LLMs are bad at drawing diagrams. Or are they?

Yeah, Mermaid is a good example of the same idea. LLMs are much better at describing a diagram in a text format than actually drawing one.

The interesting part for me was separating the semantic structure from the layout completely.

Not gonna lie, I actually greatly dislike Mermaid.

Virtually every diagram it makes that's over 6 nodes becomes a spaghetti mess and you can barely influence layout yourself. The default of using curves, the lack of care to prevent lines from intersecting with labels and even nodes themselves, making a plurality of graph types and shapes before DB ERDs can point arrows from/to a property, docs that aren't versioned...

For something that's become nearly universal and standard, it just feels remarkably half-baked.