My experience at the minute is that AI is good at writing functionally correct code, but bad at writing clean code. This is why it quickly gets out of hand if you don't manually refine the output and rely solely on the vibes - if you keep building without refining it quickly becomes non-functional spaghetti. I've found that it's quite good at "colouring in" - you sketch out or partially sketch out what you want just using descriptive function names and then the AI can fill in the blanks. Once a boilerplate is established and you need to repeat it multiple times that's where it really shines and saves a lot of time.