How are you reviewing AI-generated PRs right now?

Genuine question, not a setup.

Since I started shipping with Cursor and Claude Code, my PRs went from 20-line diffs to 800-line sprawls, and my review process hasn't kept up.

I've talked to engineers using everything from "read it carefully and pray" to running the full test suite on every PR to skipping review entirely on AI-authored changes.

What's your current ritual? Where does it break down? What would you want a pre-merge tool to actually tell you before you hit merge?

Building MergeProtect around the answers, happy to share what I'm hearing from other teams too.