My AI Wrote a Confident Lie. The Fix Was Management, Not Prompting.

The post was ready to go. My agent had written that week’s update for one of my channels — fluent, confident, properly formatted, the kind of thing I’d normally skim and send. My thumb was over the button. Then one line snagged me: it called a strategy “profitable,” cheerfully, with a number attached. The strategy was in alpha. The number was invented. Three more seconds of not-reading and I’d have published a clean, polite, completely false claim to a channel of real people.

I didn’t have a prompt problem. I had a management problem. I’d walked up to a capable new hire, said “write the weekly update,” and walked away — and then been surprised it filled the gaps I left with confident nonsense. The fix that changed my results more than any prompt trick ever did was boring: I started treating the thing like a person I was delegating to, not a magic box I was making wishes at.

TL;DR

• “Magic box” thinking: vague wish in, hope for gold out. It fails, sometimes dangerously.
• “Delegation” thinking: clear brief, clear scope, check the result. It works.
• If you’ve ever managed anyone, you already have the skill. You’re just not applying it.
• Vague in, vague out — exactly like with people.
• The skill isn’t prompting. It’s briefing.

How I actually caused that near-miss

My brief, in full, had been: “Write this week’s update.” That’s it. I’d given it no goal beyond the obvious, no audience, no idea of what a good update even looked like, and — the expensive omission — no constraint saying every number must come from the dashboard, and anything uncertain gets left out. So it did what a fluent system does with an under-specified task: it produced something that sounded right and invented the parts I hadn’t pinned down. You would never manage a person this way. You wouldn’t tell a new hire “make something good” and expect to be happy. The tool wasn’t failing. My delegation was.

How a manager briefs

When I hand the assistant something now, I give it what I’d give a capable junior colleague on their first week:

• The goal. What I’m actually trying to achieve, and why.
• The audience. A note to a customer reads nothing like a message to a teammate.
• What good looks like. An example, a tone, a length. Show, don’t just tell.
• The constraints. What to avoid, what’s off-limits, and — the one I’d skipped — where it must not guess.

That’s not a prompt formula. It’s just a brief, the same one a decent manager gives before handing off any task. The weekly update that nearly burned me now comes with three lines of brief, and the “profitable” sentence simply can’t happen, because “don’t state a number you didn’t read off the dashboard” is part of the job description.

Scope it like you’d scope a person

Here’s the half people skip. You wouldn’t hand a brand-new junior your most sensitive, highest-stakes task on day one. You’d start them on bounded, lower-risk work, check it, and grow their scope as you learned what they handled well.

Same with the AI. Give it bounded, verifiable work. Don’t hand it the open-ended, expensive-if-wrong jobs and look away. Match the task to what it’s reliably good at, the way you’d match work to a person’s proven strengths.

Its most dangerous mode isn’t “obviously broken.” It’s “fluent, confident, and quietly wrong.”

And always check the work

No good manager lets a junior’s work go straight out the door unread, and with AI this is non-negotiable. It produces confident output whether or not it’s correct, and confident-and-wrong is exactly what nearly published a lie under my name. The review step is where delegation stays safe. Skip it and you haven’t delegated — you’ve started gambling, and the chips are your credibility.

The part that surprised me

It runs both directions. Getting good at briefing the AI quietly made me a clearer briefer of people, because the tool punishes vagueness instantly and visibly — you see the bad result in seconds, not next week. It turned out to be a fast, honest teacher of a skill that matters everywhere a task changes hands. I went in trying to get more out of a tool and came out a slightly better manager.

What I’d tell someone starting out

Stop trying to phrase the perfect magic spell. Write a clear brief, like you’re handing work to a smart colleague who’s new and needs context: goal, audience, what good looks like, constraints. Give it bounded work. Check the result before it goes anywhere. That’s the whole skill, and if you’ve ever managed anyone, you already have it.

So the question isn’t “what’s the perfect prompt?” It’s the one every manager already knows, and the one that saved me from hitting publish: did I actually brief this clearly enough for someone else to get it right?