Reviewing Code With an Assistant

Where automated review helps, where it misleads, and how to ask for specific feedback.

4 min read

Ask for a specific kind of review

A general "review this" produces generic advice about naming and comments. Ask instead for a review focused on one dimension: correctness, error handling, security, or performance.

Running three focused reviews produces far more usable feedback than one broad one.

What it catches reliably

Unhandled error paths, missing null and boundary checks, resource leaks, inconsistent return types, and obvious injection risks from string-built queries.

Copy-paste divergence, where two similar blocks differ in a way that looks unintentional.

What it misses

Anything requiring knowledge of the wider system: whether this function is called concurrently, whether the data is already validated upstream, whether the business rule is right.

Architecture. It will comment on the code you pasted, not on whether that code should exist.

Use it before the human review

Run the automated pass first so that your colleagues spend their time on design and correctness rather than on missing error handling.

Do not paste the feedback into the pull request unedited. Decide which points are real in your context and act on those.

Frequently asked questions

Tool for this

CodeGPTGenerate, explain, review and debug code.

Open the CodeGPT page

More in Developer Tools & Code