After my Claude Design review, four PwA members have independently asked about design styles that don’t look blatantly AI-ish.
Or, more precisely: how to brief AI when the visual is clear in your head but not on your tongue.
I’ll know it when I see it, I just don’t have the words to describe what I really want.
— Daniela
I can picture exactly how I want it to look, but I struggle to translate that vision. I eventually adapted an existing design, and while the result is fine, it doesn’t feel like something I’m particularly proud of.
— PwA Reader
We usually learn to like before we learn to explain why. And that's okay.
Don’t ever let it make you feel like you’re not ready to start experimenting. None of us were born with a design vocabulary.
We are not covering centuries of visual design in one Substack post. Instead: hero sections with prompts and code, plus a workflow for translating every possible I’ll know it when I see it into design direction for your AI.
We’ll do that 11 times, using the exact same website in 11 different design styles.
You’ll also enjoy this article if you’re curious what Kimi K3 can do while the internet continues its Claude-and-GPT hype cycle.
Or, if your brain likes looking at beautiful things.
AI-ish Design and Claude-ish Design
AI-ish design (also: Claude-ish design) is the readily identifiable default aesthetic of interfaces generated by AI coding tools without clear human art direction.
Some telltales: the indigo-to-purple gradient style that Claude is terribly fond of. Identically rounded cards with soft shadows. Badge pills everywhere. Tailwind, Tailwind, Tailwind.
But I think that definition misses its most important tell: we look at it and feel nothing. Annoyance, maybe. But mostly nothing.
I saw far more indigo websites in 2025 than I do now. So: progress. But plenty still leave you with that same blank feeling.
What Changes When YOU Direct it
The table below shows what changes when we guard our taste like a hawk:
Why I Chose Kimi K3 as the Model For this Experiment
Because Kimi K3 is shockingly good at design even before you give it a single Skill. That’s because it can see its own work.
Kimi K3 arrived in July 2026 from Moonshot AI: 2.8 trillion parameters, open weights.
It runs an agentic visual loop. Which basically means it iterates between code and live screenshots. On July 17, that capability helped put it at #1 on Arena’s Frontend Code Arena, ahead of Claude Fable 5.
Yes. Fable 5 got beaten at frontend design by a Chinese open-weight model.
Hey, I’m Karo Zieminski 🤗.
AI PM and builder. I write Product with Attitude, an AI newsletter for tens of thousands of readers across 146 countries, helping them develop critical AI literacy the only way it sticks: through practice.
Welcome to the 399 of you who have joined us since last week! Here’s what you might have missed:
What ChatGPT Work, Perplexity Computer and Claude Cowork assume about us, the users
Stop discussing Pangram without reading the AI-assisted Craft Manifesto
Subscribe for free, and I’ll keep bringing you deeply researched,
human-written articles like this.
And if you share this article with 3 friends or colleagues,
I’ll thank you with 3 months of a premium subscription.
The Experiment: Kimi K3, the Same Words, 11 Hero Sections
Eleven hero sections, one newsletter website, identical copy and subscription structure, built 11 times in 11 different design styles. I wanted the visuals (and textures!), not the words, to do all the talking.
Design Style 1: Conceptual Editorial Minimalism
Design Style 2: Nocturne Editorial Surrealism
Design Style 3: Cinematic Frosted-Glass Editorial
If any of these designs catches your eye, please share or restack this guide so more people can find it.
Design Style 4: Punk Editorial Collage
Design Style 5: Surrealist Watercolor Editorial
Design Style 6: Surrealist Retro Collage
Design Style 7: Futurist Glass Editorial
Design Style 8: Surreal Literary Editorial
Design Style 9: Swiss Halftone Editorial
Design Style 10: Botanical Luxury Minimalism
Design Style 11: Alpine Cinematic Editorial
The Experiment: Part 2, Your Turn
And now I’d love to bring YOU into the experiment.
Close your eyes (and please don’t scroll back!).
Name the first two designs you can still see. Those are the ones that left the strongest imprint on your brain.
Which two designs appear first?
You can download all of the designs above (and more) from PwA Design Resources. Find the password here.
And if you’d like to build from scratch, here’s a workflow I recommend.
3 Steps to Direct Your Next Redesign: How to Build Stunning Websites When Your Taste Is Clear but Your Vocabulary Isn’t
Visual Prompt Engineering
Visual prompt engineering is the practice of pairing images with written instructions to give AI clearer context and guide its output.
I recommend reading Figma’s Visual Prompt Engineering Guide.
First rule: if the design vocabulary isn’t coming, don’t force it. Switch to visual language instead.
The visual reference-to-prompt workflow turns an 'I'll know it when I see it' into design direction your AI can execute.
Browse sites like Cosmos or Behance and scroll until you see something that, well, stops the scroll. Save it for later. I mean it, I’ve made this mistake often enough to know better: save that thing immediately (that’s rule number 2). You can use tools like Cluing, bookmark it, anything.
If one of the saves feels close to what you want for your homepage, use it as a visual reference in the prompt. I’ve seen people paste it into ChatGPT first and asking it to describe the design in precise language. They were happy with the results. But in my tests, giving Kimi K3 the image AND the prompt worked even better.
Rule #3: prepare your UX copy (the words people see on your site) before you prompt.
Using today’s example, this is the copy I gave Kimi K3:
Now hand it over to your AI. Just don’t paste in one of those giant “ultimate web design prompts” from the internet. Instead, ask the agent to interview you, one question at a time, until it has enough context to understand what you want. Until you both do.
You are a team of senior product designers, art directors, UX experts, and design engineers. Interview me one question at a time until you are at least 95% confident you understand my requirements.
That’s AI-assisted craft in practice: let the model do the assembly work, but keep the standard. (And please read the manifesto!)
If you choose to use Kimi K3, I accessed it via Perplexity Computer. Here are some other options:
I hope this helps!
You Might Also Enjoy
Claude Design for Substack: 8 Ways to Use It by AI Meets Girlboss
Starting a Substack? Without Branding Your Writing Is Invisible by AI Meets Girlboss
Absolutely love it! Thanks Karo :)
I truly believe that everyone has their unique aesthetic taste and judgment about text (even when they claim they don’t).
There’s a big gap between being able to tell that something looks good and being able to prompt for the good. You’ve really helped bridge that gap.
11 variations from one prompt is a great way to expose a model's default biases. What's interesting is that most AI models default to the same 2-3 "safe" visual patterns unless you actively fight against them with negative prompting. Curious whether you tried asking Kimi K3 to explicitly avoid its own most common outputs from the first few rounds. That kind of self-referential negative prompting has worked well for me when iterating on UI variants.