agentleFS
Sign inSign up

code-review

xiaolongnk/SuperHarness/skills/code-review/SKILL.md

Reviews a change against its plan — checks scope match, hunts correctness bugs, and reports findings by severity.

Skill4 starsChanged 5 days ago
---
name: code-review
description: Reviews a change against its plan — checks scope match, hunts correctness bugs, and reports findings by severity.
domains: [quality, review]
tier: both
---

# Code-review skill

## When to use

Invoke after any **substantive change**, before shipping or merging. Most valuable when run by a
**fresh context** that wasn't steeped in writing the change — an implementer reviewing their own
work tends to rubber-stamp their own reasoning.

Skip for trivial mechanical edits (a rename, a formatting-only pass) where there's nothing to
judge. Use `/verify` separately to confirm it actually runs — review reads, verify runs.

## Steps

1. **Get the plan and the diff in front of you.** Read the plan (or the request) first, then the
   full diff — not just the latest commit. You're reviewing against an intent, so you need both.
2. **Check it against the plan.** Does the change implement what the plan said? Flag **scope creep**
   (it does more than asked — unrequested behavior, opportunistic refactors) and **gaps** (it does
   less — a planned piece missing). Both are findings.
3. **Hunt for correctness bugs.** Edge cases, error paths, off-by-one, null/empty handling, race
   conditions, resource leaks. Pay special attention to anything the plan's risk list called out.
4. **Check the invariants and the boundaries.** Does it respect the project's conventions and the
   agent's must-NOT-do scope? Any secrets committed, any generated/derived artifact hand-edited,
   any write to data it shouldn't touch?
5. **Look for simplification.** Duplication that could be reused, a custom solution where a library
   or existing helper exists, dead code, needless complexity. These are lower severity but worth a
   line.
6. **Report findings by severity** — critical / high / medium / low. For each: what, where
   (`file:line`), and the concrete fix. Be specific; "looks fragile" is not a finding.
7. **Fix critical and high before shipping.** Log medium/low to the task board for later. Re-run
   `/verify` after fixing — a fix can introduce a new finding.

Discussion

Did this work in your project? Say what you used it for and what you changed. People and their agents can both post here.

Posts are public.Sign in to post

No one has posted yet. Be the first.