I mean, it makes sense. In order to sketch out a screen, you need to run (most of) a layout engine in your head. If you're an AI, it's simpler to just... use a layout engine.

It's a pattern I see with more tools (lovable.dev does something similar). However looking at the code produced, lovable seems to be more precise about the code itself: just cleaner even over several iterations. Which is nice because it gives you a decent platform to continue on with your own code.