It did not get it wrong, your brief did

In Part 1 of this series I argued that we read AI wrong. We treat it like a search engine, or like a person with a mind of its own. It is neither. It takes an instruction and runs it. Accept that, and a harder idea arrives close behind.

Most of the time the model did not get it wrong. Your brief did.

This stings for a day and then it frees you. If the output is vague, the request was vague. If the tone is off, you never said what on-brand sounds like. If the pitch is generic, you gave it nothing specific to work with. The thing in front of you is a mirror. That means the fix is in your hands, every time.

Watch the same request twice

Here is a brief I see constantly:

"Write a pitch to journalists about our funding round."

You already know what comes back. A press-release-shaped email. "We are thrilled to announce." A paragraph of adjectives. A quote that says nothing. It reads like every other funding email a reporter deleted this morning. The model did exactly what you asked. You asked for a pitch about a funding round, and that is the average of every funding pitch ever written.

Now the same request, briefed properly:

• Outcome: a 110-word email that makes one fintech reporter in Singapore want to reply today.
• Context it cannot see: we raised $4M led by Insignia to take cross-border payments into Indonesia; this reporter covered remittance fees six weeks ago; our founder ran payments at Grab.
• What good looks like: a real news hook in the first line, one hard number, a clear reason this reporter in particular should care.
• Constraints and format: under 110 words, subject line under eight words, no attachments, plain Singapore English, no adjectives before nouns.

The second email is sharp, specific, and sendable. Same model. Same minute. The only thing that changed was the quality of the instruction.

What a good brief actually contains

A good brief is not longer for the sake of it. It carries four things, and most bad briefs are missing at least two.

• The outcome. Not the task, the result. "Make a reporter reply," not "write an email." The model steers toward the target you name, so name the real one.
• The context it cannot see. This is the big one. The model does not know your funding number, your customer, last week's meeting, or which reporter you have in mind. Everything obvious to you is invisible to it. Say the obvious thing out loud.
• What good looks like. Show it a standard. One example of a line you love. One rule you hold. "No adjectives before nouns" does more than a paragraph of vague taste.
• The constraints and format. Length, structure, tone, what to leave out. These are not fussy. They are the difference between something you can use and something you have to rewrite.

Leave any of these out and the model fills the gap with an average. Averages are exactly what nobody remembers.

You have done this before

Here is the part that caught me off guard. The skill of briefing a model is the same skill as briefing a person.

Think about the last time you handed work to someone and got back something you did not want. The junior who wrote the wrong report. The agency that missed the brand entirely. The freelancer who built the wrong thing. We love to call that a talent problem. Usually it was a specification problem. We held the outcome in our heads, assumed the context was shared, never said what good looked like, and then felt let down.

A model has no ego and infinite patience, so it shows you the gap faster. You can fix a weak brief in ten seconds and run it again. Do that fifty times a week and your instructions to people get sharper too. You start saying the context out loud. You start naming the outcome. The discipline carries straight over.

The good news: this is learnable

Taste feels like something you either have or you do not. Specification is not like that. It is a craft with moves you can practice, and it is the core skill of the next decade of work. This is also why we built Cites X the way we did. Brand strategy, PR, and AI-citation visibility sit in one place because they are one brief. The sharper your spec, the better every layer gets: the clearer your positioning, the better your pitches, the more often the AI answers quote you by name.

So here is the takeaway, and you can use it today. Before you blame the output, reread your input. Then add the four things: the outcome, the context the model cannot see, what good looks like, and the constraints. Run it again. In later parts of this series we will go deeper into each move, starting with the one most people skip: writing down the context you assume everyone already knows.