Reviewing

How to review text, code, and pull requests with an agent.

Almost anything can be reviewed. You might already know that you can copy non-confidential documents into a chatbot and ask for a review; your agent can do a lot more too. You can also use local models for help (a little) with private material.

Reviewing text¶

You can ask for reviews of documentation, plans, and just about any text in a repository. You can ask for a general review, which will catch a variety of things like grammar, and you can also ask for reviews into specific things. Some examples:

Check all the code snippets and make sure they are valid.

Look for things that could use links.

Is anything outdated?

Scan the repo and collect all the public API, check the documentation and look for gaps. Is everything covered?

You are an NSF reviewer. What is the weakest part of this proposal?

Don’t mix tasks in a single context; if you have a specific ask, that will affect a general one. “Review this and also check for links” will likely focus nearly entirely on links, for example.

Reviewing PRs¶

AI has gotten really good at reviewing PRs / changesets. There are several options:

Review tool¶

Most harnesses have a built-in option, usually /review or similar. You might need to give it further information, like a PR number, branch, etc.

This is a great place to start. Try it out on a PR.

Adversarial review¶

Another thing you can try is an adversarial review. Try something like this:

You are an adversarial reviewer for the new feature in this branch. Do you see any problems? Anything that could be done or written more cleanly? Can you break it?

This is a real prompt too; it was used (more than once) while developing nox#1131, parallel session support.

Simplifying review¶

Claude Code has this as /simplify, but you can prompt it yourself. The key idea is to use four subagents^[1], each looking for a specific type of simplification. These are:

• Altitude

• Reuse

• Simplification

• Efficiency

They report back, and any findings that check out as correct are applied.

Reviewing AI with AI (Rubber Duck)¶

AI review is a powerful tool for AI-generated code. To be most effective, you should have two models of similar level from different model families. By using one as the implementer, and the other as the reviewer, you can get better results than a single model alone, much of the way to the next model class above. This is called Rubber Duck, and it is built into GitHub Copilot CLI; GitHub reports that Claude Sonnet 4.6 reviewed by GPT-5.4 closed 74.7% of the gap to Claude Opus 4.6 running alone.^[2]

If you try this with matching models, they will generally just praise the work. If the reviewer is a model class above the implementer, it does work (but then you could have just used the better model to do the implementation).

This is also why it’s really nice to let reviewers know what model(s) you used in PRs.

Human in the loop¶

AI writes code really fast, often code that looks really well thought out (and sometimes is, depending on lots of factors). Review is more of a bottleneck. AI is also good at helping review; you can review code much faster with an AI helping you.

An important, rather new bottleneck is actually design; is some new feature you’ve thought about for 10 minutes a good fit? Are you willing to support this API forever? Is the extra code worth it? This was less noticeable when a new feature took much longer and more effort.

One thing you can do is to use AI to simulate downstream. You can ask it to adopt your new feature in a library that uses yours and report on how well it works.