Everything you leave out, it decides for you.
And it decides average. This is the one habit that separates people who get real value from AI from people who conclude it is overhyped — and it works the same way for text, images, and video.
“I tried ChatGPT for proposals. It was generic.” Of course it was. Look at what it was given.
Almost every disappointing AI result traces back to the same thing, and it is not the model. It is that the request contained a fraction of what a human colleague would have needed — and unlike a colleague, AI never says “hang on, which client? what budget? what tone?” It just fills the gaps silently and hands you something confident.
There is a mechanical reason for this, and it is worth understanding once. A model predicts the most likely continuation of the text in front of it. When the text in front of it is thin, the most likely continuation is the most average one — the blandest, most common version of that kind of writing that exists. Vagueness in does not produce randomness out. It produces the statistical middle.
Generic output is not the model failing. It is the model doing exactly its job on a starved input.
How much are you actually leaving out?
Here is a request that feels perfectly reasonable — the sort of thing anyone would type on a busy afternoon. Tap each question to see what got decided on your behalf.
“Write an email to the client about the venue change.”
Eleven words. Feels complete. It is not — every question below had to be answered before a single sentence could be written, and you answered none of them.
The same task, five times, one model.
Nothing changes here except how much you say. Drag through the versions and watch a useless answer become a usable draft — and note where the jump is biggest. It is not between one and two. It is the moment real documents enter the picture.
“Write a proposal.”
A blank template about nothing. Headings like “Introduction” and “Our Approach”, filled with sentences that would suit a plumbing firm equally well.
Level three deserves a warning of its own. That is the dangerous rung — detailed enough to look professional, empty enough that every number in it is fiction. A proposal that is obviously generic gets rewritten. A proposal that is confidently wrong gets sent.
The same rule runs images and video.
People assume image and video generation work on different principles. They do not. “A nice dinner setup” leaves the same blanks as “write a proposal”, and they get filled the same way — with the most average dinner setup in the training data.
The difference is that visual prompts have a reliable anatomy. Switch parts off below and watch the prompt fall apart; each part controls something specific.
A welcome dinner for 40 guests at one long table, on a Lisbon rooftop overlooking the river, at golden hour, warm string lights overhead, terracotta and white linen, low olive branch centrepieces, wide shot from the head of the table, shallow depth of field, photorealistic, editorial travel photography, no visible logos, no text overlays, no distorted faces.
Nobody writes this in one go. You write the subject, generate, see what is wrong, and add the part that fixes it. After a dozen tries you are writing prompts like this from memory — that is the whole learning curve, and it is shorter than people expect.
Five parts. Not all five, every time.
For a quick question, type the quick question. This is for the work that matters — the proposal, the client email, the itinerary review.
Who it should be
“You are an experienced DMC operations manager.”
Steers it toward a body of writing with the right vocabulary and the right worries. Cheap, and it genuinely changes the output.
What the job is
“Review this itinerary for operational risks.”
One clear task. Two tasks in one prompt reliably gets you two half-answers.
The raw material
[the actual itinerary, attached]
The highest-value line in any prompt. Attach the real thing rather than describing it — you are changing what every prediction is built on.
What good looks like
“A table: issue, which day, how serious, suggested fix.”
Describe the shape you want or you will get prose you then have to reformat by hand.
The guardrails
“Use only the attached document. Do not invent times or rates. List what is missing.”
This is the line that stops confident fiction. It should be in every prompt that touches client-facing facts.
“That is more work than writing it myself.”
For one email, sometimes true — so write the email. The maths changes completely the moment a prompt gets reused. Your proposal prompt is written once and used every week for a year. Your company brief is written once and pasted into every important chat from then on.
There is also a shortcut nobody mentions: you do not have to type any of it. Dictate. Ramble for ninety seconds about the client, the constraint, the tone you want and what you are worried about, and let the mess be the prompt — AI is unbothered by rambling and will use every scrap of it. Most people write short prompts because typing is tedious, not because they have nothing to say.