I don't know, sometimes I feel like I've been misled, I come back to the code that was generated, I find so many bad choices, claude never subtracts, it keeps piling on new code. Some minor things get declared as unbreakable rules, regardless of context. It memorizes the tiniest exception as some big decision and all new code follows those new rules it made up. Good luck cleaning up all the nonsense to all who just vibe coded their apps. It has to be steered constantly, if you care about understanding the code that is. The LLM will happily use gibberish variable names, long functions that mix responsibilities, inconsistent code style and so on. Give it a prompt to review the code over and over again, ten times in a row, it will always find something, that one race condition that wasn't handled, or it will warn about something that is clearly not important, but it won't realize that it showed rendering code into the core logic component. It can fix it once it's pointed out, it's good for refactoring and doing sweeping changes across the codebase, but it's better to be on top of things, because it can lose its way behind your back.