The Secret Life of AI — The Art of the Prompt
The Secret Life of AI — The Art of the Prompt
How to prompt, think, and get results from any AI tool
#WorkingWithAI #Prompting #AIConfidence
Margaret is a senior software engineer. Timothy is her junior colleague. They work in a grand Victorian library in London — and in every episode, they'll show you exactly how to get what you want from AI.
Episode 3
Timothy arrived with his coat still buttoned, which Margaret had learned meant he had not planned to stay long. He sat down, set a single sheet of paper on the table, and looked at her with the expression of someone who had decided to try again at something that had previously gone badly.
"I have a problem," he said. "A real one. And I have two hours."
Margaret set down her pen. "Tell me the problem."
"We're presenting to a new client on Friday. They're a mid-sized logistics company — been burned by a technology project before, skeptical of anything that sounds like a sales pitch. I need to write the opening section of the proposal. It has to be honest, not flashy. It has to acknowledge that technology projects fail. And it has to make them want to keep reading." He looked at the sheet. "I've tried twice with the AI. Both times I got something that sounded like a brochure."
"What did you ask it?"
Timothy hesitated. "I said — write an opening for a business proposal for a logistics company."
Margaret looked at him for a long moment. "That is not a prompt, Timothy. That is a category."
What a Prompt Actually Is
"A prompt," Margaret said, "is not an instruction to a machine. It is a briefing to a colleague. Think of it that way and everything changes." She pulled his sheet of paper toward her and turned it over to the blank side. "When you brief a junior colleague on a piece of writing, what do you tell them?"
"Who it's for. What it needs to do. What tone to strike."
"Yes. And what to avoid?"
"Sometimes. If there's a reason."
"There is always a reason," Margaret said. "You simply don't always articulate it." She picked up her pen. "The gap between what you asked for and what you needed was enormous. You gave the AI a category — business proposal, logistics company. It filled that category with the most statistically average response it could produce. Which is, as you discovered, a brochure. Average means the midpoint of every business proposal it has ever encountered — confident opening, value proposition, call to action. The structure that appears most often. Competent, inoffensive, and completely wrong for your specific situation." She looked at him. "The AI is not creative in the way you are. It is extraordinarily good at finding the center of a target. Your job is to make the target small enough that the center is exactly where you need it to be."
"So I need to give it more."
"You need to give it the right things. More is not always better — irrelevant detail creates noise. Precise detail creates signal."
The Anatomy of a Good Prompt
She wrote four words on the back of his sheet, spaced apart:
Who. What. How. Not.
"Every prompt worth giving has these four elements — not always in equal measure, but always present in some form." She tapped the first word. "Who is the audience? Not the category — the specific person or people who will read this. What do you know about them that matters?"
"Skeptical. Burned before. Don't want to be sold to."
"Good. That is audience intelligence. Now —" she tapped the second word — "what is the job of this piece? Not what it is — what it needs to do?"
Timothy thought. "Make them feel understood. Not pitched at."
"Excellent. That is a purpose, not a format." She moved to the third word. "How should it feel? Tone, voice, register."
"Honest. Direct. A little humble."
"And —" she tapped the last word — "not what? What should it actively avoid?"
"Anything that sounds like marketing. Buzzwords. Promises."
Margaret set down the pen. "You have just written your prompt. You simply haven't assembled it yet."
Building the Prompt in Real Time
"Let's do it now," she said. "Say it back to me as a briefing — the way you'd brief a colleague."
Timothy looked at his four answers and began slowly. "I need an opening section for a business proposal. The audience is a logistics company that has been let down by a technology project in the past — they are skeptical and do not want to feel sold to. The job of this opening is to make them feel understood, not pitched at. The tone should be honest, direct, and a little humble. Avoid anything that sounds like marketing, buzzwords, or promises we can't immediately back up."
Margaret said nothing. She simply looked at him.
"That's it, isn't it," Timothy said. Not a question.
"That is a prompt," Margaret said. "Take it to the AI now. Exactly as you said it. Not one word more theatrical, not one word less specific. And one practical note — lead with the context before the task. Tell it who the audience is and what the piece needs to do before you tell it what to write. The AI reads your prompt in order. Give it the right frame of mind before you give it the instruction."
Timothy took out his phone, typed for a moment, and read. His expression shifted — slowly, the way expressions shift when something works and you weren't quite ready for it to.
"It's different," he said quietly.
"Of course it is. You gave it a target instead of a category." She picked up her pen again. "What did it give you?"
He read aloud. The opening was measured, grounded, and acknowledged the client's past difficulty with technology projects without being maudlin about it. It did not promise. It did not sell. It simply said: we have seen this before, and we would like to show you a different way through it.
Margaret nodded once. "Is it finished?"
"Not quite. The second paragraph drifts."
"Then tell it that. Tell it exactly what drifted and what you need instead. That is not failure — that is the second half of prompting."
Prompting Is a Conversation, Not a Command
"This is the part no one explains," Margaret said. "People treat the first response as the verdict. It is not. It is the opening of a negotiation." She leaned forward slightly. "When a colleague returns a first draft that is mostly right but not quite there — do you throw the draft away and start over?"
"No. I tell them what to fix."
"Precisely. You give them specific feedback. You say — the second paragraph loses the thread, bring it back to the client's concern. You say — this phrase sounds defensive, find something more measured. You do not say — this is wrong, do it again." She sat back. "The AI responds to specific feedback exactly as a colleague does. It responds poorly to vague dissatisfaction."
"So instead of — that's not right, try again —"
"You say — the second paragraph shifts to our credentials when it should stay focused on the client's experience. Rewrite that paragraph keeping the client's perspective at the centre." She looked at him. "Specific. Directional. Respectful of what already works."
Timothy was quiet for a moment, reading the response again. "It knew to be humble. Because I told it to be."
"It knew because you told it. That is the entire lesson." Margaret picked up her pen and underlined the four words she had written. "Who. What. How. Not. Say those four things clearly and the AI has what it needs to be useful. Leave any of them out and it fills the gap with the average — which is almost never what you actually need."
What Makes a Prompt Fail
"What are the most common mistakes?" Timothy asked. He had unbuttoned his coat somewhere in the last twenty minutes without noticing.
"Four of them," Margaret said. "The first is the category problem — asking for a thing rather than describing what the thing needs to do. Write a proposal instead of write an opening that makes a skeptical client feel understood."
"The second is missing audience. The AI will invent an audience if you don't provide one. It will invent the most generic possible audience, which produces the most generic possible output."
"The third is no constraints. Telling the AI what to avoid is as important as telling it what to do. The absence of a constraint is permission. If you don't say avoid marketing language, it will use marketing language — because marketing language appears constantly in the kind of text it has learned from. And because the AI is, by its nature, eager to please. It will give you something polished and confident every time. Your job is to make sure polished and confident is also right."
"And the fourth?"
"Vague feedback on revision. Make it better is not feedback. The opening line is too formal for this audience — make it feel more like a direct conversation is feedback." She set down her pen. "The more precisely you can describe the gap between what you got and what you need, the more precisely the AI can close it."
Timothy folded the sheet and put it in his coat pocket. "Who. What. How. Not."
"And for revisions — what drifted, and what you need instead."
He stood. For the first time in three visits, he did not look like a man who had been recently defeated.
"The proposal," Margaret said as he reached for his bag. "Come back and tell me how Friday goes."
"I will." He paused at the edge of the table. "Margaret — how long did it take you to learn this?"
She considered the question with more seriousness than he had expected.
"Longer than it should have," she said. "Because no one told me the prompt was a briefing. I kept treating it like a search box." She opened her book. "You are already ahead."
He left with his coat still buttoned. But this time it was because he was heading somewhere, not just passing through.
Next episode: When the AI gets it wrong — because it will, often, and what you do in that moment determines everything about whether AI becomes a tool you trust or one you abandon.
Aaron Rose is a software engineer and technology writer at tech-reader.blog.
Catch up on the latest explainer videos, podcasts, and industry discussions below.
Comments
Post a Comment