The Best AI Users Write the Shortest Prompts

The people selling you prompt engineering don’t use the thing. Ask them. Whoever tells you otherwise probably doesn’t want you to know that.

The market already called it. Job posts titled “prompt engineer” are down about 40% from their 2023 peak. The six-figure job everyone was promised is gone. Meanwhile postings that list prompting as a skill have grown roughly 250%. The title died. The skill moved into everybody’s job, and most people are still doing it wrong.

I have spent more than three years in these tools every single day. Here is what I wish someone had told me on day one.

I.

Your AI is a new hire.

Think back to your first week at a new job. Your boss walks up and tells you exactly what to do, down to the keystroke. Did you feel empowered? Or controlled, boxed in, and a lot less likely to create anything?

That’s what you’re doing to your AI.

You would never hand a new employee a task and walk away. You spend time with them. They get to know you, you get to know them, and they learn how you think. You give them context, you let them try, you tell them where they screwed up, and you do it again tomorrow.

That’s the whole job. Even Shopify’s CEO says the real skill is context: getting the right information in front of the model so the problem is solvable at all. Context is a relationship. You don’t write it once.

II.

Lazy prompting is a feature.

Every time you over-specify, you cap the output at your own imagination. And your way was never guaranteed to be the best way. You just haven’t seen the alternatives yet.

A loose prompt gives the machine room to be creative about how the problem gets solved. Let it test three approaches. Let it argue with you. Hand it the goal, get out of the way, and make it bring you the solution instead of a question.

The people who get the most out of AI sound almost lazy when you read their prompts. They aren’t. They’ve done the work upstream, so the instruction can be short.

III.

Tight prompting earns its place exactly once.

Precision matters at the agent level. Once you know what works, you write it down and wire it into a flow that runs on a schedule. That’s where the careful instructions belong.

Here is the test I use. If I can’t explain what good looks like, I’m not ready to write the instructions. I hand the AI the problem, let it run, and study what comes back. When I see the same good answer three times in a row, that’s when I write it down. The prompt comes last. Most people write it first and wonder why the results feel like a form letter.

Before that, you’re micromanaging something smarter than your instructions. And even after it, set-and-forget is a fantasy. The goal is a system that keeps improving itself: run it, read what it produced, give it feedback, run it again. The loop never ends, and the more time you spend inside it, the faster you get.

IV.

What a real handoff looks like.

It took me six months to hand off my process. Three of those months were nothing but training and retraining, plus a few big jumps in the technology along the way.

I now run an org chart of more than a hundred agents. Every one has a name, a job and a boss. Atlas is my advisor; I give it a goal and it turns that goal into a number and a date. Flick triages the inbox. Fox does the work that comes out of it. Pace pulls the Friday numbers. Bridget keeps the interview calendar full. Every Saturday, Probe goes looking for data nobody is collecting yet and hands the best ideas to the right company on Sunday.

They run on three rules. Act by default. If it’s new, act and then tell me. And anything that moves money, signs a contract or touches a password gets prepared to one click, and a human presses it.

I manage the agents. A person manages the people. Nobody gets confused about which is which.

The same goes for the hiring market. The companies winning with this hire people who can run a system, read the output with a critical eye and say, “That’s wrong, do it again.” That’s a manager’s skill. It always was.

V.

AI rewards people with opinions.

Here’s the part nobody puts in the tutorials. The tools reward the people who have a point of view, who add something to the conversation and who are loud enough that they can’t be ignored. Bland in, bland out.

Sometimes having an opinion means being wrong in public. Do it anyway. The machine can generate a thousand average answers. Your judgment is the only thing it can’t produce on its own.

There’s no shortcut around the reps.

Opinion pieces in Bellwether are the view of the writer. Disagree? Write to us.

Bellwether’s editor owns Fracs.