Last week I asked a model for forty buttons. Not literally forty, but variation after variation of one component, all generated in about ten minutes. Then I sat and looked at them for far longer than that.
It would have taken a day by hand a few years ago: a different radius, a different weight, the label in mono, the label in sans, a version with a little bracket on the corner, a version without. Now it's ten minutes of typing.
And then I just looked at them. For way longer than ten minutes. That's the part nobody automated. The making got cheap. The choosing didn't.
01 The work moved
I used to think my job was production. Push the pixels, build the file, ship the screen. That was the thing I traded for money, I was good at it, and for a long time it felt like the skill.
It isn't the skill anymore. I can describe a layout and get a working version back before my coffee's cold. I can stand up an MVP in an afternoon that used to be a two-week sprint with a handoff in the middle. The tools, Claude Code, Cursor, v0, are genuinely incredible, and I'm not going to do the thing where a designer pretends they're not.
But here's what happens when production gets free: you generate ten good options and you still have to know which one is right. The bottleneck doesn't disappear, it just moves up, into judgment, into the quiet voice that says no, not that one, the other one, and usually can't explain why.
That voice is taste. I've never found a way to outsource it, and I have genuinely tried.
02 Why my site looks like a terminal
People ask why my portfolio reads like a command line. Monospace everything, and a status line blinking at the top like the page is a process that hasn't finished running.
The honest answer is not that it converts. I have no evidence it converts. A clean SaaS template would probably convert fine.
I made it look like this because that's where I actually live. I spend my days in editors and consoles, designing tools for people who also live in editors and consoles, operators fixing hardware in a data center, where a wrong click costs real money. That world has a texture: mono type, dense information, no decoration that isn't load-bearing. I find it beautiful, so the site looks like the inside of my head instead of a conversion funnel.
That's a taste decision. You can't brief it. You can't A/B test your way to it, because A/B testing only tells you which of two things you already made performs better. It can't generate the conviction that made you build the weird one in the first place.
03 Taste is your history, compressed
Here's what I think taste actually is, under the hood.
It's every interface I've ever used and quietly hated. It's the finance years at AllianceBernstein and American Express, living inside the enterprise tools nobody chooses and everybody depends on, absorbing what dense-but-usable feels like before I had words for it. It's the operator tools at Meta, where I learned that trustworthy and pretty are different things, and the wrong one gets someone hurt. It's a thousand small reactions, that's too rounded, that gradient is trying too hard, that's the right gray, that I never wrote down and couldn't fully justify if you asked.
Left alone, a model drifts toward the average of everything it has ever seen. That average is exactly the thing I'm trying to avoid. My job, the part that's still mine, is to have a point of view about which direction to push away from the average, and that point of view is built from my specific, non-transferable history of liking and not-liking things.
You can't prompt that into existence. It isn't information the model is missing, it's wear, and I put in the years to get it.
04 The uncomfortable part
I'll be straight: this is also the thing I'm betting my livelihood on, and some days that bet feels thin.
Because taste is invisible. It doesn't show up in a diff. No one claps when you reject the obvious version and sit in the discomfort of this isn't right yet for another hour. Clients pay for outcomes, and increasingly the outcome can be generated by anyone with a good prompt. The value I add is in the editing: the choosing, and the no. Editing has always been the hardest thing to put a price on.
But I keep landing in the same place. When everyone can make everything, what's scarce is knowing which thing is worth making, and being willing to hold that line when a client would rather you didn't.
Forty buttons took ten minutes. Knowing which one to ship is the whole job, and it's still just me doing it.
After a decade of thinking my value was how fast I could produce, that turns out to be a strangely reassuring place to land.