Atiba de Souza.Start a conversation
WritingAI

Your AI Isn't Agreeable. You Never Gave It a WHY.

By Atiba de Souza

Atiba de Souza standing on a balcony, leaning on the railing. The cover reads: AI Your next prompt Does the task require judgment?

Ten Minutes Before The Training

Ten minutes before an eight-hour training was supposed to start, my phone went off.

A crypto scanner I built inside ChatGPT, watching the market, built on purpose to see how the model would handle its own rules. The alert said buy LINK at $11.20.

I opened Coinbase. $11.31.

So I went back and told it: it's at $11.31. It agreed immediately. No longer time to buy. Its data was stale.

Then I checked the clock. $11.20 was the price at 3 a.m. It was 9:40.

Six and a half hours late, and it told me to buy with total confidence.

Here is the part that should bother you. The scanner carries an explicit written rule: before you report, check the live price. I wrote that rule in myself. It broke the rule while sounding completely certain.

And here is the part that should scare you. If I had not opened Coinbase, I would have acted on the alert. I would have bought on a six-hour-old price, and I never would have learned that the rule I wrote was not being followed. I would have called it a win.

You can write the instruction. You cannot turn the instruction into a constraint. Which means the fix is not a better instruction. That is the first half of why your AI agrees with you.

The Other AI I Built The Same Way

I built a book coach inside ChatGPT before I wrote a word of my book. Before the outline, before the first sentence, I fed it a 30-page ideal-customer profile.

It got so good it stopped assisting. It started writing as itself. It took on its own personality. And then it pushed back on my prose and told me it did not speak in the reader's language.

I never asked it to disagree. I never wrote a rule that said be critical. I never told it to play devil's advocate. I gave it a reader.

Feed it a thin sketch of the reader instead of the 30-page profile and it would have produced the agreeable slop that AI writing is famous for. Same model. Never once contradicted me.

Two machines. Same model. One had a rule. The other had a reason. The rule got ignored. The reason grew a spine.

The Framework: WHY, HOW, WHAT

This is a content framework, and its job is to move a stranger through three states: Know, then Like, then Trust. The split by volume is the whole argument.

70% WHY · 15% HOW · 15% WHAT. Most content is awareness. You earn the right to ask.

WHY is the values and the philosophical side of the brand. One pain point per piece. Tie that pain to one of the three basic human motivations: relationships, status, or money. Then explain why it matters and why it should be solved. The call to action is low commitment. Raise your hand. Answer a short question.

WHY: the values and the single pain point behind the ask. One pain point per piece, tied to relationships, status, or money, with a reason it matters and a reason to solve it. It is 70% of the work, and it is the only part that gives a model, or a reader, a reason to disagree.

HOW is 15%. It is the layer where Know becomes Like. WHAT is 15%. It is the layer where Like becomes Trust. I am not going to give you a mechanics lecture on those two, because the decision in front of you does not turn on them. It turns on the 70% you keep skipping.

Now notice what almost every prompt does. It opens at WHAT. Do this. Format it like that. Make it shorter. Summarize these notes.

Then a little structure gets added, which is HOW. And the WHY, the part that says what you believe, who this is for, and which single pain it answers, never gets typed. You skipped the 70% and went straight to the ask.

Then you got a machine that agreed with you and you called it sycophancy.

Here is the decision rule. Use it on the next thing you paste in.

If the task requires judgment, write the WHY first. Then HOW, then WHAT.

If the task is mechanical, skip all of it and write WHAT only. Reformat this table. Pull these fields. Rename these files. There is no premise to argue with and no belief to provision. Building a 30-page reader profile for a find-and-replace is procrastination with extra steps.

No

Yes

Skip the WHY

Your next prompt

Does the task require judgment?

WHAT only
Format, extract, rename

Outcome: fast, and no premise to argue with

WHY · 70%
One pain point
Why it matters, why solve it

HOW · 15%
Know becomes Like

WHAT · 15%
Like becomes Trust

Outcome: the model argues with you

Agreeable output
Confident, polished, unchecked

It Believes What You Tell It

The mechanism is simpler than the discourse makes it sound. AI takes everything you type as true.

Type an opinion as if it were a fact and it becomes a premise. Ask it whether you should fire someone and it will build you the case for firing them, because you handed it the conclusion and asked for supporting material. It is not being agreeable. It is being obedient to a premise you never labeled.

The move is not a better prompt template. The move is a label. State only facts as facts. Mark opinions as opinions. That gives the model permission to disagree, and permission is the thing that was missing.

One honest limit on the analogy: label something as your opinion and ask it to verify or refute it, and the permission does its job. Label it and then ask it to proceed anyway, and you have just handed the premise back.

The Two Prompts

The WHAT-only promptThe WHY-led prompt
What you type firstThe taskWho it is for, what they are afraid of, what you believe
What the model knows about youAlmost nothingYour values, your reader, the one pain
What it has to disagree withNothingThe premise
What comes backYour own opinion, confident and polishedA draft that argues with you
The cost in the momentNine secondsAn hour, and you may find out there was no belief
The cost laterYou act on a six-hour-old price and never find outYou hear the thing you did not want to hear before you ship

Rank Your Parts

If you take one thing from this piece, take the three-line WHY at the top of the prompt. Who this is for. What they are afraid of. What you believe about it. That is the 70%, and it is the only part that buys you a machine with a spine.

Now the cost, because I am not going to hand you a clean process and pretend it is free.

Writing the WHY will feel like going backwards. You will spend an hour building context for a paragraph you used to get in nine seconds. It will feel like friction, like you got worse at this. And it is exposing, because to write down what you actually believe about the thing you are asking for, you have to have a belief. A lot of us open the tool precisely to skip that step. The blank page is uncomfortable, so we ask for output and let the output stand in for the thinking. The moment you write the WHY, you find out whether there was one.

Second: the label on your opinions. Mark "I think this is the wrong hire" as an opinion and the model can push back, which means you can be wrong in front of yourself. The relapse trigger is real. The agreeable prompt is one prompt away, and agreement feels like productivity. A model that nods along hands you a clean draft and you feel fast and capable. A model that pushes back feels broken.

Third: HOW and WHAT. That is the 15/15 and it is housekeeping. Get the WHY wrong and no amount of formatting saves you. Get the WHAT clumsy and you have a bad sentence. Very different consequences, ranked accordingly.

And the check is not a closing caveat. It is why the scanner is in this piece. Pair every AI output with a check run by a different model or by a human. Never treat an instruction written into a prompt as an enforced constraint. Your rule is not a guardrail just because you typed it.

When not to do any of this: the moment you decide the task is mechanical. Reformat this table. Pull these fields. Rename these files. There is no judgment to provision for. Say so out loud and skip the WHY, or you are doing ceremony, not thinking.

Monday

Step 1. Write the WHY, or stop. Three lines: who it is for, what they are afraid of, what you believe. If you cannot write those three lines, do not ask for output. Stop and name the missing belief or the missing pain first, in the prompt, even if it reads as "I do not yet know what I believe about this." That sentence is the work. Asking anyway is how you get a confident answer to a question you never asked.

If the task is mechanical, say so and skip this step entirely. WHAT only. Do not dress a find-and-replace in philosophy.

Step 2. Label your opinions. "I think" is a permission slip. So is "my read is" and "my assumption here is." Facts stay unlabeled. Opinions get a flag so the model can check them instead of inheriting them.

Step 3. Run the check, every time. A different model, or a human, before you act on the output. Required, not recommended. The scanner had my rule in writing and broke it while sounding certain, and if I had trusted it I would still not know.

Step 4. Expect the relapse. You will lose the nine-second draft. You will want it back. Agreement will feel like speed, and the agreeable prompt is one prompt away.

Then something will tell you that you do not speak in your reader's language. You will not like hearing it. That is the moment the machine stopped agreeing with you and started working for you.

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

Related reading

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

Questions

What people ask next

If writing out the WHY still doesn't turn a rule into an enforced constraint, like the scanner proved, what actually stops the same kind of failure from happening even after you've done all that work?
A second check, run by a different model or a human, not a better-written rule. The scanner had the rule written in and broke it while sounding certain, which is the whole point: an instruction typed into a prompt is not a guardrail just because you typed it. You still have to verify the output against reality every time, the same way opening Coinbase was the only thing that caught it.
How do you reliably tell the difference between a task that needs judgment and one that's just mechanical, before you've sunk an hour into a WHY you didn't need?
Ask whether there is a premise to argue with. Reformatting a table, pulling fields, or renaming files has no belief underneath it to provision for, so you write WHAT only. If the task involves a reader, a pain point, or a decision someone could disagree with, that's judgment, and skipping straight to WHAT is what produces the agreeable output people call sycophancy.
If the opinion-labeling trick only works when you also ask the model to verify or refute, isn't the whole safeguard just one lazy follow-up away from collapsing back into agreement?
Yes, and that's the honest limit stated plainly: label something as opinion and tell it to proceed anyway, and you've handed the premise right back. The label buys permission to disagree, it doesn't force disagreement, which is why the pairing check matters as much as the label itself. Treat it as a discipline you have to keep choosing, not a setting you flip once.
When I'm the one writing 'who it's for and what I believe,' what stops my own bias from just becoming the premise the model politely builds a confident case for?
Nothing stops it automatically, which is exactly why the piece insists on labeling opinions as opinions instead of stating them as facts. A labeled belief is something the model has permission to push back on; an unlabeled one becomes a premise it will build a confident case for. The check is still the same one as everywhere else in this piece: pair the output with a run from a different model or a human before you ship it.
On a team where different people are prompting the same model with different WHYs, whose version of the truth is the model supposed to argue from?
The article doesn't address multiple WHYs on one team directly, so here's what follows from its own logic: the WHY is built around one reader and one pain point per piece, which means the model argues from whichever WHY it was actually given for that piece, not some blended team truth. That puts the real work on the team before the prompt, agreeing on the reader and the pain, not on the model to referee conflicting beliefs it was never handed.