How to Turn Aesthetic Judgment into Parameters

A lot of designers believe AI-generated work always misses something — taste.

Take Glassmorphism. One execution feels elegant. Another feels mechanical. That distinction is hard to articulate. It's years of trained intuition.

You prompt "Glassmorphism style." AI gives you a result. You prompt it again. You get something different. Both technically qualify as Glassmorphism, but they're not the same thing. Try prompting "Brutalism" three times and you'll probably get three completely different answers. Professional terminology solves the efficiency problem. It doesn't solve the consistency problem.

This is the most underestimated challenge in AI-assisted design: it's not that AI isn't smart enough. It's that AI has its own interpretation of every word you use.

We're facing a real engineering problem. We can't ask designers to supervise every AI output like a film director, refining step by step — that's too slow, and it defeats the purpose of using AI. But we also can't accept every output being AI's interpretation of "Glassmorphism" on that particular day — that doesn't scale, and it doesn't work for team collaboration.

What we need is this: whoever triggers it, whenever it's triggered, AI consistently produces work at the professional standard of this team — not AI's interpretation of a word.

The answer isn't finding more precise vocabulary. The answer is translating your aesthetic judgment into parameters that AI cannot misinterpret.

And that requires working across six dimensions at once.

Let me walk through what this translation actually looks like.

Say my product needs to feel "restrained, professional, but not cold." Any designer gets it immediately. AI doesn't — how restrained is restrained? Professional in what industry?

I need to translate that feeling into parameters:

When "restrained professionalism" gets translated into these parameters, something important happens: the designer's aesthetic judgment becomes a constraint that AI can actually follow.

That's the real value of Design Tokens — not just as an organizational tool for design systems, but as the parameterized encoding of aesthetic judgment. The same logic applies to any style vocabulary. "Glassmorphism" in your system is no longer just a word. It's a set of parameters:

Once visual consistency is solved, the next dimension is dynamic interaction.

A CTA button used in production needs every state defined. What happens on hover? What state does it enter when a form is submitting? How does it communicate failure? Without constraints, AI drifts.

So: every component needs its complete states defined.

Base states: Default, Hover, Active, Disabled, Focus.
Business states: Loading, Error, Empty.

In agentic system design, there's an additional layer of states that simply don't exist in traditional specs:

These aren't just visual problems. They're the moments where users build — or lose — trust in an AI system. A "Waiting for Human" state that isn't clearly communicated can silently stall an entire workflow. Production-grade design systems need complete specs that both human designers and AI outputs are held to equally.

Spacing matters more than most people think.

Left unconstrained, AI drifts — it'll use 13px, 13.5px, 15px, whatever feels approximately right, even if you've told it to strictly follow existing spacing conventions.

So down to spacing, we require everything to come from multiples of a 4px base: 4, 8, 12, 16, 24, 32, 48px. And it gets written into design.md.

Spacing gives a digital interface its rhythm. With this constraint in place, AI-generated pages naturally develop the breathing room that professional design has.

Typography is the easiest thing for AI to get wrong.

Without constraints, AI freely mixes type sizes — 15px here, 17px there, 16px somewhere else. The deeper problem is that type hierarchy determines which information reads as more important. A lot of the time, when people admire a high-quality design, typography is doing most of the work.

So we need a complete type scale: from Display to Caption, every level with a fixed size, weight, and line height. AI can only choose within that framework.

Weight and line height matter as much as size. An H1 at line-height 1.2 feels completely different from 1.5. Leave these undefined and AI gives you a different answer every time.

One dimension that's easy to overlook: copy.

Copy is the most direct conversation between your design and your users. Without it, design is just a shell.

So our Content Guidelines need to treat AI as an audience too. Error messages should explain what went wrong and what the user should do next. Empty states shouldn't just say "nothing here" — they should explain why, and point to what comes next.

In agentic systems, copy faces a challenge that didn't exist before: how do you explain to a user what an AI is doing right now? That copy needs to find a balance between being transparent enough and not burying people in technical detail.

Design QA has always been part of the process. Now we can bring an Agent into it.

Layer 1: Design QA Agent (automated)

Everything with a standardized answer gets handled by the Agent:

Anything out of spec gets flagged immediately.

Layer 2: Human Review (judgment)

What passes the first layer goes to Human Review. This is where the higher-order quality calls happen:

Together, these two layers turn QA from a human checklist into something that actually works at scale: Agent handles everything with a right answer, Human handles everything that requires judgment.