Somebody here said programming is about designing and reusing libraries rather than coding from scratch like LLMs. But that choice has always been a tradeoff between abstraction and bloat vs performance and debuggability. Writing 50 lines of intentional code is often preferable to adapting a 50000 line library. The trick to using LLMs is to not ask for too much. Claude 3.5 could reliably code tasks that would have taken 5-10 min by hand. Claude 3.7 is noticeably better and can handle 10-30 min tasks without error. If you push those limits you will lose instead of save time.