How to Write AI Prompts: A Simple Framework for Everyday Work

Write an AI prompt the way you would brief a capable new colleague who knows nothing about your situation: say what you want done, who it is for, what material to use, what the result should look like and what to avoid. Then read the answer and refine it. Most disappointing results trace back to a missing piece of that brief rather than to the tool itself.

The five parts of a useful prompt

1. Task. Start with a clear verb: draft, summarize, compare, rewrite, list, classify, explain.
2. Context. Who will read the result, why it is needed and what the reader already knows.
3. Material. The text or facts to work from, clearly separated from your instructions with quotation marks, a heading or a line such as "Source text below".
4. Format. Length, structure (bullets, table, short paragraphs), tone and reading level.
5. Constraints and examples. What to leave out, terms to avoid and, where style matters, a short sample of what good looks like.

Not every prompt needs all five. A quick definition needs only the task. A client email or a report summary usually needs every one.

Weak and strong prompts side by side

Weak prompt	Stronger prompt	What changed
Write an email about the delay.	Draft a short, polite email to a client explaining that their website launch moves from Friday to next Wednesday because the hosting move needs more testing. Offer a short call. Plain language, under 150 words, apologize once.	Audience, reason, new date, next step, length and tone are all defined.
Summarize this report.	Summarize the report below for a manager deciding whether to extend the pilot. Give five bullet points on cost, results, risks and open questions, and say if any of these are not covered.	The purpose decides what counts as important, and gaps are flagged rather than filled.
Give me blog ideas.	List ten article ideas for a newsletter read by owners of small cleaning companies. Each should solve one practical problem. Avoid generic productivity tips.	A defined reader and a clear exclusion prevent bland suggestions.

Techniques that improve answers

Give a specific role

"You are an editor checking this text for clarity, for readers who are not accountants" works better than a vague "act as an expert", because it tells the model which standard to apply.

Show an example

Pasting one or two samples of the style or structure you want, sometimes called few-shot prompting, is often quicker than describing it in words.

Split big jobs into steps

Ask for an outline first, agree on it, then request one section at a time. Long, all-in-one requests tend to drift off course or end too early.

Let it ask questions

Close with "Before you start, ask me up to three questions about anything unclear." This brings missing details to the surface before they turn into wrong guesses.

Ask for assumptions on tricky tasks

For comparisons or calculations, ask the tool to list its assumptions before giving the answer, so you can spot a faulty premise quickly.

Save standing instructions

Many tools let you store instructions that apply to every chat, sometimes called a system prompt or custom instructions. Use them for things that never change, such as your usual audience, spelling preferences and phrases you never want to see.

Mistakes that waste time

• Vague goals. "Make it better" gives the model nothing to aim at. Say better for whom, and in what way.
• Trusting facts without checking. Language models can produce confident errors; the section on hallucinations in our explainer on large language models explains why.
• Demanding an exact word count. Models plan in tokens rather than words, so ask for a range. Our guide to tokens in AI covers the details.
• Pasting confidential material. Client data, passwords and unreleased figures do not belong in a tool your company has not approved; offline AI chatbots explained covers safer options.
• Stopping at the first draft. Follow-ups such as "shorter", "more concrete" or "drop the second paragraph" are part of the process, not a sign of failure.

A reusable prompt template

Copy this into a note and fill in the brackets each time:

• Task: [draft / summarize / rewrite / compare] [what]
• Reader: [who will read it and what they need to decide or do]
• Background: [two or three facts the tool must know]
• Material: [pasted below, clearly marked]
• Output: [length range, structure, tone]
• Avoid: [claims, phrases or topics to leave out]

The same structure works for writing jobs of every size. Our guide on how to write a letter with AI applies it to formal correspondence, from payment reminders to cover letters.

Questions people ask

Do longer prompts give better results?

Only when the extra words carry information. Relevant detail helps; padding, repeated instructions and polite filler do not.

Do I need special phrases or tricks?

No. Clear, specific language beats any formula. If a prompt would confuse a person reading it cold, it will probably confuse the model as well.

Should I write prompts in full sentences?

For anything beyond a quick lookup, yes. Full sentences carry relationships between ideas, such as "because" and "unless", that a string of keywords loses.