A code review skill that catches what you'd rationalize away

The most consistent way AI-written code goes wrong: the same model both writes it and reviews it. Same blind spots, same rationalizations, same "looks fine to me." The Code Review Helper is the cheapest fix for that.

When to use it

• You're about to merge an AI-generated diff and want a real second pass.
• You're self-reviewing your own PR before requesting human review.
• You're reading legacy code you're about to change and want to surface what's actually load-bearing.

What it does that "review this code" doesn't

The skill is opinionated about what counts as a real finding versus a nit:

1. Correctness first — off-by-one, null/undefined, wrong type assumption, resource leaks. The boring bugs.
2. Edge cases the change misses — empty input, concurrent calls, retry behavior. The things your tests don't cover because you wrote the tests right after writing the code.
3. Hidden contracts — API signature changes, schema migrations, idempotency assumptions. The "this is technically a breaking change" stuff.
4. Test coverage gaps — new branch with no test, assertions that would pass on return null.
5. Naming and intent last — it surfaces these but doesn't lead with them.

It explicitly skips style nits a linter handles. You want findings, not noise.

The trick that makes it work

Run it with a different model than wrote the code. Claude wrote your code? Ask GPT to review. Vice versa. Same model = same statistical center = same blind spots. Different model catches what the author rationalized away.

When not to use it

• You want approval, not feedback. The skill will find things.
• The diff is huge and you have no context. Skim first, give it the risky 200 lines.
• Style-only review. That's what prettier/eslint/gofmt are for.

Install it

Drop the skill file into your Claude Code or Cursor setup. The full version is on the Code Review Helper page.