How to write a good prompt

A “prompt” is just what you type into the box — your request. And the single biggest thing standing between people and a useful answer isn’t the AI; it’s a vague request. The good news is that writing a better one is a small, learnable craft, not a talent. This page is the hands-on version: the anatomy of a clear prompt, with real before-and-after examples you can copy the shape of. If you want the broader mindset — the handful of habits that make you good at this over time — that lives next door on Use AI well. Think of that page as the principles and this one as the concrete craft of the prompt itself.

The anatomy of a good prompt

Almost every weak prompt is weak for the same reason: it assumes the AI knows things it can’t possibly know — who you are, what you’re trying to do, what a good answer would even look like to you. It can’t read your mind or see your situation. A good prompt simply hands it the things a thoughtful human helper would have asked for. There are five of them, and you rarely need all five at once:

• Context. Who you are and what this is for. “I’m a parent…,” “I’m studying for a high-school biology test…,” “this is going to my landlord….”
• What a good answer looks like. The length, the tone, the level. “Keep it short and warm,” “explain it like I’ve never seen the topic,” “be blunt.”
• An example, if you have one. A sample of the thing you want — an old email you liked, a sentence in the right voice. Showing beats describing.
• The format you want. A bulleted list, a table, three options, a single paragraph, a draft I can paste as-is.
• Permission to ask you questions. One line — “ask me anything you need before you start” — turns a guess into a conversation.

That last one does more than it looks like it should, so it gets its own section below. But the headline is simple: a vague prompt makes the AI guess, and it will guess confidently. A specific prompt lets it actually aim.

Before and after: a parent drafting an email

Here’s the difference in practice. Say you want help writing a note to your child’s teacher.

Vague: “Write an email to my son’s teacher about his grades.”

You’ll get an email — a generic, slightly stiff one, because that’s all the request supports. The AI doesn’t know your son’s name, what’s actually going on, or how you want to come across.

Better: “I’m the parent of a 10-year-old, Marcus, whose math grade dropped this term. I want to email his teacher, Ms. Reyes, to ask for a quick call and find out how I can help at home — warm and collaborative, not accusatory, about four sentences. Give me a draft I can send, and ask me anything you need first.”

That version carries context (who, what, the actual situation), the shape of a good answer (warm, not accusatory), the format (four sentences, send-ready), and the door held open for questions. The draft that comes back is usable instead of a starting-over point — because you told it what you actually needed.

Before and after: a student studying a topic

Same move, different goal.

Vague: “Explain photosynthesis.”

You’ll get a tidy paragraph — possibly pitched at a textbook you can’t follow, or so simple it’s useless for your test. The request didn’t say who’s asking or why.

Better: “I’m a 9th grader with a biology quiz Friday. Explain photosynthesis like I’m new to it, in plain language, then give me three questions to test whether I actually got it. If I get one wrong, explain that part again a different way.”

Now it knows the level (9th grade, new to it), the purpose (a quiz), and the format (explanation, then a self-quiz, then a second pass). You’ve turned a flat definition into something that teaches. Notice that none of this required knowing a single technical term — just describing your situation honestly.

Iterate — it’s a conversation, not a vending machine

The most common mistake is treating the first answer as the final answer. These tools work best when you go back and forth: read what you got, say what was off, and ask for the fix. “Make it shorter.” “Too formal — loosen it up.” “That second point isn’t quite right; here’s what I actually meant.” Each reply lands in the same conversation, so the AI keeps everything you’ve already said and adjusts from there.

You don’t have to write the perfect prompt up front. A decent first prompt plus two rounds of “not quite, more like this” beats agonizing over the opening line. The skill isn’t crafting one flawless request — it’s steering.

A better-shaped answer is not a truer one

Here’s the line worth underlining, because it’s the one people most often miss. A better prompt gets you a better-shaped answer — more relevant, better aimed, in the format you wanted. It does not make the answer more true. A beautifully phrased request can return a confident, well-organized, completely wrong reply. Prompting controls the shape of the output, not its accuracy.

So the craft on this page and the habit of checking facts are two separate jobs, and you need both. When the answer contains a fact you’d act on — a date, a dosage, a legal point, a citation — verify it somewhere real, no matter how good your prompt was. The page on why AI makes things up explains exactly why a fluent answer and a correct one aren’t the same thing, and how to catch the gap.

One more practical limit: on long tasks, the AI can only “hold” so much of the conversation at once. Past a certain length, earlier parts start to fall out of view — which is why a great prompt buried thirty messages back can quietly stop being followed. That ceiling is the context window, and knowing it’s there tells you when to restate the important bits rather than assuming they’re still in front of it.

Where to go next

Tried one of these and it didn’t land — or know a clearer way to put it? Tell us — an education resource only earns trust by being checkable.