In my experience LLMs are surprisingly bad at CSS beyond a very basic level. They work fine if you need to change the color of a button, but when it comes to actual styling work, even intermediate stuff like position:absolute or CSS grid, Copilot or even CC default to outputting correct-looking gibberish really quickly.

That's telling about CSS design. Folks here on HN are talking about how they purposely ask LLMs about APIs that don't exist, and they hallucinate with a better and more intuitive design that they would come up with on their own.

I don't know the best solution for the problem, but CSS is a very convoluted one.

It could also be that there is a dirth of high quality CSS training data in comparison to JavaScript et al.

I wouldn't be surprised if the negative developer sentiment toward CSS is reflected in training datasets.

My guess is it’s because CSS is so dependent on context. Especially layout styles only make sense for a specific structure of HTML elements, which might be stored in an entirely different file and directory.