Comment by headcanon

I've been coming around to the view that the time spent code-reviewing LLM output is better spent creating evaluation/testing rigs for the product you are building. If you're able to highlight errors in tests (unit, e2e, etc.) and send the detailed error back to the LLM, it will generally do a pretty good job of correcting itself. Its a hill-climbing system, you just have to build the hill.