Tme to pull out the deep neural nets! Isn't the problem that you're describing a bit like generating, say, words given an image? Sure, there are several differences, but it's an interesting problem.

Easier than caption generation? - There's only one possible outcome for HTML/CSS --> website look - And we know what this outcome is (just render the thing) - not a lot of "words" in the vocabulary... there are only so many tags and styles we can use (at most a couple hundred), with most of the complexity coming from the combinations of these elements

But in some sense, it might be a harder problem, since the idea of web development is combining these "words" in a way that makes sense. So there might be a harder-to-learn grammar.