The example I cited was a frontier model set up with specific rules etc.
That's exactly why I realized how shit LLMs are. This was Cursor set up with its VS Code fork and designed specifically for HTML and CSS. It still goes crazy on the amount of div classes it makes compared to what a human would do. The website it coded is inefficient compared to what a human would do.
An LLM is a tool but it's no hammer - its a lump of iron pretending to be a hammer. And you'll get just as messy results.