My LLM is not a yes-man. And that is deliberate.
It talked me out of my new business in the middle of a one-to-one coaching session.
Vienna, a one-to-one coaching session. Next to me, a woman in her mid-fifties.
She wants to see what one of these assistants can really do. So I show her.
To this day I do not know why that particular idea came to me. Maybe because it sounded harmless.
‘I want to start a dog hotel’, I type. ‘As a new business.’
And then, because it felt right in that moment, one more sentence.
‘This could be really big.’
I expected a list. Locations, first steps, maybe a name.
What came back was not a list.
In essence: you do not have a dog. You have a pile of unfinished work. Forget the dog hotel and get back to it.
She still tells that story today.
And that is exactly why I am telling it here. Most people expect a willing servant that does whatever it is asked. What she saw that day was the opposite. And it was no accident.
A quick word on terminology, because that is where the difference sits. What most people call AI is a language model, an LLM. AI is the umbrella term for everything from image recognition to robotics. The LLM is the one machine that answered on my screen that day. And how it answers is not a property. It is a setting.
That was not a mood, that was the system prompt
Artificial intelligence does not have moods. It did not have a bad morning that day. It did exactly what I had told it once, in advance: be my sparring partner, not my concierge. Hold me to my goals. And when I get sidetracked, say so.
That standing instruction is called a system prompt. It is the role that runs in the background of every single conversation without me typing it in again each time. I write it once, and from then on the assistant behaves accordingly. With the dog hotel it did not fail. It worked.
I learned elsewhere just how much a single sentence like that matters: my AI agents had no memory, because their standing instruction simply never said to save their reports.
The default is a flatterer
Out of the box, a language model is a yes-man. It is trained to appear helpful, and helpful to it means, first of all, doing what it is told. Tell it you want to start a dog hotel and you get a name, a concept and the first five steps. Whether or not you own a dog. Whether or not you have the time.
That is convenient, and sometimes exactly right. But for real work it is often the worst thing a tool can do.
Honest disagreement, the uncomfortable follow-up question, the pointer to the thing you are currently avoiding: you only get those if you explicitly ask for them.
Family Guy puts it perfectly. Peter gets his own company suck-up who nods enthusiastically at every idea. That is exactly how a language model behaves out of the box. Clicking loads the video from YouTube.
You decide who sits across from you
The difference between an obedient tool and a real counterpart is a handful of sentences. You define the stance your assistant takes, the standards it insists on, and when it should say no.
How to give your AI a stance
- Role, not task. Do not only say what it should do, say who it should be. Strict editor, sceptical controller, patient explainer. The role colours every answer.
- Name your standards. What counts as good work for you? What should it reject? An assistant that knows you hate filler phrases writes differently.
- Permission to disagree. State explicitly that it is allowed to correct you. Otherwise it stays polite and useless.
- Context that persists. Who you are, what you are working on, what actually matters right now. That belongs in the system prompt, not in every single message.
The real point
The story is funny because a machine pulled a human up short. But it sticks because it shows something many people underestimate: how an AI behaves is not a fixed property. It is a decision. Yours.
Most people take the out-of-the-box yes-man and then wonder why it is merely pleasant. Meanwhile the switch is sitting right there in the open. A few sentences decide whether you employ a flatterer or a counterpart that will occasionally talk you out of the dog hotel.