AI Doesn't Read Minds.It Reads Notes.

The quality of your answer is decided before you press enter. Here is how to write the prompt.

Somewhere right now, a very capable AI is producing a completely useless answer.

Not because it is dumb. Because somebody typed "summarize this report", hit enter, and walked away believing that was an instruction.

You have probably done it. Everyone has. You paste a document, ask for a summary, and get back a paragraph that is smooth, confident, but missing the one thing you actually needed. You try again with different words. Same result. By the third attempt you have typed more than the summary would have contained, and you have quietly filed AI under "overrated".

Here is the twist. The AI did not fail. It obeyed.

The two word note

Imagine you had a helper who could go to the shop for you. Capable, fast, endlessly willing. One catch: the helper knows nothing about you except what is written on the note you hand over.

You write:

It comes back with two litres of full fat. You wanted a small carton of oat milk for your coffee. Was the helper wrong? Not really. The note was two words long, and the world contains a lot of milk.

So you rewrite the note:

Same shop. Same helper. Perfect result. The helper did not improve. The note did.

Every AI chat works the same way. The note you type has a name: the prompt. And the model cannot see your head, your inbox, or your job. It has your prompt, and only your prompt. Every vague word in it is a blank, and blanks get filled with guesses. In small talk, guesses are free. In real work, a guess is how you end up confidently wrong.

Six lines that do the heavy lifting

So what goes into a good prompt? Look at prompts that work and prompts that flop, and the pattern is almost embarrassingly consistent. The good ones cover six things. None of them are technical. All of them are briefing. And the recipe is identical whether you use ChatGPT, Copilot, Claude, or Gemini.

1. The role. Tell it who to be. "You are a health writer who explains research to patients" changes everything that follows. Ethan Mollick, the Wharton professor who studies how people actually work with AI, gives the same advice: treat it like a person, and tell it which person to be.
2. The task. One verb. Explain, compare, shorten, draft, check. If you cannot pick the verb, the model certainly cannot.
3. The background. Paste the real thing: the report, the leaflet, the notes. A description of a document is your opinion of it. The document itself is the facts.
4. The output. Name the shape. Under 150 words. A table. Three bullets. An email draft. These are different products, so order the one you want.
5. The reader. Every answer is for someone. A nervous patient, a rushed manager, a room full of specialists. Same content, very different clothes.
6. The rules. Your do's and don'ts. "Flag anything you are unsure of. No jargon. Never invent references." This line protects you, and it is the one most people skip.

Put together, "explain this medicine" grows up into:

Six lines. Forty seconds of typing. This is most of the trick.

The part that sounds made up

If this sounds like just a nice theory, here is the part that has been measured.

In 2021, researchers ran a revealing experiment at one of the field's main research conferences. The study is linked in the sources below. They gave an AI model a few solved examples to learn from, and changed nothing except the order those examples appeared in. Accuracy ranged from 54 percent, basically a coin flip, to 93 percent, close to the best systems of the day.

Read that again. Same model. Same examples. Different arrangement of the prompt. The study tested one narrow thing, and the models in it were small by today's standards, but the wider lesson has aged well: what you hand these systems, and how you arrange it, changes what comes back.

OpenAI's own prompting guide says the quiet part plainly: give the model the source material, and it invents less. Background shrinks the guessing. It does not delete mistakes, which is why a human still checks the answer. But it moves you from gambling to briefing.

Watch two prompts grow up

Makeover one: explaining trial results

Why it works: the tool stopped choosing the audience, the length, and the tone. You chose.

Makeover two: meeting notes

Why it works: the AI knows the shape of the answer before it writes a single word. No guessing, no filler.

Five mistakes most people make, and how to fix them

Most prompts that do not work fall into one of these five patterns. If your AI keeps giving you unhelpful answers, check which one you are making.

• Being too vague. You typed something short and hoped the AI would figure out the rest. Fix: be specific about the one thing you most need from the answer.
• Describing a document instead of sharing it. You wrote "I have a report that says..." instead of pasting the actual text. Fix: copy and paste the real thing. Your description is your version of it. The document is the facts.
• Asking for too much at once. You put five different jobs into one message. Fix: break it into steps, one message per job.
• Forgetting to say who it is for. The AI wrote a perfectly fine answer for the wrong person. Fix: name your reader. "Explain this to a patient" and "explain this to a doctor" are very different jobs.
• Treating the first answer as final. The first reply is a starting point, not a finished product. Fix: respond with small adjustments. "Make it shorter." "Add the cost." "Sound warmer." Most good answers come after one or two rounds of this.

Most of these are one line in a prompt or one habit in a workflow. We'll take them apart properly over coming issues.

Remember it as PROMPT

You do not need a template. You need one word -- and it is the thing you are already writing:

• P -- Purpose: what you want done
• R -- Role: who the AI should be
• O -- Output: what you want back
• M -- Material: the document or facts you are sharing
• P -- Person: who the answer is for
• T -- Terms: your rules and limits

Run through them before you press enter. Skip the ones that do not apply. But when the answer matters, ten seconds with PROMPT saves you far more time than sending the same vague message three times.

It was never about the machine

Here is the strangest part of the AI era: the skill that gets the most out of these tools is not technical. It is a skill you already have. Explaining the job clearly before someone starts.

You already know how to do it. You do it with colleagues, contractors, couriers, anyone who acts on your behalf. The machine simply needs the same courtesy, in writing, every time.

A prompt, in the end, is just the note you hand the machine. Write a better PROMPT, and you will get a better answer.