The Ladder, Pulled Up
For most of the first decade of my career, I was in spec docs. Not the glamorous part of design. The part where you’re sitting at 11pm walking a button backward through three nested containers to make sure its spacing is on the 8px grid. Not in spirit, but actually, mathematically, in the file. Where you’re in a forty-minute conversation with an engineering lead about whether a value is twelve or fourteen because the design system says one thing and the implemented component says another and somebody has to be wrong on purpose. Where you’re auditing your own Figma file the night before a review because if anything is off by a pixel, somebody on the team will see it, and they should, because that’s what makes the system actually a system.
I have always operated from taste, but the taste was built in those rooms. Not in the workshop where we sketched the concept. In the long, unglamorous tail of making the concept hold up to its own rules.
The other thing I did, for years, was build working prototypes for founders. Not mocks. Real, clickable, demoable prototypes. The kind you could put in front of an investor or a customer to argue about a concept. A week from sketch to a thing on a URL. I built dozens of those. The skill of taking something that did not yet exist and dragging it across the line where it became real enough to fight over was a muscle I built one rep at a time.
Both of those rep streams have been automated. Not partially, not later. Now.
The dominant claim in the design conversation right now is that taste is the moat. AI can produce. It can iterate. It can stamp out competent variations of almost anything you point it at. What it can’t yet do is pick which version was the right one. Nobody has convincingly explained how it could. Judgment survives. Taste survives. The work shifts from making to choosing, and the people who can choose well are the ones who keep the job.
I think the strong version of this is correct. I have spent the last year using AI tools harder than anyone I know, and the bottleneck has not once been the model’s ability to generate something. The bottleneck has always been knowing which of the things it generated was actually right.
But the people saying this loudest are the people who already have taste. And almost nobody is asking the next, obvious question. The one that gets quieter the more carefully you ask it.
Where does the next person’s taste come from?
Taste is not a static asset. It’s residue. It’s what’s left over after a particular kind of work has been done a particular number of times. The taste I have for spacing came from years of losing arguments about an 8px grid until I could see the grid in my sleep. The taste I have for what makes a prototype feel real instead of fake came from building a hundred of them, mostly badly, sometimes well. The taste I have for what to put in front of a user came from sitting through research sessions and watching what didn’t work, over and over, until I started knowing why before the user said it.
I came up the long way. Most of the senior designers I respect did. And we are, almost without realizing it, the last generation produced by the apprenticeship that has just been kicked out from underneath everyone behind us.
The “taste is the moat” story is comforting if you happen to already have some. It is much less comforting if you are trying to figure out where any of it is supposed to come from.
The apprenticeship that produced my generation ran on specific work. Junior designers spent years polishing pixel grids and groomed Figma files until they could see geometry the way other people see grammar. They built clickable prototypes for ideas that mostly went nowhere, which is exactly what made them useful. You cannot learn how to make a thing feel real if every thing you make has to ship. They held the line on design systems, often for teams that didn’t always want one held, and in doing so absorbed how a system actually works under stress. They translated between designers and engineers, which is its own slow education in what a medium is actually made of.
Each of those reps produced a specific layer of taste. Each of them is being automated, and the org chart is reshaping around it.
You can see this in how teams are hiring. Most companies I talk to are now staffing at one of two ends. Senior designers who already have judgment and can direct, and a small number of fast, code-fluent generalists who can move quickly with agents. The middle, where taste used to actually get built, is hollowing out. There is no longer a serious junior-to-senior pipeline at most places, because the work that used to fill it is being done, faster and cheaper, by a tool.
I happen to be okay. I climbed enough of the ladder, early enough, that I’m now in exactly the spot the new system rewards. The same systems-thinking that made me good at design systems makes me good at directing agents. The work I did adhering to other teams’ frameworks taught me how to specify behavior precisely enough to get something specific back. The strategic, design-thinking position my career drifted into is, almost embarrassingly, the perfect spot from which to orchestrate models all day. The career I built without intending it turns out to be one of the most useful careers to have right now.
That is the part that should make us uncomfortable. I am not okay because the system is fair. I am okay because I happened to climb the ladder before someone pulled it up.
So what does the new apprenticeship look like, given that the old one is gone? I don’t fully know. Nobody does yet. But I have a working theory in three parts, and I think it is at least the right shape.
The first move is that the core rep changes. The old rep was production. Make the thing, again and again, until you can feel what’s wrong with it. The new rep is critique. Not the soft “this feels off” version. The other one. The kind a senior used to give a junior over their shoulder, where the diagnosis lands so precisely that the fix is almost automatic. When an agent generates a destructive action button, the weak critique is “make it feel more serious.” The strong one is: the verb is too soft for the consequence, the type weight doesn’t carry the gravity, the affordance reads like marketing copy instead of a product surface, the empty state assumes a level of trust the user has not earned. That kind of precision is teachable. It is also almost nobody’s job to teach it. The thing the agent does for you is volume. The thing you have to learn to do is push back against the volume in language exact enough to fix it.
The second move is that what I’d call pre-prompt taste matters more, not less. The opinions you walk into the room with, before the model hands you fifty plausible options, are the ones that survive contact with it. Without them, you accept the default, which is invariably an average of every interface ever made. Competent, generic, slightly soulless, slightly wrong. With them, you have a center of gravity. You know what you’re after. You can reject most of what’s offered without thinking about it. You can describe the missing thing precisely. The way you build pre-prompt taste is the way it has always been built. By looking, hard, at things that are good. By reading the original sources of the patterns the model has compressed into a single greasy average. By copying, deliberately, not as the lazy move but as the serious one. The library is back. So is the practice of sitting with something good for long enough to finally see what makes it good.
The third move is harder to make sound serious because it skirts the edge of a sermon, but: sometimes you have to take the tool away on purpose. Not as a refusal of the new thing. As training. Athletes train under load; designers can train under constraint. Once a quarter, take a flow you would normally hand to an agent and audit it the long way. Write the spec by hand. Read the research transcripts straight, not summarized. Sit through your own onboarding the way a new user would, with no shortcuts. The point is not nostalgia for the way it used to be. The point is that the rep deficit is real, and the only thing that has ever closed a rep deficit is reps.
I do not know exactly what the new studio looks like, and anyone who tells you they do is selling something. What I am reasonably sure of is that AI labs are not going to figure it out, because it is not their problem. The companies that figure it out will have a reason their juniors get good. They will name the new equivalent of polishing pixel grids. The work that is generative for taste even if a model could do it faster. And they will require their seniors to teach it. Not as a kindness. As the job.
A generational gap is forming, and it is quiet. Most teams haven’t noticed because their senior designers, the ones produced by the old system, are still producing. We are the holdouts. When we leave, or retire, or burn out, the gap is going to be loud.
The ladder is up. Someone has to put it back down. Probably us. The ones still on it, still calling taste a moat, still writing posts like this one.
It would help if we said so.