Comment by viccis

Seems ironic to say about this, but it's not: Don't anthropomorphize LLMs; they are just really really good story telling machines.

Telling it to "blur your eyes" doesn't mean it has eyes to blur. It means that it is telling a story that involves itself (a front end engineer) looking at a website and blurring his eyes. What it will continue to complete this with is the tokens with the highest probability of following such a statement. That is to say, it understands what it means to "blur your eyes".

I'd also add that it can take screenshots with headless browsers, and it can blur them with image manipulation tools, and finally it can examine those with its multimodal capabilities.

>Does it work? Maybe. I'm just really skeptical after reading through that repo that any of this leads to actually better user interfaces.

Anecdotally? Yes.

>I'm pretty sure I'd have better luck just telling the LLM explicitly what I want, because experience in UI/UX is still better than what an LLM would slop out on its own.

If you are competent enough to do professional level UI/UX design spec, then yes, this is not for you. You'd be better off using tailored tooling at that point. This is for people who haven't spent years generating UX spec documentation.

If you don't generate exact specs, these LLMs will, in my experience, generate stuff that looks like generic forms that people made with PHP tutorials in 2005.