Learn

AI basics#1

The first move.

Make the AI interview you before it does the work.

Brian Stewart1,200+ ad accountsfounder and operator

The prompt

I want you to [task]. Before you start, ask me whatever you need to do this really well.

Get the 9-page PDF: the interview taken apart, a worked ad brief, and the have-ready sheet.

The guide, straight to your inbox, with your skills library link.

What an AI Ad Operator does

You direct. AI executes. That is the whole job, and it is the opposite of how most people use these tools: type a request, take whatever comes back, and decide the AI is average. The output was average because the request was.

This is the first move of the basics: the habit that makes every other AI-for-ads move work. It takes one sentence, it costs nothing, and it changes what comes back from the same model on the same task. If you have only ever chatted with ChatGPT, start here. If you already run ads with AI, check that you are actually doing it, because most people skip it.

By the end of this guide, twenty minutes from now, you will have four things:

  1. The sentence, in muscle memory.
    You will stop typing a request without it.
  2. Your brief.
    Five answers about your business, saved as a file, ready to paste into every ad task from now on.
  3. Three hooks you could run this week.
    Built from your facts, with the second pass most people skip already done.
  4. A way to tell good AI output from average.
    Once you have seen the interviewed version next to the thin one, you cannot unsee it.

Twenty minutes, with the tool open

Read it once. Then open Claude, ChatGPT or Gemini beside it and run the ad prompt in “Run it on an ad” on your own business. The have-ready sheet at the back is the part you keep.

Add one sentence before any real task

The prompt

I want you to [task]. Before you start, ask me whatever you need to do this really well.

It stops guessing and starts asking. Instead of an answer you get four or five questions: who this is for, what you are offering, what you can prove, what has worked before, what it must not say. Answer them, and then it does the task. Same model, same task, different output, because now it has the brief.

That is all the move is. The rest of this guide is what to do with the questions, because the quality of the answers you give is the quality of the ad you get back.

  1. Write the task in one line.
    Plain words. “Write three Meta ad hooks for my meal-prep delivery service.” Not the whole backstory; the interview is where the backstory goes.
  2. Add the sentence.
    Word for word. “Ask me whatever you need” is the part that matters. “To do this really well” tells it the bar.
  3. Answer in bullet points.
    Short is fine. Specific matters more than polished. It is the brief a good copywriter would ask for, and you would not write that in paragraphs either.
  4. Let it do the task. Then keep going.
    The first answer is not the one you use. “Then run it twice” is the second pass, and it is where the good version comes from.

Questions, not an answer. That is the point.

The first reply looks something like: “Before I write these, a few questions. 1. Who is this for, specifically? 2. What exactly are you offering, and at what price? 3. What proof do you have that it works? 4. Do you have an ad or post that performed well that I should learn from? 5. Is there anything I must not say or claim?” Answer in the same numbered order. If it asks something you cannot answer, say so. A gap you name is better than one it fills.

An AI answers the question it was given

Give it a thin question and it fills the gaps with averages: the average audience, the average offer, the average ad. It is not being lazy. It has nothing else to use. Every gap in your request gets filled with the most common thing anyone in your category would say, which is exactly the thing your customers have already scrolled past a hundred times.

The interview replaces the averages with your numbers. Your customer instead of a demographic, your price instead of “affordable”, your result instead of “proven”. The model was always capable of the specific version. It had no way to get there from what you typed.

Thin promptWith the interview
“Write 3 ad hooks for my meal-prep service.” Done.The same line, plus the sentence. It asks five questions before it writes a word.
Who it writes for: “busy professionals”Who it writes for: the person you described. In “Run it on an ad”, a nurse on 12-hour shifts who orders takeout four nights a week.
The claim: healthy, delicious, convenientThe claim: $9 a meal against the $18 she spends on takeout, 45 grams of protein, first week half off.
The proof: none, or inventedThe proof: yours, because you gave it, and only yours.
The hooks: the same three every business in your category getsThe hooks: built on your enemy, your number and your offer.

One caveat, and it is the important one. The interview only works if you answer it truthfully and specifically. A vague answer to a good question gets you the average back, with extra steps.

Run both, once

Type the thin version of your task and keep what comes back. Then type the same task with the sentence, answer the questions, and keep that too. Put them side by side. The difference is the whole argument for this guide, and you will never skip the sentence again.

What it will ask, and what to have ready

The questions vary a little by model and by task, but for an ad they come down to five. Here is what each one is really asking, what a good answer looks like, and the weak answer most people give.

It will askA good answerThe weak answer
Who the ad is forOne specific person: their situation, what they do now instead of buying from you, what they would say the problem is.A demographic. “Women 25 to 45.” The model cannot write to an age range. It can write to a nurse coming off a night shift.
The offerOne line, with the price if there is one, and what exactly they get.The category. “Healthy meals delivered.” That describes your competitors too.
The proofA number, a name, a result. Your price against the alternative, a customer’s words, a before and after you can show.Adjectives. “High quality, trusted, proven.” If you have no proof, say so. It will write around the gap instead of inventing one.
What’s worked beforeYour best hook or ad, pasted in whole, with a line on why you think it worked.Nothing. Then the model starts from zero, and so do you.
What it must not sayClaims you cannot make, words you do not use, anything your industry regulates, anything you promised a customer you would never do.“Nothing, go wild.” Then you spend the second pass deleting things instead of improving them.

What is the click for?

Where does the ad send people, and what counts as a result: a purchase, a lead, a message? If it doesn’t ask, tell it anyway. A hook for a $9 trial and a hook for a $900 program are different sentences. Part 1 of the Meta Ads playbook is about deciding this before you spend a dollar.

Don’t use the first answer

When it gives you something good, don’t use it yet. The first answer is the model’s best average of your brief. The second pass is where it gets specific about its own work. Ask:

The second pass

What would make this better?

Then run it again. Two passes beat one prompt, every time. Most people stop at the first answer, which is why most AI-written ads read like AI-written ads. Three more follow-ups that earn their keep. Use one, or all of them:

  1. “What did you assume that I didn’t tell you?”
    It will list its guesses. Correct the wrong ones. This is the fastest way to find the gap in your own brief.
  2. “Which of these would you cut, and why?”
    It is a better critic than it is a writer. Make it rank its own work before you do.
  3. “Rewrite the strongest one three ways: shorter, number first, as a question.”
    Now you have versions with one thing different in each, which is what a real test needs.
First passSecond pass
“Coming off a 12-hour shift, the last thing you want to do is cook. So don’t.”“Off a 12-hour shift. Dinner’s already at your door. $9, not $18.”

Shorter, the number in it, the enemy named. Same brief, one more question.

It drafts. You decide.

The AI does not choose the hook, set the budget or publish the ad. You pick what goes live, and you run it as a real test: one thing different per version, a decision rule written down before launch, nothing edited mid-flight. Part 7 of the Meta Ads playbook has the test checklist.

Five ways this goes wrong

  1. Answering with a demographic.
    “Women 25 to 45” is not a person. Describe one customer you could name.
  2. Skipping the proof.
    Leave the proof question thin and you get adjectives. Leave it out and some models fill it in, which is worse: now there is a number in your ad that isn’t true.
  3. Pasting nothing that worked.
    You already ran the test. Your best post is data. Give it the data.
  4. Taking the first answer.
    One pass gets you the average of your brief. The second pass is the whole point of “Then run it twice”.
  5. Letting it decide.
    It drafts hooks. It does not know your margin, your refunds or your regulator. Every claim gets read by a person before it runs. That is the operator’s job, and it does not go away because the drafting got fast.

Keep the brief

The five answers you gave are not a one-time thing. They are your brief, and every ad task from now on starts by pasting it in. Save them as a file: brief.txt, a note on your phone, anything. Update it when the offer changes, or when a new review says it better than you did.

That file is the beginning of a system. Turned into a skill, the same brief is something your agent reads on its own, so you stop pasting and start directing. But the file comes first, and it comes from this interview.

Point the same move at the account, not the ad

Once a week: “I want you to review last week’s ad results and tell me what to change. Before you start, ask me whatever you need to do this really well.” It will ask for the numbers that matter. Paste them in. That is the weekly review in Part 9 of the Meta Ads playbook, with the AI reading and drafting and you deciding, which is exactly how the accounts I run work.

The have-ready sheet

  • Who it’s for
    the specific person, not the demographic
  • The offer
    one line, with the price
  • The proof
    a number, a name, a result
  • What’s worked
    your best hook or ad, pasted in whole
  • What it must not say
    claims, words, rules
  • What the click is for
    purchase, lead or message, and where it goes

The first move, in seven lines

  • The task in one line, then the sentence: “Before you start, ask me whatever you need to do this really well.”
  • Answer every question it asks, in bullet points. Specific beats polished.
  • Every number in the brief is yours. If one appears that you did not give it, cut it.
  • “What would make this better?” Then run it again.
  • Versions with one thing different in each, and a decision rule written down before launch.
  • A person reads every claim before it runs. It drafts; you decide.
  • Save the brief. The next task starts by pasting it in.

Next reads

Brian Stewart at his desk on a video call, explaining with both hands
September 1, 2026

This is move one.

Personal AI Agent

I set up a personal AI agent live, start to finish, on standard off‑the‑shelf tools.

When
Tuesday, October 20, 2026, 6–8 pm CT
Length
About 2 hours, on Riverside
Price
$47, one time
Replay
Included, by email within 48 hours, and kept on brianwith.ai for 12 months.