Has anyone had Claude review their code?
I'm working on a side project solo, and since there's no one to review it, I've just been tossing PRs at GPT, but I'm not sure if I should trust it. What model do you all use to run reviews?
5 answers
Agreed. Same here—mine's a side project, so I'm opening PRs and reviewing them all by myself, so I just toss them to Claude lol
Honestly, Claude is better than GPT at code review. Especially when you paste in a whole long file, it misses the earlier context far less often. But the catch is that both are good at sounding plausible, so I don’t apply review results as-is—I only use them as candidates. Of the issues raised, only about 30–40% are actually valid? The rest often mention nonexistent methods or bring up edge cases that were already handled. So I set up my workflow like this: 1) Don’t just give the PR diff—paste in the related files and test code too. 2) Split into 'bug/security/performance/readability' categories and ask about each separately. 3) For every point raised, push back with 'Give me an example input where this actually causes a problem' to verify it. This greatly increases the rate at which nonsense gets filtered out. Bottom line: trust, but verify.
Well, that's a bit... Before we even talk about review quality, you should probably think about copy-pasting code you wouldn't show anyone else as-is. Especially if it's company code. And the claim that 'AI reviews are better' is over the top—reviewing isn't just pointing out style; it's done by someone who knows the architecture or requirements context, so how is a model supposed to replace that? It's good for convention checks or catching minor bugs, but I don't agree if you're saying it replaces reviewers.
I’ve set up a dedicated system prompt just for reviews, something like: 'You are a strict senior reviewer. Minimize praise, back up critiques with line numbers, and if you’re not sure, explicitly note that it’s a guess.' Since I added this, it no longer says things like 'Great code!' and is much more useful. I’m using the latest version of Claude, and sometimes I cross-run it with open-source models and only pick out the critiques where they disagree.
Got a source? Like a benchmark comparing review accuracy. Let's go by data, not vibes.