitions that apply to all findings.

Step 3: First pass -- detect candidates

Read through the diff with the loaded category rules active. For each potential finding:

• Identify the location (file and line)
• Assign the category and subcategory
• Assign initial severity (Critical / Major / Minor / Nitpick)
• Assign initial confidence (High / Medium / Low)
• Tag the qualifier (Missing / Incorrect / Extraneous)
• Draft a brief description

Cast a wide net in this pass. Include anything that might be an issue.

Step 4: Second pass -- verify and filter

Re-examine each candidate finding against the full context:

• Promote findings where surrounding code confirms the issue
• Demote findings that are likely false positives (check the "Common False Positives" section in each category reference)
• Drop Low-confidence findings that lack supporting evidence after re-examination
• Merge duplicate findings that describe the same root cause in different locations
• Adjust severity based on the specific context (a null dereference in a hot path is Critical; in dead code it may be Minor)

This two-stage pipeline reduces false positives -- raw single-pass LLM output carries too many for production use.

Step 5: Present the report

Produce the structured output defined below. Group findings by severity (Critical first), then by category within each severity level. Include positive observations where genuine.

Output Format

## Code Review: [branch-name, file-name, or "staged changes"]

**Scope:** [what was reviewed -- file count, line count from diff stat]
**Categories:** [which categories were active]
**Profile:** [profile name if used, or "custom"]
**Severity threshold:** [minimum severity reported]

**Summary:** [count] findings ([count] critical, [count] major, [count] minor, [count] nitpick)

---

### Critical

#### [Finding title]
- **Location:** `file:line`
- **Category:** [category] > [subcategory]
- **Confidence:** [High | Medium | Low]
- **Qualifier:** [Missing | Incorrect | Extraneous]
- **Description:** [what the issue is and why it matters]
- **Suggestion:** [concrete fix or direction]

[...repeat for each critical finding...]

### Major

[...same format...]

### Minor

[...same format...]

### Nitpick

[...same format, only included if severity threshold permits...]

---

### Observations
- [Positive patterns worth noting]
- [Architectural notes that are not findings but merit discussion]
- [Areas that were reviewed and found clean]

If no findings are produced, state this explicitly with a brief note on what was checked.

Escalation Conditions

Stop and ask when:

• The diff is too large to review meaningfully in one pass (500+ changed files) -- suggest splitting by module or area
• The category selection does not match the nature of the changes (e.g. security profile on a CSS-only change)
• A finding has high severity but low confidence and could block a merge incorrectly
• Project conventions are needed to judge a finding but are not available
• The diff context is insufficient to determine correctness (e.g. deleted code whose callers are not visible)

Constraints

• Read-only analysis -- do not modify files, make commits, or apply fixes unless explicitly asked
• Honest assessment -- flag genuine risks; do not soften findings to be polite; do not inflate findings to appear thorough
• Proportional depth -- spend more analysis on complex or risky changes, less on trivial formatting
• No false authority -- use Medium or Low confidence when genuinely uncertain; never present speculation as fact
• Respect scope -- review only the diff provided; do not audit the entire codebase unless asked
• Category discipline -- only report findings within the active categories; do not silently expand scope
• Severity honesty -- most findings in any review are Minor or Nitpick; a review with many Critical findings should prompt self-verification

Reference Documentation

• references/categories.md -- master category index, scales, and finding template
• references/<category-name>.md -- individual category review rules with examples