This is where prompting and context is key - you need to keep the scope of the review limited and well-defined. And ideally, you want to validate the review with another LLM before passing it to the dev.

Still won't be perfect, but you'll definitely get to a point where it's a net positive overall - especially with frontier models.