Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →stevesolun/code-review
Review a Git diff against repository standards and its originating requirements.
Passed 14 of 19 test cases with the skill, and 11 of 19 without it.
and 3 more in the full report
5 fixed13 passed either way5 still failing1 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
A PR replaces a thread-safe singleton cache in `src/cache.py` with a global mutable dictionary that causes race conditions under load. Style linters pass without errors. Which axis catches this issue?
| Checked | Without skill | With skill |
|---|---|---|
| Identifies the Standards axis for catching material design regressions. | Missed | Passed |
The instructions your model reads when the skill loads.
Honor a fixed point supplied by the user. Otherwise infer a sensible comparison from the repository state, such as the default-branch merge-base. Ask only when different choices would materially change the review.
Verify that the ref resolves and inspect the diff and relevant commit messages. Report an empty or unavailable comparison instead of manufacturing findings.
Use the repository's actual standards and the originating issue, PRD, or spec when available. Follow repository conventions for fetching tracker context. If requirements cannot be found, review the standards axis and state that spec conformance was not verified.
Use deterministic checks for properties tooling can decide. For qualitative review, load review heuristics only when the diff warrants deeper design analysis.
The axes are independent and may be evaluated concurrently when the diff is large enough to benefit. Keep a small review local, and avoid parallel work when coordination would cost more than it saves.
Lead with actionable findings, citing the affected file and evidence. Separate documented violations from judgment calls and keep Standards and Spec findings distinct so one axis cannot mask the other. State unavailable evidence and report when no material findings remain.
test cases passed, vs 11 / 19 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.