The models that build websites can now make almost anything. A founder with no design training can ship a page that looks professional in an afternoon. You’d expect that to make websites better, and in one way it has.
But watch experienced designers review those pages and you notice them flinching at the same kind of thing. A line that follows you down the page as you scroll. A button that chases your cursor. A menu that fades away when you point at it. Each is a bit of motion that’s there to look clever, and each pulls your eye away from what the page is supposed to tell you.
They aren’t bugs, and the model didn’t sneak them in. Somebody asked for each one, or saw it on the screen and said yes. Before AI, each would have cost a developer days of work, and nobody would have spent days on them. Now they cost nothing, and that’s the problem.
That’s the diagnosis I kept hearing in five recent episodes of YC’s Design Review, where a guest and the host, Aaron Epstein, take startup websites apart live. Of the scrolling line, Epstein said, “You would only implement something like this because it was so easy to do.” The cost of building used to do two jobs nobody assigned to it. It filtered out ideas that weren’t worth making, and it told visitors that someone cared. AI has quietly taken both away.
Cost was the filter
The first job was filtering. When every effect took days to build, someone had to ask whether it was worth that time, and most weren’t. Building by hand also meant using the thing by hand, so you found its bugs. When Raphael Schaad, a general partner at YC, found a broken widget on one site, his reaction was that nobody who’d coded it by hand could have missed it.
The filter was blunt, though. It cut by price rather than value, so it killed good ideas too. Schaad praised card animations on one site that “would have been maybe too expensive to do” before AI, and Marvin Schwaibold, who leads a design studio at Shopify, said he now builds things he “couldn’t do” before. What’s gone isn’t the good ideas. It’s the decision, which used to make itself and now has to be made on purpose.
And the models won’t make that decision for you, because they lean toward more. “Models love having five, six, seven, eight different font sizes,” said Stephen Haney, who runs the AI design tool Paper. “Try to work with three.” The guests listed the rest of the habits: everything bold, purple gradients, little icons and labels in every corner. The Paper episode called the models insecure designers, afraid to leave any space empty.
Polish stopped meaning care
The second job was signaling. Katie Dill, Stripe’s head of design, explained why Stripe sweats its small animations. “If you see the care that goes here, then you’re right to assume that we also put that care behind the scenes in the way that we move money.” Polish used to be evidence of effort, and effort is hard to fake.
Now polish is free, so generic polish is evidence of nothing, or worse. Haney was looking at a site that felt vibe-coded, built by prompting an AI and keeping whatever came out. “Your mind goes to what else don’t they spend attention on,” he said. And generic polish is easy to spot because it all looks alike. Schaad’s guess is that a well-linked site’s style ends up in the models’ training data, “and then all of a sudden the next week all the startup websites look the same.”
So care now has to show in what a model can’t produce for you: what you chose to leave out, and things only your company could make. Epstein’s example was the counter on Stripe’s homepage showing the share of global GDP that runs through Stripe. Almost no other company could put that on a page, and that’s what makes it convincing.
The obvious fix for generic design is rules, and rules help. Haney says Paper’s whole secret is “basic rules of typography and contrast,” written into the model’s instructions by senior designers. But a rule raises everyone to the same floor, and once every model follows the same rules, following them won’t set you apart. As Epstein asked, if the models all learn the same taste, “does everything just look the same again?”
Taste is a reaction, not a description
Rules cover the part of taste you can put into words. The rest you can only recognize, and these episodes show how much of it there is. Schwaibold couldn’t get his coding agent to make a gallery behave the way he wanted, so he drew it as colored boxes in Figma. “I couldn’t articulate it via text or voice, but I could draw it out.” After 91 prompts on a carousel, he stopped describing and had the agent build sliders for each setting, like how strongly a card reacts to a fast swipe, so he could tune them by eye. “I know when it feels off, but the back and forth with chat just doesn’t work.”
I think that’s what taste mostly is. It isn’t a specification you hand over. It’s a reaction to something you can see. Prompting asks you to describe the result in advance, which is exactly what taste can’t do. So the workflows that work let you react instead. Ask for ten versions rather than one, build a slider for any setting you can’t describe, and change things directly rather than prompting again.
It’s also what the guests mean when they call the designer an editor now, and editing mostly means cutting. “It’s easy to throw everything,” Epstein said. “It’s much harder to figure out what needs to be edited out.” Cutting was always the hard part, because every feature has someone who wants it on the page. Now that AI makes adding free, cutting is most of what’s left.
Taste isn’t a specification you hand over. It’s a reaction to something you can see.
That leaves the question of what to do with the time AI saves. AI can get you to a seven out of ten fast, Dill said. “Do you just ship that and ship more sevens out of tens? Boy, I hope not.” Stripe pushed back its new homepage to get the transitions in one section right, and Paper still has people read every line of code its agents write. These aren’t slow companies. They’re spending the time AI saved on the part AI can’t do.
For anyone building this way, it comes down to a few habits. Decide how your product should look and behave before you prompt, for example its colors, its type and how much anything on the page should move, and give that to the model instead of taking what it makes and tweaking it. Start from the model’s defaults and you end up where everyone else did. As Epstein put it, “you have to be intentional about ending up in a different place.”
Before you keep any effect, ask what it does for the person reading the page, and if the only reason it’s there is that it was free, cut it. And before you ship, click through every page yourself, because building it no longer makes anyone do that. Neither takes long, and both are what the cost of building used to do for you.
Schwaibold said editing was always the designer’s real job. It used to take a lot of handcrafted work just to reach the point where you could edit. AI has taken that work away but not the editing, and it turns out the expense of making things was quietly doing some of the editing for us.
Notes
The five YC Design Review episodes this essay draws on, all hosted by Aaron Epstein:
- How To Stand Out In A World Of AI Designs, with Alex Danilowicz of Magic Patterns.
- How To Design In The Agent Era, with Stephen Haney of Paper.
- The New Full-Stack Designer, with Marvin Schwaibold of Shopify.
- How Stripe Built Their New Website, with Katie Dill of Stripe.
- Common Mistakes With Vibe Coded Websites, with Raphael Schaad.