Forty Good Options, One Hour, and the Only Skill Left

I gave an agent a task last week. It came back with forty good options in under a minute, and then I sat there for an hour. That hour is the whole story of the next ten years, compressed into one afternoon at my desk.

Making was the bottleneck. It isn't anymore.

For almost all of working life, making was the hard part. You got paid because you could do the thing other people couldn't do fast: write the page, build the model, draft the plan, lay out the deck. Skill meant production. Speed plus craft of output. That was the whole game, and most careers were built inside it.

The machines took that and made it cheap. Not perfect. Cheap. Fast, plausible, often genuinely good, sitting there the second you ask. My forty options were not garbage. Several of them, a year ago, would have been a solid afternoon's work from a competent person.

That is the part people keep getting wrong. They wait for the output to be bad so they can feel safe. It isn't bad. It's fine. It's fine at a volume and speed no human can match, and that is precisely what changes the math. When the thing you were paid to produce becomes abundant, your value doesn't disappear. It relocates. The question stops being can you make it and becomes do you know which one is right and why the other thirty-nine aren't.

Value floods toward whatever stays rare

There's a dry idea in economics that explains the whole afternoon: the scarce input captures the value. In any process, the rare ingredient is the one that gets paid. Everything abundant gets commoditized down toward its cost.

Watch how this played out before. When electricity got cheap, the money didn't go to people who could generate power. It went to the people who knew what to build with it: the factory line, the cold chain, the radio. The scarce thing wasn't the current. It was the judgment about where to point it.

Making just became the cheap electricity of knowledge work. So the scarce input moved. It moved to the capacity to look at forty competent options and feel, with something close to your whole body, which one is true and why the rest are merely acceptable. That capacity has an old, unglamorous name. Taste. Judgment. Discernment.

This is not a soft skill in the dismissive sense. It's the hard one. It's the only part of my hour the agent couldn't do, and it's the part the whole result depended on. The model could generate. It could not prefer. Preferring with reasons is now the work.

Why "learn the tools" is the wrong panic

The common reflex right now is to chase the tools. Learn the prompts, learn the platforms, stack the subscriptions, stay ahead of the feature drops. People treat fluency with the machine as the safe harbor.

It isn't, and here's the mechanism. Tool skill is the abundant thing. The whole point of these systems is that they get easier to operate, not harder. The interface you sweated over this quarter gets a friendlier version next quarter. Anything designed to be accessible cannot also be your moat. You can't build scarcity on top of a thing whose entire purpose is to become common.

The other failure is more subtle. When you can generate forty options for free, the temptation is to ship the first one that clears the bar of acceptable. Volume becomes a substitute for choosing. You produce more and decide less, and the work drifts toward a kind of confident blandness, because nobody upstream actually preferred anything. They just approved what showed up.

That's the trap. The machine removed the cost of making, and a lot of people responded by also removing the cost of thinking. The first was a gift. The second is how you become replaceable by the exact tool you thought was protecting you.

Taste is a muscle, and here is the gym

People talk about taste like it's a gift you're born with. It isn't. It's trained, and it trains the same boring way everything does: high volume of exposure, plus the discipline of saying why.

Rick Rubin is the cleanest example I know. He can't play an instrument. He's produced some of the most important records of the last forty years. His craft is almost entirely judgment: knowing the take that lands when the room is exhausted and every version sounds fine. Decades of listening built an instrument out of his attention.

You build the same thing on purpose. Three habits.

First, raise your exposure. Look at far more good work in your field than you produce. Read the great memos, study the records that lasted, take apart the campaigns that worked. You can't recognize the strong option if your reference set is thin.

Second, force the why. When you pick one of the forty, write the one sentence that says what makes it right and what's wrong with the runner-up. Vague preference is useless. Articulated preference compounds, because next time you can reuse the reason.

Third, keep a stake in the outcome. Taste only sharpens when you live with what you chose. Ship it, watch what happens, feel the miss. Judgment with no consequence stays a hobby. Judgment that costs you something becomes a sense.

"But the machines will learn taste too"

The honest objection: won't the models get judgment as well? They're already trained on human preference. Give it time and the agent picks the right take itself.

Partly, yes. The floor will keep rising. The defensible position is not generic good taste, the kind you could read out of a style guide. That gets absorbed. The defensible position is judgment rooted in something the model has no access to: your specific situation, your customer, your stakes, the thing you're actually trying to make true in the world.

A model can tell you which option is broadly excellent. It cannot know that option three reads as arrogant to the particular people you're trying to reach, or that option seven quietly contradicts a promise you made last year, or that the safe one is exactly wrong because your whole position is built on not being safe. That judgment lives in context the machine doesn't hold.

There's also a harder layer. Taste includes deciding which problem deserves the hour at all. The agent answers the question you bring it. Choosing the right question, the one worth forty options in the first place, stays upstream of anything you can prompt. The machine optimizes the answer. You're still responsible for the question.

The skill underneath the skill

Step back from my afternoon and you can see the shape of the decade. A long stretch where value lived in production, ending. A new stretch where value lives in selection, beginning. Most of the anxiety around all this is really the friction of that handover, people defending a kind of scarcity that quietly stopped being scarce.

The good news is that the surviving skill is the deeply human one. Not output. Discernment. Knowing what's worth doing, which version is true, when to stop. These are the things attention and care were always for, before we spent them all on manufacturing the work by hand.

So the move isn't to out-produce the machines. You can't, and trying makes you a slower version of them. The move is to hand them the making and reclaim the choosing. Let the volume be cheap. Make your judgment expensive. That's the trade that's actually on the table.

My hour wasn't wasted time after a fast result. The hour was the result. The forty options were the cheap part. Knowing which one, and why, was the only part anyone needed me for. That's the work now. At MARSA that's the whole bet: hand the production to the agents, and put real weight back on the human judgment that decides where they point. The part that stays rare is the part worth getting good at.