How to Write Better AI Prompts: A Practical Guide

AI assistants are good at following instructions and bad at reading minds. Most disappointing answers trace back to a prompt that left out something the writer knew but never said. The fix is not a secret formula. It is the same thing that makes a good brief for a human.

In short: Tell the assistant who the answer is for, exactly what you want, what form it should take and what material to use. Then improve the first answer instead of accepting it.

1. Say who it is for and why

The same question deserves different answers for different people. Tell the assistant who will read the result and what it is for.

Weak: Explain inflation.

Better: Explain inflation to a 15-year-old who has just started an economics class. Use one everyday example and keep it under 150 words.

2. Be specific about the task

“Help me with my CV” could mean a dozen things. Name the action you want: rewrite, shorten, critique, compare, list, translate, find errors in.

Weak: Make this email better.

Better: Rewrite this email so it is polite but firm. The goal is to get the invoice paid this week. Keep it under 120 words and keep my opening line.

3. Describe the output you want

If you care about length, structure or tone, say so. Otherwise you will get the assistant's default, which tends to be longer and more list-heavy than you need.

• Length: “in three sentences”, “about 300 words”.
• Format: “as a table with columns for cost, pros and cons”, “as plain paragraphs, no bullet points”.
• Tone: “friendly and direct”, “formal, for a legal audience”.

4. Give it material to work from

Assistants do far better working from text you provide than from memory. Paste in the document, the data, the job description or the error message. If you want writing in your style, include a sample of your writing. One good example of the output you want is often worth a paragraph of description.

Be careful with what you paste. Do not share passwords, confidential company information or other people's personal data unless your organisation has approved the tool for that use.

5. Treat the first answer as a draft

The biggest difference between casual and effective users is that effective users reply. Tell the assistant what was wrong and ask again:

• “Too formal. Make it sound like a text to a colleague.”
• “Keep the second paragraph, cut the rest by half.”
• “Option 2 is closest. Give me three more in that direction.”

A template you can adapt

I am [who you are / your situation].
I need [the specific task].
It is for [audience and purpose].
Here is the material: [paste text or data].
Please give me [format, length, tone].
Avoid [anything you do not want].

You will not need every line every time. For quick questions, a single clear sentence is fine.

Two techniques for harder problems

Ask for the reasoning. For anything involving calculation, logic or a judgement call, ask the assistant to explain how it reached the answer. The explanation makes mistakes much easier to spot.

Ask what is missing. End a complex request with: “Before you answer, ask me any questions you need.” This often surfaces the detail you forgot to include.

Check before you rely on it

A well-written prompt improves relevance. It does not guarantee accuracy. AI assistants can state incorrect facts with complete confidence, including invented quotations and references. For anything that matters, verify the key points using the method in how to fact-check AI answers.

A worked example, from first try to useful answer

Imagine you run a small bakery and want an email announcing new opening hours.

First attempt: “Write an email about our new opening hours.” The result is polite and generic. It invents hours, runs long and sounds like a bank.

Second attempt, with context: “Write a short email to customers of a neighbourhood bakery. From 1 March we open at 7:00 instead of 8:00 on weekdays; weekend hours stay the same, 8:00 to 14:00. Warm, plain tone. Under 90 words. One subject line.” Now the facts are right and the length fits.

Third step, refining: “Good. Make the subject line shorter, mention that fresh bread will be ready from opening time, and remove the exclamation marks.” Three small instructions turn a decent draft into one you can send.

Nothing here is a trick. The improvement came from supplying facts the assistant could not know, stating limits, and editing instead of starting again.

Adjusting the approach to the task

• Summarising. Say who the summary is for and what they will do with it. “For a manager deciding whether to read the full report” gives a different result from “for my own revision notes”.
• Learning something. Give your current level and ask for an example, then ask to be quizzed. Explaining an idea back and asking what you got wrong is especially effective.
• Editing your own writing. Ask for specific changes, such as “shorten by a third without losing any facts”, and ask for a list of what was changed.
• Working with data. Paste a small sample, describe the columns and state the exact output you need. Check calculations yourself; assistants can make arithmetic slips.

Weak prompts and stronger versions

Mistakes that waste time

• Asking for everything at once. Split large jobs into steps and check each one.
• Leaving out the audience. The same facts read very differently for a child, a customer and a specialist.
• Starting a new chat to fix a small flaw. Say what to change and keep the rest.
• Pasting confidential material. Check your organisation’s rules, and remove names and personal data you do not need.