Taste Is the Last Skill

Last month I needed one headline for a landing page. I asked a model for options and had forty of them in about ninety seconds. Then I sat there reading. A few were clearly bad. Most were fine. I could not tell you which one was good, and I lost the rest of the afternoon to a decision that five years ago would have been settled by whichever line I had the energy to write.

The making took ninety seconds. The choosing took four hours. I think that ratio is the whole story of the next few years.

Production Got Free. Selection Didn't.

For most of history, the expensive part of creative work was making the thing. A record needed a studio. A film needed a crew and a budget. An app needed months of someone's life. We complained about that cost constantly, and we never noticed it was doing a second job for us.

It filtered. When making something costs you weeks, you only start the ones you believe in, and you abandon the ones that stop being worth it halfway. The budget was a judgement machine that ran automatically, without anyone having to be wise.

That machine is gone. Not weakened - gone. The amount of stuff that exists goes up, the average quality goes down, and neither of those is because people got worse. The filter was the cost, and the cost has fallen through the floor. So the filter has to move somewhere else, and the only place left is inside the person doing the work.

What Taste Actually Is

Taste gets discussed as though it means preference - liking obscure records, wearing a lot of black, having opinions about typefaces at parties. That is not it. Preference is what you enjoy. Taste is being right about what works.

You have seen it in a senior engineer who looks at two designs and says the second one is going to hurt in about six months, and cannot fully justify it, and turns out to be correct. That is not intuition in the mystical sense. It is a compressed model of a few hundred previous designs, most of which they got wrong at the time and had to live inside afterwards. The same faculty shows up in a kitchen, in a mix, in a paragraph.

Preference is what you like. Taste is being right about what is good, usually before you can explain why.

And a large part of it is subtraction. It is the same argument as the one for deleting code: the hard judgement is never which thing to add, it is which of the forty perfectly acceptable things to throw away. Adding is a decision anyone can make. Removing requires you to be sure.

Why It Doesn't Automate

Three reasons, and none of them are about models being weak today.

1. There is nothing to check against. AI systems get good at whatever has an honest test attached - code runs or crashes, tests pass or fail. 'Is this good' has no test. A model can produce a thousand variants and has no way to score them, so it converges on competent and stops there.
2. Models are trained toward the middle. What you get back is, roughly, the average of everything ever written on the subject: pleasant, fluent, unsurprising. Taste is almost always a departure from the average. Something built to be the middle cannot reliably prefer the exception.
3. Taste is a set of refusals, and refusals are personal. The real question is never 'what is best in general' but 'what is right for this thing, made by me, for these specific people.' A model does not know who it is for. It is answering a much easier question than the one you are actually asking.

Why It Doesn't Teach Fast Either

You can teach the rules in an afternoon. Contrast, hierarchy, cut the adverbs, one idea per function. Rules are genuinely useful and they are not taste. Taste is the thing that tells you which rule to break this time, and it cannot be handed over in a lecture because it is not information. It is calibration, and calibration requires error.

You build it by making a lot of work that isn't good, watching it land badly in front of someone whose opinion you care about, feeling that properly, and going again. The unpleasantness is not a side effect of the process. It is the process. Nothing calibrates without a gap between what you meant and what came out.

Which is where this gets uncomfortable. The reps that used to build taste are exactly the work we are now handing off. Juniors wrote the clumsy first version and had it pulled apart in review, and that is where the model got built. If the first version now arrives already competent, the error never happens, and there is nothing to learn from.

You cannot calibrate on work you did not do.

I notice it in myself already, which is the part I would rather not admit. I accept generated code I would not have written. Not because it is better - because reviewing is passive and writing is active, and judgement goes soft in the passive mode. It is easy to nod along to something reasonable. It takes real attention to notice that reasonable is not the same as right.

How to Build It Anyway

None of this is a reason to go back to doing everything by hand. It is a reason to be deliberate about where your reps come from now.

• Finish things. Taste comes from consequences, and unfinished work has none. The project that nobody saw taught you nothing about what works.
• Say why. 'I don't like it' teaches you nothing next time. Force the sentence: the rhythm is wrong, there are three ideas here competing, this is a solution to a problem we don't have. Naming the reason is how a vague reaction turns into a model you can reuse.
• Keep the rejects. Save the thirty-nine you didn't pick, and read them again in a month. You will find out whether you were actually right, which is the only feedback that matters.
• Put constraints back on purpose. One option. One hour. No regenerating. Constraints reintroduce the cost that used to do your filtering, and they force a decision instead of a survey.
• Spend deliberate time with work that is genuinely excellent, including outside your field. Not to copy it. To reset where your sense of 'good' is sitting, because a feed of competent output drags that line downward all day without telling you.
• Keep doing some of it by hand, slowly and worse than the machine would. That is not nostalgia. It is maintenance on the only instrument you have.

The Last Skill

I don't mean last as in most advanced. I mean last as in remaining. Piece by piece, every part of creative work that could be written down precisely enough to explain has been handed over, and that handover is mostly fine. What is left is the part that was never written down: looking at forty acceptable options, knowing which one is the one, and being willing to stand behind that in a room full of people who liked a different one.

The forty options are only going to get cheaper. The person who can look at them, choose, and say why is close to the entire job now. And that person is still assembled the slow way, out of a great deal of bad work done on purpose.