My AI Agent Did Exactly What I Asked

The site you are reading was built by an AI agent. I wrote about why I stopped renting my website. This is about the part that almost went wrong.

My first direction to the agent was three words: premium keynote speaker.

It did exactly that. The page that came back was polished and confident and a little bro-y. Everything I asked for was there. None of it sounded like me.

It built what I described

My first instinct was to blame the tool. Then I read my own instruction again.

“Premium keynote speaker” describes a category. It says nothing about the person, the reader, or how the page should feel when you leave it. A human designer would have asked me those questions over coffee. The agent did not ask. It filled the gap with the most likely answer, and the most likely answer for “premium keynote speaker” is the guy on the stage with the headset mic.

It built what I described. I just hadn’t described me.

So I described me

The second brief was different in kind, not just in length:

Magazine-level editorial, not corporate. A voice somewhere between Obama, James Baldwin and Anthony Bourdain.

Three names did more work than three paragraphs could have. Obama for the calm. Baldwin for the honesty. Bourdain for the curiosity, and for refusing to talk down to anyone. Same agent. The next version was the one I approved, and it is the one you are looking at.

A no fixes one line. A why fixes the rest.

Partway through, the agent offered a tagline: Ordinary people can do extraordinary things.

It sounds kind. It isn’t. It quietly tells the reader they are ordinary, and it puts me above them.

So I said no, and I said why. That second part mattered more than I expected. The agent did not just drop that line. It stopped writing lines like it. A no corrects one sentence. A reason corrects every sentence after it.

This was never really about AI

For most of my career, vague instructions have survived because good people filled the gaps. They asked a follow-up. They read the room. They knew what you meant.

I learned to stop relying on that years ago, working with developers on the other side of the world. When the team wakes up as you go to sleep, a vague message costs a full day. You learn to write the brief that answers the follow-up before anyone asks it.

That turns out to be the exact skill agents need. They just shrink the loop from a day to a few seconds.

What I put in a brief now

The goal behind the task. Not “build a homepage.” Who it is for and what they should do next.
What good looks like. A reference, a voice, a reader. Names work better than adjectives.
The hard lines. What it must never do.
The why behind every no. So the next answer is better too.

Prompting is just management on fast-forward. The people who get the most out of these tools will not be the ones with the cleverest prompts. They will be the ones who already knew how to brief a team, and now get the answer back before lunch.

What did you ask for this year, from a person or a machine, that came back exactly as written and nothing like what you meant? Reply to the newsletter and tell me.