Notes6 min read
AI closed the skill gap. It widened the taste gap.
Models favor the most common answer, so unsteered output lands on the average. You fix it by training your eye with side-by-side comparisons and writing down what you choose, not by prompting harder.
From the course · lesson 0.4
Ask an AI agent for a landing page and you'll get something competent. A clean sans-serif font, a soft gradient, three cards with icons, a button that says "Get started." Nothing is wrong with it. It's also generic: you've seen it a thousand times.
This happens for a reason, and the reason has little to do with your prompt. Once you see it, the fix gets much clearer.
Why AI output looks generic: it drifts to the average
A language model writes by predicting what usually comes next, based on patterns in what it was trained on. Anthropic's team calls the design side of this "distributional convergence." Safe design choices, the ones that work everywhere and offend no one, dominate the web. So when you give a model no direction, it picks from that safe middle.
Anthropic even named the usual suspects. On their list:
- Fonts: Inter, Roboto, Open Sans, Lato and default system fonts
- Color: purple gradients on white backgrounds
- Motion: little or none
- Layout: cookie-cutter designs with no character tied to the product
The good news from the same post: the models are very steerable. Tell one specifically what to avoid and what to try instead, and results improve right away. The catch is the word specifically. You can only steer toward something you can see and name. That's taste.
Ira Glass's gap, redrawn
The radio host Ira Glass has a well-known piece of advice about the gap every beginner goes through. When you start making creative work, your taste is already good. You can tell your work isn't as good as the work you admire, and that gap is painful. His advice was to push through it by making a large volume of work until your skill catches up with your eye.
AI has changed the shape of that gap. The skill side now closes almost instantly. An agent can produce competent work on the first try, in any style you name. The other side didn't move. Someone still has to look at the output and say whether it's good, why, and what should change.
If anything, that side got harder. A beginner used to judge one rough draft of their own. Now you're judging twenty polished options from an agent, and polish makes weak ideas look finished.
Practitioners describe taste in a way that fits this new job. Dylan Field, Figma's CEO, calls it knowing what's good and being able to say why. Designer David Hoang adds a useful test: calling something "modern" or "on brand" without explaining why isn't taste. The producer Rick Rubin is the extreme example. He has said he has no technical ability and knows nothing about music. Artists hire him for the confidence he has in his taste, and for his ability to say what he feels.
That's the role you play with agents. You don't need to do the craft. You need to recognize good work, explain it in specific words, and insist on it. We cover the bigger picture in building got cheap, judgment didn't.
Comparison beats study
The fastest way to train that eye is comparison, and research on how people learn styles shows why.
In a 2008 study, psychologists Nate Kornell and Robert Bjork asked people to learn the painting styles of 12 artists, six paintings each. Some saw each artist's paintings grouped together. Others saw them mixed, one painting from each artist in turn. Then everyone looked at paintings they hadn't seen before and named the artist.
On the first round of the test, the mixed group got 61% right. The grouped group got 35%. And here is the strange part: 78% of participants said the grouped approach worked as well or better, even after their own scores showed the opposite.
The authors suggest the mixing works because it puts different styles next to each other, so you learn what makes each one different. Seeing six paintings by the same artist in a row feels smooth and easy, and that easy feeling fools you into thinking you've learned.
Two lessons follow for taste:
- Don't scroll through "ten great dashboards" in a row. Put different approaches side by side and make yourself choose.
- Expect it to feel harder. That effort is the sign it's working.
Try a taste duel today
A taste duel is a two-option comparison where you have to pick a winner and explain why in specific words. It takes about fifteen minutes, and you can do it right now.
1. Pick one small screen. Keep it narrow. For a made-up testimonial app called Kudos, try the thank-you screen a customer sees after leaving a testimonial. Use whatever product you're working on.
2. Get two genuinely different versions. Screenshot two real products you use, or ask your agent (new to this? start with Claude Code for designers):
Make two versions of the thank-you screen that differ in one big
decision, not small styling. Label them A and B. Don't tell me
which one you prefer.
3. Pick in under ten seconds. Trust your first reaction. You'll explain it next.
4. Write why, in three sentences, with vague words banned. Name what you can actually point at: the size of the type, what your eye lands on first, the spacing, the exact words.
5. Note one thing you'd steal from the loser. The losing option usually gets something right.
6. Switch to a different kind of screen and repeat. Do a pricing page next, then an onboarding email. Mixing categories is the point, as the painting study showed.
Here's the difference a banned-word list makes, on the Kudos thank-you screen:
- Vague: "B feels cleaner and more modern."
- Specific: "B repeats the customer's own words back in large type, so they see what they wrote mattered. A leads with a big checkmark and 'Success!', which feels like a form receipt. B's single button, 'Close', respects that they're done."
Save every duel in a file called taste.md. After ten or so, read them back. The reasons that keep coming up are your taste becoming visible. Those are the beginnings of rules you can hand to your agents, which is what a DESIGN.md is for.
What a written taste can't do
One honest limit. A 2026 position paper at an ICML workshop argues that real connoisseurship can't be fully written down, only handed to someone competent at the moment of judgment. We think that's right, and it's a reason for hope rather than worry. Your written rules will catch a lot of what you care about, and they'll steer agents away from the average. The final look, the one that says "no, not yet," is still yours.
In lesson 0.4 of Chief of Agents, Taste 101, you run these duels with an agent, learn to spot the signs of average output, and leave with the first version of your own taste profile.
Sources
- Improving frontend design through Skills (Anthropic, November 2025)
- Ira Glass and what every successful person knows (James Clear)
- Figma CEO on how anyone can get good at design in AI (Behind the Craft, April 2026)
- The role of taste (and what it means) (David Hoang, 2023)
- Learning concepts and categories: Is spacing the "enemy of induction"? (Kornell and Bjork, Psychological Science, 2008)
- Mise en Place for Taste: Recipes, Connoisseurship, and Cultural Competence in Generative AI (Jun Li, ICML 2026 Culture x AI workshop)
Learn this inside Claude Code.
Join the waitlist for Chief of Agents. You'll get the free beginner's handbook and skills right away, and hear first when the course opens.
Keep reading
6 min read
Building with AI got cheap. Judgment didn't.
When AI can build an app in an afternoon, the hard part is judgment: what to build, what good looks like, and who will pay.
7 min read
Write your taste down: a DESIGN.md your agents follow
How to get AI to follow your design taste: write a DESIGN.md of short, checkable rules your agents read, then review screens against it.
6 min read
From designer to builder: directing a crew of agents at work
What working with AI agents looks like day to day after ten years in design: the work of direction, and what stayed the same.