Comment by nickd2001

Am using Claude to attempt to do refactoring and find bugs. Sometimes its fantastic, finding issues instantly that'd take a lot of trawling or insider knowledge otherwise. Other times it gets obsessed about irrelevant things, makes suggestions that for some other obscure but non obvious reason don't work in practice. The generated code sometimes has excellent ideas I wouldn't have thought of. Other times it has places for bugs to lurk e:g if a directory isn't there, make it. Er, no thanks I want you to blow up if the dir isn't there because if it isn't , something else major went wrong. The trick is knowing when its going to be good and when hopeless and take you down a rabbit hole. Perhaps that is a meta skill on the part of the human developer. But I'm not optimistic about things improving, its the nature of how it is. The AI doesn't know personally the previous devs on the team, their programming tastes, the discussions they had at planning etc. Its got no context.