Atiba de Souza.Start a conversation
WritingMarketing

Your Team Asks AI to Write. You Taught Them That Was the Whole Job.

By Atiba de Souza

Atiba de Souza standing outdoors in a park-like setting, wearing a white shirt and a cap. The cover reads: MARKETING A request lands What gets delegated

My AI told me my writing was wrong

I built a book coach inside ChatGPT before I wrote a word of my book. Not a prompt I reused. A coach. The first thing I did was feed it a 30 page ideal-customer profile, before I asked it for a single sentence.

I'll tell you what happened. It got so good that it stopped assisting and started writing as itself. It took on its own personality. And then it did something no tool has ever done for me. It pushed back on my prose and told me it did not speak in the reader's language.

Sit with that. The thing I built was arguing with me about my reader.

Now here is the part I could not see at the time. Feed that same coach a thin sketch of the reader instead of the 30 pages, and it would have produced the agreeable slop that AI writing is famous for. It would have agreed with every sentence I wrote. It would never once have contradicted me. I would have called it a good tool and published a worse book.

The 30 pages were not setup. They were the difference between a machine that agrees with me and one that can argue.

Now pull up your marketing team's prompt history.

Write me a LinkedIn post about the launch. Draft an email for the webinar. Give me five subject lines.

Every one of those is a request. Not one of them is a thought.

30 pages of reader profile before the first sentence of the book. One hour to a vetting guide, a legal-coach skill, and a launch checklist once the wiring was fixed. Neither of those came from a better prompt.

The tell is always in the ask

Most people use AI as a vending machine. You punch in a request and take what comes out. AI is designed to be a thinking partner.

A vending machine gives you exactly what you asked for. That is the whole promise. It has no opinion about whether you should have asked for it.

Neither does a model handed a task. It gives you the draft. Competent, agreeable, and indistinguishable from every other team's version of the same request, because no team put anything of their own in front of it.

Here is the part that stings. When the output is flat, your team blames the model, or the prompt, or the tool. The output is flat because they delegated the typing and kept the thinking in their heads, where it never reached the machine. Your team is not bad at prompting. They are handing a thinking partner a typewriter's job, and a typewriter has never had an opinion.

The fear that keeps them asking for drafts

Now the uncomfortable part, and I am on your side here.

Asking AI to write is safe. Asking AI to think is exposed.

Think about what a marketer gives up in the switch. The moment they type "here is our buyer, here is what we believe about their problem, argue with me," they have to have a point of view first. Their judgment is on the table. If the angle is weak, it is not the machine's fault anymore. It is theirs.

Compare that to "write me a draft." If the draft is boring, the tool is boring. Nobody gets blamed. Nobody has to have an opinion at nine in the morning.

Then there is your fear, not theirs. You bought AI for speed. Every version of this change sounds like slow down. If your team spends an hour arguing with a model about an angle, what exactly are you paying for?

You are paying for the judgment. It is what you were always paying for. You just stopped noticing when the drafts got fast enough to hide it. Be honest about your part in this. You celebrated the speed. You asked for more output, not better thinking, because output is the thing you can count on Friday.

And here is what happens when a team hits the wall and nobody names it. A pharmacy owner of seventeen years opening a wellness clinic came to a live workshop I ran. She had already done the hard part. She had written her business brain document. When she came on, she said, "I had my brain, but I'm having trouble making sure that all of my projects know my brain." She was certain the problem was her prompting.

It wasn't. The work was not missing. The wiring was. Her projects had never been pointed at the brain file, so every project started from nothing, and the model met her business for the first time every session.

She was one rewrite away from deciding the brain document had been wasted effort and going back to prompting from scratch. That is the fork your team is standing on right now, and it has nothing to do with whether they are good at prompting.

The framework: delegate thinking, not tasks

Most delegation advice is about handing off the task with better instructions. Write the SOP, hand it off, check the box. That gives you a team that executes, right up until the situation changes and nobody can decide without you.

Delegate thinking, not tasks. Ownership is the destination. You want people who can decide, not people who can follow a step.

Delegating thinking: handing over the judgment, not just the output. You supply the context someone needs to decide, then let them decide and own what comes back. The destination is ownership: people who can decide without you in the room.

The CASE Method is the mechanic under this. This is the worldview that makes the mechanic work, and the spine of that worldview is three things: vision, openness, and asking great questions. If that is how you lead people, it is also how you lead a model. AI is just the newest place it shows up, and it shows up the same way.

Move one, vision. Give it the point of view before you give it the request. The 30 page ideal-customer profile came before the first sentence of the book. Not a better instruction. A richer picture of who the work is for. And vision that lives in one document nobody's project can see is not vision in practice, which means every project gets pointed at that file, every time.

Move two, great questions. Turn the first message from an order into a question. "Write me a post" is a request. "Here is our buyer, here is what we believe about their problem, where is this angle weak?" is a question. Same tool. Different job.

Move three, openness. Let it push back, and answer it. Mine told me my prose did not speak in the reader's language. I did not want to hear it. Answering cost me my own sentences. I rewrote the prose. Openness is not a mood. It is paying for the pushback with a revision.

Now rank them, because they are not equal. Move one is the biggest. Without context there is nothing to think with, and moves two and three have nothing to argue about. Move three is where the compounding lives, because that is the only place judgment actually changes. Move two is the one teams will do badly, because a sharpened request still feels like a question. And the prompt wording your team is pouring energy into? Smallest lever in this piece. Nobody ever wrote a better book from a better sentence in a prompt box.

So pick the leg your team is failing, and decide what you stop doing about it. If the briefs are one line and nobody can name the buyer, you are failing vision. Stop accepting one-line briefs this week. If they are writing beautiful prompts for work that was never pointed at anything, you are failing great questions and chasing the wrong problem. Stop rewriting prompts until you have checked the wiring. If the first draft goes straight to review, you are failing openness. Stop shipping the first output.

The task

The thinking

A request lands

What gets delegated

Vending machine. Output arrives. Nobody owns it.

Context goes in first: buyer, business, point of view

The model pushes back. The team argues. Judgment sharpens.

The team owns the output

That is the fork, not the whole map. The map is at the end.

What it looks like on both sides of the line

The momentInconsistent delegationConsistent delegation
The first message"Write me a post about the launch.""Here is our buyer and what we believe about their problem. Where is this angle weak?"
What goes inA topic and a toneThe reader profile, the business brain, the point of view
The draftArrives finished and goes straight to reviewArrives after an argument and gets sharper in it
Who owns itNobody, which is why it reads like nobody wrote itThe person who argued for the angle
Six months inThe team is faster at producing the same thinkingThe team's judgment has visibly compounded
When strategy changesNew prompts, from scratchUpdate the context, everything re-aims

When not to do this

I would rather you use this well than use it everywhere.

Not every task deserves a thinking partner. Reformatting, summarizing a long thread, cutting one approved message into channel variants. That is typing. Ask for the typing and move on. Delegating thinking on work with no judgment in it is waste, and it teaches your team that the whole idea is theater.

Do not start before you have a written point of view. If nobody on the team can say plainly who the buyer is and what you believe about their problem, do that first. No prompt fixes an empty brief.

And do not roll this out to the whole team on day one. One deliverable. Prove it there.

Monday

First, the wrong Monday, so you can recognize it. Pick the deliverable. Write a sharper prompt. Hand it to the model. Ship the draft. Count how many pieces went out this week. That is "write the SOP, hand it off, check the box" wearing a new shirt, and it is how you ended up with fast output nobody can defend.

Now the real one. One deliverable, not the whole calendar.

1. Pick the piece of content you produce most often. The one everybody quietly agrees is flat. The resistance is that you have to say out loud that the flat thing is yours. That is the cost of admission.

2. Build the ideal-reader profile in the AI, and let it interview you. Do not write it from memory in one sitting. Let the model ask questions and pull the detail out of your team. This is slower than the brief you could have typed in ten minutes, and that is the point, because the ten minute brief is three sentences wide and you have never seen how thin it is. Mine ran 30 pages. Yours will be longer than you expect.

3. Before anyone rewrites a single prompt, check the wiring. Is the project actually pointed at the brain folder, and do the instructions name the file explicitly? This is the step teams skip because it feels like setup. It is setup the way a foundation is setup. The pharmacy owner of seventeen years opening a wellness clinic had written her brain document and still got work that did not sound like her business, because no project had ever been pointed at it. Once the folder was attached and the file named in the instructions, she produced a vetting guide, a legal-coach skill, and a launch checklist inside one hour. Not after weeks of practice. One hour after the wiring.

4. Change the first message from a request to a question. "Here is our buyer, here is what we believe, here is the offer. Where is the weak spot?" Have that argument first. Only then ask for the draft. This is judgment on the table at nine in the morning, for the person who has to type it, in front of you.

Give it one deliverable cycle, then one yes or no question decides whether you continue or go back: can the person who owns that deliverable defend the angle without you in the room? Yes, you widen it to a second deliverable. No, you do not roll it out to the team, because you did not roll it out to a team, you rolled out a faster prompt.

The price

Expect it to feel slower before it feels better. That feeling is not a signal that it is failing. It is the feeling of thinking being visible for the first time.

Expect one person to resist, and expect it to be your fastest draft writer. Speed is the identity that put them in the room. They have been the hero of every deadline, and asking them to slow down and argue with a machine reads as a demotion. Do not leave them to guess. Say the thing out loud: the speed was never the value. The speed was hiding the value, and the value is the judgment. Give that person the first visible win on purpose, and take the volume trophy off the wall while you are at it, because you were the one handing it out.

Expect your own rewiring to be the hardest one. You have to stop praising output in the meeting where they are waiting to hear whether thinking counts.

And here is your floor, so you are not guessing. The compounding does not start after months of practice. It starts the moment the wiring is fixed. The pharmacy owner of seventeen years opening a wellness clinic did not get faster at prompting over weeks. She produced three real assets in the hour after her projects were finally pointed at what she already knew. The work was never the missing piece. The connection was.

context never wired in

pushback refused

Request

Context: buyer, business, point of view

A question, not an order

The model pushes back

The team answers with a revision

Output the team owns

Agreeable slop nobody owns

Monday morning, one deliverable, and one question answered after: can the person who owns it defend the angle without you in the room?

If the answer is no, you never had delegation. You had typing, and it was fast.

Pick the deliverable today. Then stop accepting one-line briefs.

<!-- cluster-links:begin -->

Related reading

<!-- cluster-links:end -->

Questions

What people ask next

Okay, so the context document is the big lever — but who actually owns keeping that thing updated as the business or the reader changes, and what stops it from going stale the same way the prompts did?
Whoever owns the judgment owns the document — it's not a shared file that updates itself any more than a brief writes itself. It goes stale the exact same way the prompting did: nobody checks the wiring. The fix isn't a new tool, it's the same discipline as move three — treat the document as something to interrogate and revise when the reader or business shifts, not something you write once and assume still points at the truth.
If the model pushing back is the whole point, how do you tell when it's actually right about my writing versus when it's just being contrarian and I should hold my ground?
The test isn't whether the pushback feels right, it's whether it's grounded in the context you actually gave it. A thin reader profile produces agreeable slop; a rich one produces pushback tied to specifics about the reader you defined, and that kind is worth a revision. If the objection can't point back to something concrete you fed it, it's not openness working, it's noise, and that's when you hold your ground.
This sounds great for one person arguing with their own model, but how does it work across a whole team — do I really expect everyone to spend an hour debating an angle before every piece of work?
No — the hour happens once, at the wiring stage, not on every piece of work. The vetting guide, the legal-coach skill, and the launch checklist all came together in an hour after the wiring was fixed, not after an hour of debate each time. Once the shared context is built and every project is pointed at it, individual exchanges get shorter because the biggest lever, vision, has already been paid for.
How do I diagnose which of the three failures my team actually has, since a flat draft could look like missing vision or a lazy question or just nobody willing to push back?
Look at what's missing, not at the draft itself. If the briefs are one line and nobody can name the buyer, that's vision. If people are writing beautiful prompts for work that was never pointed at a context file, that's great questions, and no amount of prompt-polishing will fix it. If the first draft goes straight to review with no argument in between, that's openness — and each one has a different stop-doing, not a different prompting trick.
Practically, how do I get people past the fear of putting their judgment on the table when asking for a draft has always felt safer and faster?
Stop rewarding the thing that makes the fear rational — the fast, agreeable draft you can count on Friday. Stop accepting one-line briefs and stop letting first drafts go straight to review; when those two things are off the table, people have no safe way left to avoid having a point of view. The speed doesn't disappear, it just shows up after the judgment instead of instead of it.