How do you balance initial building speed with long-term code maintainability when using AI?

The speed of going from zero to a working demo with AI is addictive. But the real headache starts when you have to refactor that code after two weeks of user feedback.

I realized that without strict naming rules good docs and clear testing modifying early AI code takes twice as long as writing it manually.

Now I force myself to slow down right after the initial build to clean up the logic before touching any new features.

Do you clean up your AI-generated code right after the prototype works, or do you wait until things start breaking?