You really have to a) use Opus and b) use the frontend-design skill for decent results.

https://github.com/anthropics/claude-code/blob/main/plugins/...

Obviously this is a personal preference, but the multiple layers of beveled grey on the Qt UI is not something I like, as it forces a lot of grouping on the eye where it doesn't serve any purpose.

I would go with the original, Apple or the Win11 one. Material would be good, what's with the lavender shades?

I always try to reduce the palette: say two background shades max, no drop shadows, only as many foreground colors as needed and if it seems to bland, add more bells and whistles.

This begs for a modern version of https://csszengarden.com/, where the CSS is generated by different LLMs and prompts.

Qt is heavily represented in training data. Qt has existed for decades and the model has likely seen Qt tutorials, screenshots, source code, discussions, etc. As a result, "Qt application" is a highly coherent concept in the latent space. "Qt app" is almost like a named distribution.

This is mostly the fault of the model, a lot of them have been trained to generate HTML in a specific style. Claude's is pretty distinct for example, I think the new DeepSeek copies it. Some of them can generate more humanlike HTML like Kimi K2 IIRC, which I feel is the model with the least amount of post-training in general.

It's necessary if you don't want it to generate HTML with images and other assets you don't have of course, that's why they use emojis or meticulously handcrafted SVGs, or WebAudio synthesized sound which pretty much no humans did before.

I donno. They all look ugly.

When making small tools for myself, I just tell it to use Svelte and then wrap it up using Tauri - no graphical cues whatsoever. And they usually comes out pretty good by my taste.

If you're just having fun with it, there are a whole bunch of other things that produce interesting options, like asking it to theme according to a movie (think Clockwork Orange, Backrooms, anything with a strong aesthetic), or throw screenshots and photos at it and use it as a "design system" (magazine/print layouts can work well with this on stronger models).

Only a small anecdote, but I'm 2 projects into telling Claude to "make it look like Google podcasts" and getting satisfactory results. Still smells like llm in parts, but overall it is not screaming low effort.

I kinda liked the Original, HIG and Windows 11 versions the most. When I think "AI slop" (in terms of web design), I think dark theme, rich purples and vibrant hues, huge headings, etc. The SaaS one kind of has that with the purples and vibrant hues; it easily looks the "sloppiest" to my eye.

Can't for personal preferences I suppose.

I think this says more about "modern" UI than it does about AI slop. The awfulness of all this comes mostly from the fact that widgets no longer have consistent shape, theme or interaction behaviour ever since desktop paridigms and original Xerox/Parc research were abandoned in favour of web slop. So yeah, this is much more Web Slop than AI Slop. AI is just amplifying it.