You're not bad at AI. You're bad at managing.
Your brand-voice agent isn't underperforming. It's untrained.
A coach DM’d me last week. She wanted me to know AI was a waste of time.
She’d tried it. Twice, actually. Once in February. Once in April. Both times the output came back beige. Both times she screenshotted the worst bit and posted it as a warning to her audience. The screenshots got better numbers than her actual writing, which she also found funny.
The thing she’d typed was: write me a LinkedIn post about leadership in my voice.
Twelve words. No samples. No rules. No definition of what good looked like. Then a verdict, delivered with the confidence of someone who has just fired a junior on his first morning.
That is the diagnosis I keep running into. I want to walk you through it this week.
Training is what you skipped.
Imagine the same coach hires a new junior writer. He arrives Monday. He’s been a stranger for forty minutes when she asks him to write a LinkedIn post about leadership in her voice. He goes away. He comes back with something acceptable but not hers. She files it as a hiring mistake and starts again.
You wouldn’t do that. Nobody would do that.
You’d hand the junior fifteen of your best posts. You’d tell him the things you’d never say (no revenue numbers, no “I help people”, never opens with a stat). You’d point at the three writers whose tone you wish was yours. You’d send him away for an afternoon and read what he came back with. You’d mark it up. You’d send him away again.
By week six he’d be writing things you could ship. By month four he’d be writing things you couldn’t tell from your own first drafts.
That is what training is. That is what training has always been, in any era, with any tool.
A model is a junior who can read faster.
The five things you didn’t give it.
(This is where you wince. Or, in our experience, where you go quiet for a moment and reach for a notebook.)
A model that produces voice-aligned, ship-ready writing needs five things from you. Skip any one and it produces beige.
One.
A role. Not “you are an AI assistant”. A narrow, specific identity: you are the copy chief for The Ghost. Your job is to protect the voice rules and catch any draft that breaks them before it ships. The narrower, the better.
Two.
Twelve rules. Testable. Hooks under twelve words. No em-dashes anywhere. Never opens with a stat. Identity-claim closes only. Each rule a thing a person could mark on a draft with a red pen.
Three.
A short description of what good looks like, in observable behaviour. Not “make it feel sharp”. Not “sound confident”. Contains one short-short-short-long-short sentence sequence per post. Closes with a fragment more than half the time. Uses one parenthetical aside per post, never two. Things a person could count.
Four.
Sixteen samples. The actual artefacts. Real posts. Real emails. The voice you mean, not the adjacent voices that sound similar. Three is not enough. The model averages across them and hands you something halfway between all three. Sixteen is the floor.
Five.
Five failure samples, annotated. The bad ones. The drafts you binned. With one line on each saying why it failed. The model learns the edge of the cliff faster from the bad than the good.
That is the brief. Anything less is the equivalent of asking the junior to write in your voice on his first morning.
Three is not enough.
The bit nobody wants to hear.
The number to run is a hundred. A hundred cycles. Hand the model a real task. Read what it comes back with. Give it specific feedback. (”Option 2 broke Rule 4. The hook was fourteen words. Rewrite.”) Send it back. Repeat.
Most people stop at three.
The output at iteration three is the model’s floor. The ceiling is at iteration one hundred. Everything in between is the part where it stops being slop and starts being usable. By iteration twenty, one in three pieces is ship-ready. By iteration fifty, four in five are. By iteration one hundred, the voice is so closely matched you’d ship without editing.
Time to a human reaching the same standard from a cold start: eighteen months, minimum, and only if the manager is good. Which is the next problem.
Time to the agent: six weeks of focused looping.
The ratio is the whole point. You are not racing the tool. You are racing the operators who are running the loop while you are giving up at iteration three.
Now the bit you didn’t want to read.
(Last fourth-wall break. I promise.)
The reason most people can’t train the model is older than the model. They have never been forced to define what good looks like in observable terms. They use words like sharper and more confident and make it pop. Those aren’t briefs. They’re moods.
The junior nods, leaves, guesses, fails. The model nods, leaves, guesses, fails. Same physics. Different speed.
The weakness is managerial. The tooling is the most generous student you have ever had. It will run the loop a hundred times for the price of a coffee. It will not complain when you ask for revision sixty-three. It will absorb every rule you write down. It cannot read your mind, and it cannot make up for the part of you that has never been forced to articulate what good looks like.
If you can fix that, the model is the easiest hire of your life. If you can’t, it will keep handing you slop, and you’ll keep posting the screenshots as warnings.
What stays on your side of the line.
Two things, properly held, stay yours forever.
Taste. Which bit earned the parenthetical. Which line will make the right client wince. Which week’s post needs the joke and which week’s needs the diagnosis. Taste is the thing the model can recognise but not originate. You are the chef. The model is the apprentice.
Risk. The willingness to put your name on the line and post the thing that will make some people in the comments unhappy. Risk is what humans get paid for in a world where intelligence and labour cost less every quarter. Everything else gets cheaper. The risk you take with your own voice is the asset.
Train the model. Keep the taste. Take the risk.
That is the work.
If your voice agent is producing slop and you’d like a second pair of eyes on the role, the rules, and the samples you didn’t write down, my Voice Audit is the door.
See you next Tuesday.
Sarra - The Ghost