Stop Hiring the Prompt

The deck was four slides. Each one looked like it came out of a partner’s office at a Big Four consulting firm. Clean headers. A two-by-two matrix on slide three. The recommendation in a callout box at the end: realign dispatch workflows for operational efficiency.

The HVAC owner read it three times. He still didn’t know what to do on Monday.

He promoted his most polished-emails employee to ops manager six months ago. The reports have been beautiful ever since. Beautiful and inert. The east-side ZIP code routing problem is still the east-side ZIP code routing problem. Six months in. Beautiful weekly summaries. Zero operational improvements.

He didn’t hire a manager. He hired a prompt.

The criterion just got commoditized

Here’s the thing. For twenty-five years, an operator hiring and promoting in a service business has had one reliable signal. The artifact. Can this person write a memo that reads tight? Can they put together a deck that holds? Does the proposal come back clean on the first pass?

The artifact was the proxy. It told you the person on the other side could think.

The artifact is finished.

Anyone with a laptop and an afternoon can now produce a memo that reads like a senior partner wrote it. The polished output stopped being the signal because the population that can produce polished output just expanded to include everyone with a browser. Same evidence. Different population producing it.

The signal moved. You didn’t.

Right? That’s the operator move that’s coming. “My new ops manager is killing it on the reports.” Yeah. Of course she is. Anyone is. The reports were never the work.

Most operators reading this haven’t redrawn their evaluation criteria yet. They’re still pattern-matching on the polished memo. They look at the four-slide deck about east-side routing and conclude there’s a thinking human behind it, because that inference has been reliable for two and a half decades. The inference broke. The pattern still fires.

The Generic Output Trap, applied to humans

There’s a failure mode in our Quality Control framework called the Generic Output Trap. We named it for AI deployments where the model produces text that reads competent but isn’t grounded in the situation. The dispatch system says “optimize routing efficiency.” The customer follow-up says “we appreciate your business and look forward to serving you.” It scans as work. It isn’t doing the work.

The diagnosis applies in both directions.

The new ops manager’s deck reads competent. It isn’t grounded in the business. The recommendation, realign dispatch workflows for operational efficiency, could literally be pasted into the deck of any HVAC shop in the country. It would be exactly as actionable for any of them. Which is to say: not at all.

A real diagnosis of the east-side ZIP code routing problem sounds different. “The Bishop crew is taking the 4 PM Greeley calls because dispatch is sequencing by job-complete time, not by drive-home zone, and we’re paying three hours of overtime a week for trucks driving the wrong direction. Switch the sequencing rule and the overtime stops.” That’s the work. The four-slide deck isn’t a smaller version of that. It’s an entirely different category of object. It’s a shape that looks like work, produced by someone who hasn’t done the work.

The Generic Output Trap doesn’t care whether the prompt is a model or a person. The shape is the same. The output scans clean. The diagnosis underneath is missing. The inspection passes because the inspection was never measuring the right thing.

What the operator habit actually is

Here’s the operator habit the artifact era built in. Polished output is proof of thinking. For twenty-five years that has been a workable shortcut, because producing polished output used to be hard. It required the thinking and the writing and the editing and the formatting. The thing on the page was the thing in the head.

Now the thing on the page is cheap. The thing in the head is the same as it ever was. The page no longer tells you which is which.

Most operators have not changed the question they’re asking. They’re still asking “can this person produce the artifact” when the answer has gone to yes by default. The question that’s still hard, and still discriminating, is upstream of the artifact. Did this person hold the situation in their head clearly enough to know what to ask in the first place.

That’s the part the laptop can’t do for you. The model can write the memo. It can’t decide what memo to write. It can produce the four-slide deck. It can’t decide whether dispatch is the right thing to analyze, whether this ZIP code is the right one to start with, what the constraint actually is.

The evidence of human judgment moved upstream of the artifact. The artifact stopped carrying it.

The reasoning question

Here’s the move. In your next hiring conversation, or your next quarterly review, swap one artifact question for a reasoning question.

Don’t ask “show me your best memo from last quarter.” You’ll get a beautiful memo. Of course you will. Anyone has one.

Ask this instead. “Tell me a decision you owned last quarter that didn’t work the way you thought it would. What did you do, what changed your mind, and what would you do differently?”

Three sentences. Or it isn’t there.

What you’re listening for is whether there’s a situation underneath the answer. Whether the person held the constraint, the people involved, what they tried, what came back, where they updated. The artifact is downstream of all of that. The artifact can be faked. The reasoning, walked aloud across three sentences without a laptop in the room, can’t.

If the answer is “I don’t really think of decisions that didn’t work, I try to make decisions that work,” that’s data. If the answer is “yeah, the September quote follow-up redesign. I thought the close rate would lift. It didn’t. I went back and watched five calls and the issue wasn’t the language, it was the timing. We shifted to a 48-hour callback and the close rate moved,” that’s also data. Different data.

You’re not asking for a confession. You’re asking for evidence that there’s a human behind the polished output. Five minutes of reasoning beats an afternoon of memos.

”But what about the people who think well and write rough”

Fair objection. The artifact-era heuristic wasn’t all noise. Some people who produced polished decks did have judgment underneath. The shortcut worked some of the time.

The shortcut also missed a lot. Every operator I know has a story about the person who couldn’t write a clean memo to save their life and ran the warehouse for fifteen years and never missed a Tuesday. The artifact heuristic let those people sit at the wrong altitude for a long time. The reasoning-question swap is also the move that finally lets the people who think well but write rough get the credit they’ve been earning quietly.

So. The downstream benefit is double. You stop hiring people who produce shape but not substance. And you start seeing the people whose substance was real all along and whose shape was never the gating issue.

What this isn’t

I want to be careful here. This is not “AI is making it harder to evaluate human work.” That framing puts the agency in the tool. It indicts the technology. It’s also a comfortable conclusion that lets the operator off the hook. “It’s not me. The tool changed.”

The tool changed. The criterion you’ve been using also changed, by being commoditized. The move that’s owed isn’t defend against AI or detect AI-written work. The move is audit the reasoning, since I can no longer audit the artifact. The agency is on the operator side. The question you ask is the lever.

Does that make sense? You haven’t been hiring badly. You’ve been hiring on a signal that worked for twenty-five years and broke six months ago. The break is structural. The fix is in the question.

The new ops manager and the east-side ZIP

Two roads for the HVAC owner with the polished new ops manager.

Road one. He keeps reading the four-slide decks. The east-side routing problem stays a problem. The reports stay beautiful. Six months from now he has a tenured ops manager who has solved nothing and produced thirty beautiful decks. The bill for that arrangement compounds quietly while the work stays untouched.

Road two. He sits down across the table and asks her about a decision she owned that didn’t work. Listens for the situation. Listens for what changed her mind. If it’s there, he found a real one and the polish is a bonus. If it isn’t there, he learned that six months ago he promoted someone who’s been running on a laptop and an afternoon.

Neither outcome is fun. Both outcomes are useful. The current state, where he doesn’t know which one he hired, is the worst of the three.

The cost of not asking is the cost of not knowing. The reasoning question is the cheapest diagnostic you’ll ever run.

The Monday move

In your next hiring conversation, or your next quarterly review, ask this one question. Tell me a decision you owned last quarter that didn’t work the way you thought it would. What did you do, what changed your mind, and what would you do differently?

Three sentences. Or it isn’t there.

Don’t grade the memos. Grade the question that came before them.

So.

Stop hiring the prompt. Start hiring the question that came before it.

The artifact is finished. The reasoning isn’t. The evidence of human judgment moved upstream. So did the place you have to look for it.

Source influences: Nate B Jones on career evidence in the AI era and the move of human judgment upstream of the artifact. Distilled and operator-translated.

Framework spine: Quality Control, the Generic Output Trap. Read the full framework.