How to Write a Prompt That Gets a Useful Answer

Most people type a question into an AI the way they'd type a Google search. A few words, hit enter, hope for the best. Then they get back a generic paragraph that sounds smart and helps with nothing, and they conclude the tool isn't that good.

The tool is usually fine. The prompt was underfed.

A language model has no idea who you are, who the answer is for, or what "good" looks like to you. It will guess, and it will guess toward the average of everything it has ever read, which is almost never what you wanted. The fix is four small habits. None of them require being technical.

1. Say who the answer is for

The same question has different right answers depending on the audience. "Explain our new vacation policy" to a new hire is one thing. "Explain our new vacation policy" to a manager who has to approve requests is another. The model cannot see the audience, so tell it.

Compare these two prompts:

• "Summarize this email thread about the office move."
• "Summarize this email thread about the office move for a teammate who was out sick all week and needs to know what was decided and what they're expected to do."

The second one produces something useful. The first produces a book report.

2. Give the context it can't guess

The model knows a lot about the world in general. It knows nothing about your week, your team, your constraints, or what already failed last time. If the information matters to the answer, paste it in. Yes, even if it feels like too much.

Planning your week is a good example. "Help me plan my week" will give you a cheerful list of nonsense. This will not:

"Help me plan this week. I have a report due Thursday that will take about six focused hours. I have three meetings Tuesday that eat the whole afternoon. I'm picking up my kid at 3pm on Wednesday and Friday. I work best on hard things before noon. Please block out the week and tell me when I'll write the report."

You added five facts. The answer becomes specific. This is the whole trick.

3. Show an example of the output you want

Describing the format you want in words is slower and worse than just showing one. If you want short bullet summaries, paste a short bullet summary you liked and say "like this." If you want the tone of a particular email, paste the email.

Say you have a long thread from a vendor and you want a two-line summary plus a recommended reply. Instead of explaining that, give it one worked example:

"Here's the thread. Reply in this format:
SUMMARY: (two lines)
SUGGESTED REPLY: (three sentences, friendly but direct)
OPEN QUESTIONS: (bullets, only the ones I need to decide)"

You just saved yourself three rounds of "no, shorter" and "no, less formal."

4. Ask it to say when it's unsure

This is the habit almost no one uses, and it's the one that saves you from the model's biggest weakness: it will cheerfully make things up rather than admit it doesn't know. A line like this at the end of your prompt changes the behavior:

"If any part of this requires information you don't have or you're guessing, mark that line with (unsure) and tell me what you'd need to answer it properly."

You'll be surprised how often it flags something. A date it doesn't actually know. A policy it invented. A name it guessed. Now you know which parts to trust and which parts to check, instead of discovering the invented one in a meeting.

Putting it together

Here is what a useful prompt looks like when all four habits are in play. The task: summarize a long email thread about a project disagreement.

"Below is an email thread between three people on my team about whether to delay a product launch. I need to brief my boss, who hasn't read any of it and has about two minutes. She cares about the decision and the risk, not the personalities. Please give me: (1) the decision being debated, in one sentence, (2) each person's position in one sentence, (3) the open questions she'll want answered. Keep it under 150 words. If anyone's position is unclear from the thread, mark it (unsure) rather than guessing."

That prompt will get you something you can actually walk into a meeting with. The ingredients were: audience (your boss, two minutes, cares about decision and risk), context (the thread itself, who the players are), format (three numbered parts, word limit), and a request to admit uncertainty.

None of this is advanced. It's just the difference between asking a stranger a vague question in a hallway and briefing a smart new assistant properly. Same person, different answer.

If posts like this are useful to you, you can get the weekly ones by subscribing on our blog.