With code, I tend to think of it as polishing. The time spent on polishing might be near zero if something isn't very important, and not bothering anyone. E.g. a Python script that generates a plot from a log file that helps me debug something. I might not even look at it. But especially when something needs review by another human, the polishing phase easily takes 90% of the time— and probably even more of the tokens. The LLM is called "generative" for a reason; you need to keep chipping away at the stuff it generated. There's also a constant hunt for weirdness hiding in corners, the you’re-absolutely-right waiting to happen. The key here is to relive your why-phase on steroids on every single line in the diff. > If I had more time, I would have written a shorter pull request. It does raise the question whether it would have been more efficient to just write the initial version by hand. For me, the jury is still out on this when it comes to high-care situations. But generating code myself also involves quickly producing something that roughly works and gradually polishing it. With the LLM, I just spend more time wearing my grumpy reviewer hat and less time wearing my "I don't feel like writing a test for this" hat.