alt Hacker NewsAs PRs get beefier I find myself using the LLM to ask pointed questions and do general surveys of the diff to flag architectural violations, smells, etc. before I read through the entire thing by hand. If you put all of your architectural and business decisions in the repo via spec-driven tools or otherwise then this becomes quite useful. It also helps if you have a vision in your head of what the system is supposed to look like structurally so you can ask the LLM specifically how the PR is handling things that you'd be on the lookout for anyway.
I've been thinking about a workflow where PRs just get automerged into a staging environment of some kind, and then tooling helps ensure that all code, ALL the code in the entire project, is regularly reviewed together.
When lines change, they are put up for review. When lines haven't been reviewed for awhile, they are put up for review. Lines are ranked based upon the complexity of the context they are in, and complex areas of the code are reviewed more often.
Basically, stop focusing on individual PRs and instead focus on continuous review of the entire system with tooling that helps identify those parts of the code that are most in need of review.
Reviewing PRs is wasteful because the optimal way to review a system does not match the way a system naturally grows through individual commits.