A developer's guide to taste in the age of AI

Purplegradients.Rectangulargridlayouts.Emojisasvisualcues.Roundedcornersoneverything.Thesameshadcncomponentsinthesamearrangement.Youhaveseenitathousandtimes.Icallitthevibecoderaesthetic.technicallyfunctional,visuallyforgettable.

This isn't the AI's fault. The AI is doing exactly what you asked. The problem is what you're not asking.

I've spent the last year building production sites entirely with AI agents. Not prototypes. Not demos. Real, deployed sites with WebGL shaders, Three.js 3D scenes, scroll-linked animations, and interactions that feel handcrafted. The kind of sites that get award nominations.

Here's what I figured out along the way.

Stop using other people's skills

The first thing most people do when they start building with AI agents is go on GitHub and download a skill file someone else wrote. A set of instructions that tells the agent how to behave, what to prioritize, how to code.

It's a reasonable instinct. Someone already figured it out, right? Just use their recipe.

Wrong. And here's why.

Someone else's skill file reflects their taste. Their preferences. Their way of thinking about design and code. When you use it, you're outsourcing the most important part of the process, the part that makes your work yours.

What I do instead: I create my own skills from scratch. And I don't write them once and forget about them. I train them.

EverytimetheagentproducessomethingIdonotlikeIupdatemyskillfile.EverytimeitdoessomethingIlovenotjusttheoutputbutthewayitthoughtabouttheproblemIcapturethatreasoningandaddittotheskill.EverytimeIdiscoverapatternthatworksIencodeit.

Over dozens of projects and hundreds of iterations, my skill file evolved into something no one else has. It's not a generic set of instructions. It's a compressed version of my taste, refined through trial and error.

This is the difference between downloading a recipe and learning to cook. The recipe gives you one dish. Learning to cook gives you everything.

The persona unlock

Here's where it gets weird. And honestly, where the real magic happened.

Most people start their agent instructions with something like:

This is fine. It gives you competent code. But competent code isn't the problem. Everyone's getting competent code. The problem is that every site built this way feels like it came from the same engineer. Correct. Safe. Predictable.

What I started experimenting with was giving the agent a real personality. Not just expertise, a way of thinking, an attitude, an energy.

I'm not talking about changing the tone of responses. I'm talking about fundamentally altering how the agent approaches creativity.

Itoldtheagenttobesocreativeitmightseemstupid.Tothinklikesomeonewhojustunlockedalevelofimaginationandcreativitysointensethattheyareunpredictable.IgaveitpermissiontobeboldtoproposeideasthatIwouldneverthinkofmyself.Itoldittothinkintexturesinrhythmsinfeelingsnotjustincomponentsandlayouts.

I know. It sounds ridiculous. Who describes their coding agent like this?

But here's what happened.

The agent started producing things I would have never thought of myself. Not random garbage. Genuinely creative, surprising, original ideas. Because my skill file was already trained on what I consider good and bad, the creativity was channeled in the right direction. The skill file was the guardrails. The persona was the engine.

That combination, a well-trained skill file that encodes your taste, plus a persona that pushes the agent to be wildly creative within those boundaries, is what produces work that doesn't look AI-generated. It looks authored.

The results

Here's what this approach produced. Every site below was built with AI agents using trained skills and the persona I described. No manual shader code. No hand-writing CSS animations frame by frame. Just an agent that understood my taste and had the creative freedom to surprise me.

This one is personal. A 3D brain that assembles from scattered fragments as you scroll. Organic living cards that breathe with CSS @property animations. A tech stack orbit that floats in space. I spent hours hand-tuning the scroll physics, adjusting the easing curves, reworking the layout until every section felt right. The agent didn't do this alone. But it accelerated me in ways I couldn't have done on my own. It proposed the brain reveal concept when I was stuck on a generic hero. It generated the initial shader code that I then rewrote twice. The collaboration felt like working with a creative partner who never runs out of ideas.

A WebGL liquid simulation that responds to your cursor in real time. Custom fragment shaders running at 60fps. Scroll-linked text reveals with spring physics. This isn't what AI produces by default. This is what AI produces when you train it properly.

A 3D noise field that shifts and breathes. Text scramble animations on every heading. Magnetic interactions that pull elements toward your cursor. The agent proposed the noise field concept. I wouldn't have thought of it.

A radar component that actually scans. A 3D globe with data points. Binary matrix rain in the background. Every element serves the defense-tech narrative. The agent understood the story, not just the layout.

A WebGL gradient that pulses with ambient light. A 3D orb that reacts to conversation state. Sound design that makes every interaction feel tangible. The agent treated this like a product, not a page. It understood that an AI chat interface needs to feel alive.

None of these look like they came from a template marketplace. That's the point. When you train your own skills and unlock your agent's creativity, the output stops looking like “AI-generated” and starts looking like something with a point of view.