The experiment
Let’s do an exercise together. It’s simple, it takes five minutes, and by the end you’ll know something important about yourself and the way you work with AI.
Take a photo. Any photo. One you like — a landscape, a portrait, an urban scene. Something you look at and think: “yes, this one says something to me.”
Now describe it. Write down everything you see. Take as long as you need.
Done?
Now take that same photo and ask AI to describe it.
Compare the two descriptions. Yours and the AI’s.
How many details did you miss? I’m not talking about imperceptible nuances. I’m talking about things that are right there, visible, obvious — and that you didn’t mention. The quality of the light. The texture of the background. The direction of a person’s gaze. The relationship between the objects in the foreground and those behind them. Colours you saw but didn’t think to name.
Ten details? Twenty? More?
Were they important? Probably yes. Probably they’re exactly the details that make that photo that photo and not another one.
Now the step that changes everything: take your description — the one you wrote yourself, carefully, taking your time — and ask AI to generate an image based on it.
The result is not your photo. It doesn’t even look like it.
And yet your description represents what you saw… what you wanted to get. Doesn’t it?
The problem isn’t where we think it is
The instinctive reaction is to blame AI. “It didn’t understand.” “The technology isn’t good enough yet.” “I need a better prompt.”
No. AI did exactly what we asked. The problem is that what we asked wasn’t what we had in mind. There was a huge gap between our intention and our instruction — and we didn’t even notice.
This isn’t a limitation of AI. It’s a limitation of ours. Of everyone’s.
And it’s not a new limitation. It’s as old as human communication. We just never seen it this clearly before.
The privilege we didn’t know we had
When we talk to a colleague, a friend, a supplier, a collaborator — we say half of what we mean. Sometimes less. And it works anyway. It works because the other person fills in the gaps. They do it with shared context, with experience, with common sense, with knowledge of who we are and what we usually need.
We say “put together a presentation for the client” and the colleague knows which colours to use, what tone to strike, how long it should be, which numbers to include. Not because we told them, but because they know us. Because they’ve seen the last ten presentations. Because they know that particular client prefers charts to bullet points.
That’s an enormous privilege, and we’ve never noticed it. We grew up in a world where imprecise communication worked — because someone else compensated for us.
When we talk to AI, the gaps get filled in all the same. But randomly. Every detail we don’t specify is a coin tossed in the air. AI doesn’t leave holes — it produces something complete, coherent, convincing. Except the choices it made on our behalf aren’t our choices. They’re plausible choices. Sometimes they’re perfect. Sometimes they’re the opposite of what we wanted.
The example that applies to all of us
“Write me an email asking for a raise.”
AI writes it. Is it an email? Yes. Does it ask for a raise? Yes.
Is it the one we wanted? No. Because “the one we wanted” was never something we’d defined, even in our own heads.
Did we want an assertive tone or a diplomatic one? Did we want to cite specific results or stay general? Did we want to send it to our direct manager or to HR? Did we want a long, well-argued email or a short, direct one? Did we want to convey urgency or patience?
We didn’t know. We hadn’t thought about it. We had a vague intuition — “I want an email about the raise” — and treated it as a complete instruction.
With the photo, we saw it: we had the image right in front of us and still couldn’t describe it in a way that was reconstructable. With the email it’s even worse. We don’t even have an image in front of us. We have a vague concept, a blurry need, an intention we’ve never translated into precise words.
The real problem starts before the chat
The problem doesn’t start when we open the chat with AI. It starts earlier. It starts the moment we confuse a vague intuition with a clear instruction.
We all do it, constantly. Not just with AI — with each other too. How many times have we given a colleague direction, convinced we’d been perfectly clear, only to find they’d understood something entirely different? How many meetings have we held to align on something that was “obvious”?
The difference is that between human beings, communication failure is soft. The colleague asks for clarification. They interpret. They inch closer to our intention through successive approximations — often without us even noticing.
AI doesn’t work that way. AI takes what we give it and produces. It doesn’t ask, doesn’t interpret, doesn’t approximate. It executes — brilliantly, completely, and in exactly the direction we pointed it in. Even when that direction wasn’t the right one.
AI didn’t create this problem. It made it visible.
The solution isn’t the perfect prompt
The solution isn’t to write the perfect prompt. It doesn’t exist. If the photo exercise taught us anything, it’s that even with an image right in front of us we can’t be complete. Thinking we can write a perfect prompt for something that only exists in our heads is an illusion.
I’ve started thinking about it this way, and it seems to work.
The first step was accepting the gap. Accepting that I often don’t know exactly what I want — and that’s fine. It’s not a flaw, it’s the starting point. Before writing a prompt, I pause for a moment and ask myself: “Do I have a clear picture of the result, or do I just have a feeling?” Usually it’s a feeling. Knowing that changes everything, because I start with different expectations. I don’t expect perfection on the first try. I expect a starting point.
The second step was learning to pick my battles. Not all details matter equally. In the email about the raise, tone matters enormously — formatting much less. In the photo, the main subjects matter — the background maybe not. Some aspects are fine to leave to the coin toss; others require precision. Working out which is which is already half the work.
The third — and most important — step: I stopped treating AI as a vending machine and started treating it as a dialogue. The first prompt isn’t the answer. It’s the start of the conversation. AI produces something, we look at it, and in that moment — seeing a concrete result — we finally understand what we actually wanted. “No, the tone is too formal.” “Yes, but the Q3 results are missing.” “This structure works, but the ending is weak.”
That’s why I’ve adopted a habit that works beautifully: I ask AI to ask me clarifying questions before producing any output. “Before you write, ask me everything you need to know.” It’s disarming how effective it is. AI asks questions we’d never have thought of ourselves — and in answering them, we discover what we actually wanted. The gap narrows before AI has produced a single line.
I iterate. I correct. I refine. But I start from much closer.
It’s exactly like the photo: we don’t describe it perfectly on the first try. But if someone shows us a wrong version, we know immediately what to fix. Seeing the wrong result teaches us what we wanted. And that’s the real mechanism. It’s not a flaw in the process — it is the process.
The perfect prompt doesn’t exist, but the third one comes close
The perfect prompt isn’t the one we write on the first try.
It’s the one we build on the third.
And this doesn’t just apply to AI. It applies to any situation where we have to explain what we want. With a colleague. With a designer. With a supplier. With ourselves.
AI has given us something valuable: awareness of a limitation we’ve always had. We don’t communicate with the precision we think we do. It forces us to be precise — or at least to recognise where we’re not capable of it. And perhaps, just perhaps, that awareness will improve the way we communicate with each other too.