I use Claude every day. When I give it a program that does something straightforward in one file and it writes it from scratch, it does great. When I have it fix issues or add functionality to small to medium sized apps that have a mostly simple design, it does great. When I point it at codebases that are 20 years old and have a lot of indirection in their design due to years of hard lessons learned and a lot (like a LOT) of cases covered, it really struggles (just read my profile to know what codebase this is). I mostly try to get it to write changelog messages, docs and tests, where it works, but I have to really wrestle with it. I can't imagine doing anything on "vibes" and it all seems quite ridiculous if you are working on hardcore library oriented software with tens of thousands of users.

If we're going to say, who cares, with LLMs we'll never need 20 year old codebases we'll just keep writing new stuff, OK you do you.

one thing that i find works really well is to ask it to research things in the codebase and write a plan first. codex with gpt-5 is exceedingly good at doing this. then ask it to write a plan for what it would do with that information, i.e., i want you to research codebase for <goal>. then write a plan for how you would achieve <goal> given what you have learned.