What tools are you using? What settings? What process? What's your code review like?

I think this varies a lot. I find with a c++ project I'm working on that the LLM needs a lot of guardrails and guidance, and still gets a lot wrong. But with a vite/js project it often one shots complex and intricate changes in large codebases.