Writing

Taste is the bottleneck

EssayJul '262 min read

When production is free, judgment is the scarce input. Notes on what taste actually is and how teams build it.

Every team I talk to has the same new problem, and most haven't named it yet: they can produce far more than they can evaluate.

Agents draft the UI, write the copy, generate the variants. Output that used to take a sprint takes an afternoon. And then it sits there, plausible, complete-looking, and unjudged, because the scarce input is knowing whether the thing is any good.

That's taste. A fast, reliable quality function: the ability to look at something and know quickly, for articulable reasons, whether it's right and what specifically is wrong when it isn't.

AI output regresses to the mean of everything it learned from, and the mean of software is mediocre. Left unjudged, generated work is competent: correct spacing, reasonable copy, familiar patterns, and no differentiation. Slop isn't broken work; that would be easy to catch. Slop is fine work, at volume, from every team simultaneously.

Which means the differentiator inverts. When every team can execute, the moat is the quality function you run the output through: how sharp it is and how fast.

The word makes it sound innate. It isn't. Every strong designer I've worked with built their quality function the same way: high-volume exposure to great work, held next to their own, with someone senior forcing the comparison. Critique is taste training. So is copying excellent work stroke by stroke to feel where your version deviates.

What's changed is the urgency. A junior designer used to spend years producing slowly, and taste developed as a byproduct of all that repetition. Now production is instant, and the reps that quietly built judgment are gone. I wrote about the same shift from the other side in The default designer now builds. Teams have to train taste deliberately or it doesn't get trained at all. Concretely:

  • Critique more, produce less. The scarce skill is evaluation; practice it in the open, on real work, with reasons attached.
  • Keep a canon. A shared, argued-over collection of what great looks like in your domain. Your agents can be pointed at it; more importantly, your people calibrate against it.
  • Never ship unjudged output. The moment "the agent made it" becomes an accepted provenance, your bar is the model's mean.

For design leaders the shift is blunt: your team's throughput is no longer your problem. Your team's threshold is. The same quality function has to sit on the constraint system, not just the artifact — that's the live question in Generative UI. The organizations that win the next few years won't be the ones that generate the most. Everyone will generate roughly infinitely. They'll be the ones whose people can tell the difference and are trusted to act on it.

Production is free. Judgment is the price.