Is a working AI prototype actually just 10% of the real engineering effort?

Everyone talks about how fast AI lets you spin up a MVP but nobody warns you about iteration two.

Building from scratch takes a few hours going back weeks later to edit old code gets confusing fast. I've stopped treating the first generated version as a finished product and started looking at it as just a rough design draft.

At what point do you guys stop prompting for new features and start manually locking down the codebase structure?