I don’t think the real problem with AI coding tools is that they have bad taste.
You hear it everywhere. Designers, developers, people on Twitter. “These tools have no taste.” And I get it. You ask Cursor or Codex to build a settings page and you get back something that looks like a 2019 Bootstrap template. Drop shadows everywhere. No rhythm. No restraint.
But I think that criticism is pointing at the wrong thing.
Taste is personal
When someone says “good design,” what they actually mean is “design that I personally think is good.” And that changes everything. Because taste isn’t universal. It’s shaped by the apps you use every day, the interfaces you stop to admire, the little details that make you think oh, that’s nice.
My taste leans toward the clean restraint of iOS and Robinhood. Flat, modern, deliberate. I love the warmth, simplicity, and playfulness of the Family app. I love the way Linear’s landing page feels both dense and effortless. Your taste might be totally different. Maybe you like the information-heavy density of Bloomberg, or the playful maximalism of a Teenage Engineering product page. None of these are wrong. They’re just different.
And that’s exactly the problem.
The issue isn’t bad taste. It’s unknown taste.
These tools don’t know the colors you like. They don’t know how much rounding feels right to you, or what kind of spacing and density and polish looks good in your eyes. So even when the output is technically fine, it still feels off. It doesn’t feel like yours.
Stuck on day one
Think about working with a human designer. Day one, their work is generic. By week three, they’re anticipating your preferences. They’ve seen what you approve, what you reject, what makes you light up in a review. They learn your taste through exposure.
Cursor, Claude, Codex: they’re all permanently stuck on day one. Every prompt is a cold start. No memory of what you find beautiful. No mechanism to learn it.
You can write a paragraph in your system prompt that says “make it minimal and modern.” But that’s like telling a chef “make it taste good.” It’s technically a direction. It’s just not a useful one.
The models are smart enough. They just haven’t seen enough examples of what you think looks good.
So they default to something generic. And generic is the opposite of taste.
Show, don’t tell
That’s the idea behind Taste. Instead of trying to describe your preferences in words, you show your AI what you actually like. Screenshot interfaces that resonate with you. The app extracts the design language underneath: the colors, the type scale, the spacing, the border radius, the overall feel.
Over time you build a library that represents your taste as a structured profile. Not a mood board for humans. A taste profile for AI. When Cursor or Codex generates UI, it doesn’t have to guess anymore. It has a real reference for what you consider good.
The outputs stop looking generic because they’re no longer starting from zero context about you.
It’s a personalization problem
The conversation about AI and design keeps getting framed as a binary. Either AI is good at design or it’s not. But that misses the point. An AI that produces interfaces I love might produce interfaces you hate. Design quality is relative to the person judging it.
We don’t need smarter models (though that helps). We don’t need more design training data. We need a way to inject your individual taste into the generation process. A layer between you and the AI that turns subjective preference into structured context.
That’s what Taste is. The missing layer.
The bet
I think as these tools get more capable, the people who get the best design output won’t be the ones writing the best prompts. They’ll be the ones who gave their AI the most context about what “good” actually means to them.
That’s why I built Taste.
