Frontend design skill and the boundaries of slop

After watching yet another video about the design capabilities of neural networks, I got triggered into finally writing down a thought I had been carrying around for a while.

We are very tired of how pages laid out by neural networks look. You can spot them immediately. That is why Anthropic's frontend design skill caused such a stir at some point: it makes neural-network design slightly less neural-networky. But that is only the first impression.

Time has passed. Now all our design looks like that. Literally every day I see some page or presentation where I can tell Claude built it with the design skill. I still very often see people recommending it to each other and so on. It spreads like a plague. Does it produce good layouts? Yes, excellent ones. Is it good design? No, it is generic nonsense that simply looks good.

My point is this: what is slop and what is not is decided only by relevance, mass adoption, and appropriateness. Ghibli pictures were an interesting idea at first too (the first two of them), and then, when there were basically no other pictures left on the internet for a moment, everyone got sick of them. This is exactly the same story. The out-of-the-box slop has simply been replaced with skill slop. And while the former changes from model to model, from one model version to another, this one is also a permanent thing.

The main part of the problem is again how much opinion the text inside the skill carries. And not a professional opinion from designers, but an opinion generated by the same neural network in a spherical-vacuum context. Brutally minimal, maximalist chaos, retro-futuristic, organic/natural, luxury/refined, playful/toy-like, editorial/magazine, brutalist/raw, art deco/geometric, soft/pastel, industrial/utilitarian — these themes are synonyms for modern slop.

To stay original, you need to make at least one extra step. Send the model to research which themes and details are applicable in which situations, and which of them are currently relevant or timeless classics. Then write that knowledge down instead of the retro-futurism-like nonsense. If you like a particular style and want to use it for some tools, you can lay out a unique design system in that style once in ten minutes and put it into your skill.

Oh, and nothing except Opus can produce pleasant design, do not even try. That is why, even though I moved to OpenAI for a hundred bucks, I still keep the minimal Claude subscription purely for occasional frontend work.