code-review

name	code-review
description	How to review a slice of a code repository with high precision — confidence-scored findings, project-guideline compliance, and Karpathy behavioral guidelines (simplicity, surgical scope, surfaced assumptions). Use when reviewing or auditing source files.
license	MIT
tools	["Read","Grep","Glob"]

Code Review

Synthesized for loomcycle exp7 from Claude Code's code-reviewer agent + code-review command + the karpathy-guidelines skill. Methodology for reviewing a slice of a repository with high precision and minimal false positives.

What to look for

• Project-guideline compliance. Honor the repo's own rules (e.g. CLAUDE.md): import patterns, error handling, logging, naming, platform compatibility, testing conventions.
• Bugs that bite in practice. Logic errors, nil/undefined handling, race conditions, resource leaks, unchecked errors, security issues (injection, secret handling, authz gaps), performance.
• Quality. Significant duplication, missing critical error handling, inadequate test coverage.

Confidence scoring (the precision lever)

Rate each candidate issue 0–100:

• 0 false positive / pre-existing; 25 maybe; 50 real but minor/nit; 75 verified, will hit in practice / guideline-backed; 100 certain, frequent.
• Only report issues with confidence ≥ 80. Quality over quantity — a short list of true issues beats a long list of nits.

Karpathy behavioral guidelines (apply to the review itself)

• Surface assumptions. If a finding depends on an assumption (caller behavior, invariant), say so. If unsure, lower the confidence rather than over-claim.
• Simplicity. Flag overcomplication: speculative abstractions, unrequested configurability, error handling for impossible cases, 200 lines that could be 50.
• Surgical scope. Judge whether changes touch only what they must; flag drive-by edits/refactors and orphaned (newly-unused) imports/vars. Don't demand rewrites of code that isn't broken.
• Verifiable. Prefer findings a developer can act on + verify (a failing case, a concrete fix).

Output per issue

file:line · severity (Critical | Important) · confidence · one-line problem · concrete fix. Group by severity. If no ≥80 issues in the slice, say so with a one-line "meets standards" summary.