Your Design System Isn't Enough
Models can read your tokens. They can't read your intent. Here are the four layers your design system is missing, and why codifying them is the real work.
I am a total Design system nerd and I think they’re one of the cleverest things the industry ever built.
I have a soft spot for the token discipline, the component libraries, the spacing logic, the motion principles. Design systems solved something really hard: how to produce interfaces at scale, with consistency. Entire teams can work in parallel and the product still feels coherent. New designers can onboard and ship without breaking things. It is a remarkable collective achievement.
Until lately, when models started being one of its increasingly important users.
The model reads your tokens. Understands your component logic. Applies your spacing rules. Renders your interface. It can also generate the design system for you. The craft your design system spent a decade encoding is now something a machine executes in seconds.
There is a catch:
You can now generate UI effortlessly and have nothing, faster than ever before. The interface comes out correct, consistent, on-brand, and somehow completely wrong, missing something.
That is because design systems were never designed to state intent.
Intent was always offered by people. The designer who had been on the product for five years and knew instinctively that this flow should feel considered and that one should feel urgent. The creative director who could look at two technically identical implementations and know which one had the right atmosphere. None of that was in the Design system. It never needed to be. There was always a human in the room.
There is not always one now.
We have done this dance before
This has happened before and the pattern is consistent.
Helvetica dominated typography for decades because typesetting was expensive. The moment Google Fonts made variety free, the landscape exploded. Products that kept defaulting to Helvetica stopped reading as neutral and started looking unimaginative. Print-quality CSS held as the marker of professional web design until Tumblr made expression cheap. Then the personal-web aesthetic shifted toward maximalism, glitch, the deliberate refusal of grid. The polished UI became the default interface.
AI collapses the cost of producing the interface itself.
You can already see the early signals. Neo-Brutalism 2.0 framed as a counter-trend to the clinical smoothness of AI-generated visuals. The return of texture and grain. Imperfection kept deliberately. The Are.na revival. The slow web. These are people reaching for something the polished average cannot give them.
I don’t know exactly what the most popular aesthetic becomes next, but I suspect the crisp UI is about to feel very dated.
Taste matters more, not less
There are two kinds of taste, and they are about to be pulled apart in a way they never were before.
Syntactic taste: knowing which typeface pairs well, which spacing feels right, which component variant fits the context. This is the taste the design system encoded. It is also the taste the model already has. It becomes commonplace, just as it is typical to expect a crisp UI.
Atmospheric taste: the felt sense of what something should be like to inhabit. Not “does this button have the right radius” but “does this product feel like arriving somewhere or being processed.” Not the visual choices but the emotional logic underneath them.
This information was never in the design system because it was too hard to write for a start. For years designers could lean on syntactic taste and call it enough design. What is arriving now means atmospheric taste, the thing most designers never fully had to develop because the system absorbed so much, becomes a really big part of the job.
Beyond Atmosphere
Atmospheric taste is one dimension of what is missing. There are others, and they share the same problem: they were always present in the work, input by the people doing it, and never made explicit enough to survive without them.
Here are a few examples:
Situational intent. Who is this person in this specific moment. The nurse at the bedside is not the patient in distress is not the administrator doing the quarterly audit. Same product, completely different needs. The design system has no way to know that.
Relational intent. What kind of relationship is this product trying to have with its user. A tool. A guide. A collaborator. These are different stances that shape every micro-decision downstream. The system encodes what the product does but never what it is trying to be.
Ethical intent. What will this product refuse to do. This becomes urgent when the thing rendering the interface is an agent making decisions at speed, without a human in the loop.
Temporal intent. What should this product build in the person using it over time. Not just completing the task. What understanding should they leave with that they did not arrive with.
The design system encodes the what. None of it encodes the why. Agents do not fill gaps with judgment. They fill gaps with the syntactic average and they guess.
The space is already evolving
The most interesting structural move I have seen from any design tooling this year is the concept of a portable intent file (DESIGN.md): a document that travels with the project and carries design rules that agents can reason from. The idea is right on point: stop keeping intent inside people’s heads, make it a document and make it machine-readable. Ensure it travels with the work. You can check out a Github repo full of them here, and make your own.
What is in those files right now is still almost entirely syntactic at the moment. Colour tokens. Spacing preferences. Component rules. That sort of thing. The four intent layers above are not in there yet. The harder part of what needs to go inside it is still an open problem.
Design is getting really serious
For a long time, a significant portion of what designers spent their days on was work the models can now do. Applying tokens. Ensuring consistency. Maintaining consistency. Necessary, real work, but not the work that requires a human who has thought deeply about what it means to live inside a product.
Deciding what something should feel. Writing the relational stance a product takes toward its users. Defining what it refuses to do, and why. Encoding the emotional logic of a flow so that an agent can render it a thousand times without losing the thread.
This is serious work. It requires taste, yes. It also requires ethics, philosophy, a genuine point of view about what technology should do to a person and what it should not.
Generative interfaces are arriving. Fluid, personalised, regenerated for each user in the moment of use. The same product producing a different surface for the patient in distress and the nurse at the bedside. The same educational interface generating differently for the student who is struggling and the one breezing through. The design system stops being a library you ship and starts being a vocabulary for the next render.
The work in front of us is to figure out what that vocabulary needs to be.
So here are three questions worth sitting with this week:
What does your product refuse to do?
What kind of relationship is it trying to have with the person using it?
How are they changed by your product?
The design system happily carried us to this point. The next era is ours to write.
Thank you, wonderful read