At what point do you stop prompt-coding and start cleaning up the mess manually?

Built a quick working prototype over the weekend using mostly AI prompts and people actually loved the idea.

But the second i tried adding a new feature everything broke down.

The state management was all over the place and styles were messy. Ended up spending weeks rewriting the core data layer just to make it usable again.

AI coding speed is awesome for early testing but how do you know when it's time to put the prompts away and write clean code by hand?