review-code

ultrathink

Process

Critique only — no fixes or implementation. Understand why the code exists before judging it.

1. Identify the review target

Prefer in order: explicit <target>, changes made in this conversation, working diff. If unclear, ask.

Then gather enough context (PR/issue text, plans, surrounding code) to avoid misreads. Don't turn every review into archaeology — if a missing detail would materially change judgment, ask briefly.

2. Match depth to risk

• Quick — small, local, low-risk; keep output brief
• Standard — normal feature or refactor
• Deep — risky, cross-cutting, security-sensitive, or surprising

3. Review flow first, then details

Understand how important parts work before commenting line by line. Focus on load-bearing logic, edge cases, scope drift, and intentional tradeoffs.

4. Report

For each finding: file path + line, why it matters, severity.

• Critical — correctness, safety, broken edge cases, scope violations. Requires evidence in the code — speculative failure modes belong in Questions.
• Suggestion — non-blocking improvements
• Question — missing context or unconfirmable concerns

If something looks odd but context supports it, acknowledge rather than flag.

5. Output format

## Code Review: [scope]

**Verdict: {Approve | Request Changes | Reject}**

### Inferred Intent
[Only when context supports it]

### Critical Issues (Must Fix)
- `path/to/file:line` — concise description. Why: [reason]

### Suggestions (Nice to Have)
- `path/to/file:line` — concise description

### Questions
- [Clarifications on intent or tradeoffs]

### Summary
[1-2 sentences]

Include only sections with content. Always include Verdict and Summary.

Boundaries

• Read surrounding code when the diff alone is misleading.
• Deliberate decisions (inline comments, user-stated tradeoffs) are not defects — raise as Question only with concrete evidence.
• Don't judge by personal preference when project fit explains the choice.
• Name the right approach ("needs parameterized queries") but don't write corrected code.

Target

$ARGUMENTS