1. Split review priority by blast radius, not diff size. A 400-line PR touching one isolated module is often faster to review than a 40-line PR touching shared components or a migration.
- Push the stuff that doesn't need judgment (naming, obvious edge cases, style) out of human review entirely, so by the time you look at it, you're only spending attention on things that actually need a human.
- Try to summarize if what is done is what was intended
- Used multiple models to surface common findings instead of starting from scratch
- We are also getting dried, that is why we try to have pre-agreggated info before looking at a PR.
btw I build dromeas.ai to solve this pain so quite biased.
Comments
Very true. A few things that worked for us:
1. Split review priority by blast radius, not diff size. A 400-line PR touching one isolated module is often faster to review than a 40-line PR touching shared components or a migration. - Push the stuff that doesn't need judgment (naming, obvious edge cases, style) out of human review entirely, so by the time you look at it, you're only spending attention on things that actually need a human. - Try to summarize if what is done is what was intended - Used multiple models to surface common findings instead of starting from scratch - We are also getting dried, that is why we try to have pre-agreggated info before looking at a PR.
btw I build dromeas.ai to solve this pain so quite biased.