The Bottleneck Was Never the Prompt

Cassie Kozyrkov ran a thought experiment in February that I haven't been able to put down. Suppose the models get perfect. Instant, correct output for any request you can make, in any field, forever. What would still be worth learning?

Her answer arrives in the first line of the essay:

"What good is a perfect answer if you don't understand the question?"

It's a better question than most of what gets written about AI, because it doesn't depend on which model is winning this month.

The filter that dissolved

Her sharpest observation is about what difficulty used to do for us:

"The difficulty of getting the answer was often a filter that ensured only those who understood the problem deeply would arrive at a solution."

That describes every credential I've ever earned. The work was the screening. Take the friction away and the screen goes with it — "the more our tools do for us," she writes, "the less we seem to understand about what we're really asking them to do."

She calls this the paradox of progress. It shows up in a plainer way at the keyboard. Watch two people use the same chatbot for a week and the transcripts look remarkably similar: short requests, plain language, nothing that would impress anyone. One of them is getting excellent work out of it and the other is getting oatmeal, and you cannot tell which is which by reading what they typed.

The gap isn't in the words. It's in what got attached, what got specified, and what got checked afterward — none of which leaves a mark in the text box. That's the filter dissolving, viewed from ground level.

Thoughtlessness, enabled

Her name for what these tools actually do is the best phrase in the piece:

"the peril of AI is the same as the promise of AI: it's a thoughtlessness enabler."

She uses the more familiar framing too — the tools are amplifiers, and "they also amplify our mistakes if we're not careful" — but thoughtlessness enabler is more useful, because it names the mechanism instead of the outcome. The machine doesn't make you worse at thinking. It removes the step where not thinking used to cost you something.

Which is why the most valuable habits with these tools are the ones that put a little friction back. Sending a second message instead of accepting the first answer. Handing over the actual document instead of describing it from memory. Asking for the strongest case against your own draft before anyone else gets to make it. None of that is a prompting technique. It's the deliberate reintroduction of the step that difficulty used to force on you for free.

The spotlight moves

The line I'd tape above a desk:

"Once the tools are no longer the bottleneck, the spotlight's on you."

Leaders used to be able to think slowly, she notes, because the analysis took the team months. Now the analysis takes ninety seconds — and the ninety seconds after that are yours, and they're the entire job. What to ask. What to hand off. What to verify before it goes out with your name on it. What to keep doing yourself because judgment, or joy, is the point of it.

She also demotes prompting in a way I'd sign twice: "Leaders, prompting is not new. It's an ancient skill." Knowing what you need done, specifying the parts you actually care about, anticipating how it goes wrong, checking the work when it lands on your desk.

Read that description again and notice there's no keyboard anywhere in it. Every clause is something you've already done to human beings — briefed a colleague, coached a contractor, left instructions for a substitute. Which is also why prompt packs age like milk. They're selling packaging on a skill you learned managing your first project.

Where I'd push back, gently

One place I part company. Near the end she writes:

"I don't see anyone building the tools that would truly elevate a leader's cognitive abilities enough to keep up. (I would love to be proved wrong!)"

Her point is that we're pouring everything into building better genies and nothing into the wisher.

I'd argue the wisher-side tools already exist. They're just not products, so nobody made a launch video for them.

Make the machine interview you before it answers: ask me the three questions that would most improve your answer, then wait. Approve the outline before you pay for the prose. Give it the real file. Assign it the job of arguing against you, in writing, before you decide. Sort tomorrow's list by who should actually own each item — you, the machine, a colleague, or nobody, because it shouldn't be done at all.

Every one of those forces a decision you'd otherwise skip, and every one runs on a subscription you're already paying for. Kozyrkov is right that no vendor is shipping this. I'd say that's fine, because judgment has never once been improved by software. It gets improved by structure, and structure is something you can impose on a Tuesday afternoon for nothing.

That's a friendly amendment, not a disagreement. Her diagnosis is the part worth carrying around: the bottleneck moved. It's sitting in our chair now. Pretending otherwise is how you end up with a technically perfect answer to a question you never should have asked.

So try this with the next real question you were about to type. Before you send it, add one sentence: Before you answer, ask me the three questions that would most improve your answer. Then wait. Answer all three. Compare what comes back to what you'd have gotten.

That gap is the whole argument. And you won't be able to un-see it.

Quotations are from Cassie Kozyrkov, "What's the Most Valuable Skill for the AI Era?", Decision Intelligence, February 12, 2026.