Judgment Over Prompts

I keep coming back to a job posting a friend of mine rewrote last spring, because it tells you something about how companies are hiring for AI that most of them don't yet understand.

She runs a small analytics team at a mid-sized services firm in Bengaluru. The listing she inherited, like a lot of 2025 listings, asked for "prompt engineering experience" as a preferred qualification. Her team had actually hired for that twice.

The first hire was a thirty-year-old who'd spent four years doing content operations at a SaaS company and, on the side, had become the kind of person who could get Claude to write a structured indemnity clause in three prompts when most people needed twenty. He was noticeably better than the rest of the team for about two months. Then the model got an update, and most of what made him special stopped being special. Not because he got worse — he didn't — but because the gap between his work and everyone else's collapsed to almost nothing. By the time he left, eighteen months later, for a job that paid more, nobody on the team could really remember what he had been specifically good at.

The second hire, made about six months after him, was the more interesting case. She'd spent six years as a senior associate at a consulting firm, where her job had been to mark up junior consultants' decks before client meetings. No prompt engineering background. By month three, she was producing output the rest of the team quietly preferred, and not because her prompts were cleverer — they mostly weren't — but because she knew, faster than anyone else, which sentences in a draft were confidently wrong.

I asked my friend, when she told me this, what she'd actually learned from the two hires, because the obvious conclusion isn't the interesting one. The obvious conclusion is "prompting is a commodity skill, hire for something else." That part's true, but it's too neat. The thing she actually learned was narrower, and more useful: the gap between good and bad work in her team wasn't about how people asked questions. It was about how fast they could tell a fluent answer from a wrong one.

That's a different skill, and it isn't new. It's the same skill a good editor uses on a junior reporter's draft, or a senior lawyer uses on an associate's memo, or a trader uses on an analyst's model. The interesting thing about AI is that it made this skill newly visible, and newly measurable. You could see, in real time, who on the team was spotting the wrong number, the misused term, the plausible-sounding claim that was about to embarrass someone in front of a client. The person who spotted those things was the one you wanted more of. The person who wrote the most elegant prompt and missed the wrong number was, increasingly, a liability.

So my friend rewrote the posting. The line about prompt engineering came out. In its place she put a single sentence she'd been drafting in her head for about three weeks: "We're looking for someone who can read a confident answer and tell us, specifically, what's wrong with it." She says it was the first time she'd written a line on a job posting that made candidates visibly pause in the interview.

The interview itself changed too. They used to ask candidates to write a prompt for a realistic task and score the structure. That exercise now lives in a Google Doc nobody's opened in a year. The replacement is blunt: the candidate gets a piece of model output on a realistic client problem — a draft memo, a financial summary, a one-pager — and the interviewer asks one question. "You've got ten minutes. Put your name on this and tell me what, if anything, you'd change before it goes to the client." There's no prompt to write. There's only the output, and the candidate's judgment of it. The ones who do well are, almost without exception, the ones who have spent their careers looking at other people's work and telling them what's wrong with it. Editors. Lawyers. Senior analysts. People who have caught a lot of bad drafts in their time and developed, sometimes without noticing, the reflex of mistrusting a fluent answer.

The onboarding change made the largest difference, and it was the most boring one. They paired every new hire with a senior reviewer for their first month, and the senior reviewer's only job was to narrate, out loud, why they were accepting or rejecting each piece of model output. Not "here's a good prompt template." Just: "I'm rejecting this number because I know the client's Q3 was softer than this implies, and I can tell from the phrasing the model doesn't know that." Slow, expensive in senior time, and, six months in, the highest-return training investment the team has made.

The new hires who went through this loop were competent, in the loose sense, on AI-assisted work in about six weeks. The new hires who only got the old prompting tutorial took closer to four months. The team is small enough that the difference isn't statistically clean, but the senior reviewers are pretty sure they know what's going on. The tutorial taught the new hire a vocabulary. The narration taught them a habit.

I think this generalises beyond one team, and I think most companies are doing the opposite of what my friend figured out. They're still posting "AI fluency" as a hiring line. They're spending their training budgets on tool tutorials. Their performance reviews credit people for the number of AI-assisted tasks they ship, not for the number of quietly wrong claims they catch. All of that rewards the skill that ages out fastest in the building. The habit of reading carefully is the same habit it was in 1995 or 2005. AI didn't invent the need for it. It just made the cost of not having it more visible, more quickly, than the cloud, or the spreadsheet before it, ever quite managed.

I'm conscious this is a sample size of one team, one friend, and one year of observation. But every other services team I talk to tells a version of the same story, and almost none of them have changed their hiring bar yet. The supply of people who can read a confident answer and tell you specifically what's wrong with it is, conveniently, much larger than the current job market suggests. They're just not on the AI-fluency listings. My friend has stopped waiting for them to be.