My team is small — three people — and most recent PRs were written with help from Copilot or Claude Code. Writing code was never the slow part; the real problem is that nobody has enough time to read every PR carefully before merging. I wired CodeRabbit into a Next.js repo already running in production to see what it would actually catch.

This isn't a rehash of what CodeRabbit's homepage already says. It's a log of real usage, based on roughly 40 pull requests that went through it.

The first thing that stood out was speed: the bot leaves its first comment almost the moment a PR opens, and the change summary is remarkably accurate — not just a list of files, but a description of actual intent, like "this PR adds an Invitation table and a unique constraint on email + organization."

The most useful part in practice was the small fix suggestions — wrong data types, missing null checks, mismatched HTTP status codes. These are exactly the kind of issues that get skipped during a tired end-of-day review, and CodeRabbit caught them consistently.

On architectural calls — say, whether to split a service out of the monolith — its feedback stayed fairly generic, along the lines of "consider scalability" without engaging with the specific context of the system. This is still a job for a senior engineer who actually understands the product.

Also, in the first week the bot was a bit chatty — leaving comments on changes that didn't really warrant them. After I pushed back a few times on unnecessary warnings, the noise dropped noticeably, matching the "learning" behavior the team describes.