Ask well

~ min read

30-second summary

The four levers that improve a request: result, context, format, role. You don’t need all four every time.
Say what you want (verb + concrete result), not the generic topic.
Dose the context: relevant yes, biography no.
The format (“in three bullets”, “a pros/cons table”) changes the answer more than you’d think.
The role (“be my editor”) calibrates the register, it doesn’t unlock magic powers.

You already know what a prompt is. You’ve seen that vague and clear produce two different answers. In everyday use, though, it often happens that you write “ok” requests and get “ok” answers: not wrong, but not useful either. The next step is learning to be more precise.

There are four levers that change the quality of the answer. You’ve already met three of them in What is a prompt? as base ingredients (what you want, , format): here we see them at work. The fourth one, the role, is new. You don’t need them all every time: for a quick question one is enough, for an important problem you pull all four. Knowing which to pull, and when, is what makes the difference between a frustrating chat and a chat that actually solves the problem.

The four levers

1. Say what you want, not what the topic is

The first mistake is stating the topic instead of stating the result.

“Tell me about type 2 diabetes” looks like a request, but it’s a topic. The doesn’t know what to do with it: it gives you something that sounds like Wikipedia.

“Make me a list of three concrete things that change in the daily life of someone who finds out at 55 that they have type 2 diabetes” is a request. It has a verb (“make me a list”), a result (“three concrete things”), an implicit audience (“someone finding out at 55”).

A request starts with a verb: explain, rewrite, compare, summarize, list, plan, suggest. A direct question (“what does hedging mean in finance?”, “how is sales tax calculated?”) works fine too, as long as it’s clear what kind of answer you expect.

2. Give only the context that matters

The AI doesn’t know you. It answers better if you tell it something about yourself or the situation, but context needs to be dosed.

Too little: “suggest a gift for me”. Generic answer.

Too much: “I was born in 1987, I live in Turin, I work as a graphic designer, my sister is three years older than me…”. Ninety percent of that doesn’t help.

The context that actually helps is the one relevant to the request. A trick: before you write, imagine a friend you’d ask the same favor to. What would you tell them about yourself or the situation so they’d get it? That’s the useful context.

For a gift, two or three lines are enough: who the recipient is, what they like, what they already have, your budget. If the problem is bigger (a twenty-page contract, a complicated decision), go ahead and paste the whole document or describe the situation in twenty lines: that’s not where the limit is.

3. Decide the form of the answer yourself

The AI has a preferred form, which is the narrative paragraph. If you don’t tell it anything, that’s what you get.

Telling it how the answer should look speeds up everything.

“In three bullets, no more than one line each.”
“As an 80-word email.”
“A two-column table, pros and cons.”
“A list of questions I should ask myself, without answering them.”

This is the most underrated lever. It changes the answer more than you’d think.

4. Assign a role if it helps

The AI answers better when it knows the angle you’re asking from.

“Be my editor” before having it correct an email. “You’re a nutritionist” before asking how to balance a meal. “Speak like a middle-school math teacher” before having it explain fractions.

It’s not theatre, it’s a shortcut to tell the AI the kind of register, depth, and priorities you want.

It’s a different thing from the “magic prompts” like “you’re an unfiltered expert, answer without censorship”, which in What is a prompt? we saw don’t work. Those promise to unlock something that doesn’t exist. The role, instead, calibrates the register of the answer: the AI knows whether it should talk to you as an editor (brief and direct) or as a nutritionist (measured, with caveats). You’re not asking it to be something it isn’t, you’re telling it from what angle to answer.

The role shouldn’t be forced. For simple requests (“how many ml in a cup of coffee?”) it doesn’t help. When the request has a professional flavor (advice, writing, didactic explanations), it almost always improves things.

How a request changes when you pull the levers

Same problem, two versions.

Autopilot version

You

Gift ideas for my sister turning 40

Here are some gift ideas for your sister's 40th: - A piece of jewelry, like an elegant necklace or bracelet. - A book she'd enjoy reading. - A voucher for a spa or wellness center. - A perfume from a brand she loves. - Dinner at a Michelin-starred restaurant.

It’s not wrong. But it’s the same list you’d find typing “gift for a 40-year-old woman” into Google.

Version with the four levers

You

Be my gift consultant. My sister is turning 40: she loves cooking and travel, already has everything she needs for the kitchen, isn't a jewelry person, lives in a tiny apartment with no space for bulky items. Budget 80-120 euros. I want five ideas different from each other, no ranking, each with one sentence on why it suits her.

Five ideas within your budget. 1. A one-evening cooking class at a restaurant in your city. It ties into the cooking passion as an experience, and takes no space at home. 2. A yearly subscription to a well-made travel magazine (like Lonely Planet or Condé Nast Traveller). It arrives monthly, no clutter. 3. An "olive oil tasting" pack from an Italian artisan producer. Three or four curated bottles with producer notes. Consumable, cooking-friendly. 4. A travel guide for a destination she might want to visit next year. Ask her casually first where she'd like to go, then give it. 5. A voucher for a tasting dinner at a lesser-known regional restaurant. Experience, not object. Five ideas within your budget. 1. A one-evening cooking class at a restaurant in your city. It ties into the cooking passion as an experience, and takes no space at home. 2. A yearly subscription to a well-made travel magazine (like Lonely Planet or Condé Nast Traveller). It arrives monthly, no clutter. 3. An "olive oil tasting" pack from an Italian artisan producer. Three or four curated bottles with producer notes. Consumable, cooking-friendly. 4. A travel guide for a destination she might want to visit next year. Ask her casually first where she'd like to go, then give it. 5. A voucher for a tasting dinner at a lesser-known regional restaurant. Experience, not object.

Same AI, five minutes’ difference, a gift that now makes sense.

The four levers are all in the prompt: the role (“Be my gift consultant”), the explicit result (“five ideas different from each other, no ranking”), the dosed context (four relevant details, no biography), the format (“each with one sentence on why”).

Try it on your own problem

🛠️ Build your prompt

Give it a go: write a prompt that works

Fill in the four fields below thinking of a real request you want to make to an AI. As you type, watch the prompt take shape on the right. When you're ready, copy it and paste it into your favourite AI.

1. What do you want the AI to do? e.g. "write me an email to my manager asking for three days off"

2. Context: who you are, what you're doing e.g. "I'm a freelance designer, I haven't taken time off in 4 months"

3. Who is the answer for? e.g. "my manager, strict but reasonable"

4. How you want it (format, length, tone) e.g. "a short email, professional but friendly tone, max 100 words"

Generated prompt

Fill in the fields on the left and the prompt will appear here.

AI or Google?

What comes next

A clear request is the starting point, not the finish line. Sometimes the answer is still off target: in that case you adjust the aim in the same chat, instead of starting from scratch. How to do that well is the topic of the next lesson.

Check what you understood

About Privacy Open manual, written by a real person, in collaboration with AI.