How to Write Better AI Prompts: A Practical Framework

When an AI assistant gives you a bland, generic or wrong answer, the cause is usually the request rather than the tool. Language models respond to what you give them. A one-line prompt leaves them guessing about who the answer is for, how long it should be and what "good" looks like. A few extra sentences remove the guesswork.

Here is a framework that works across the major assistants, whether you use it for writing, analysis or planning.

The six parts of a strong prompt

1. Role: who should answer

Tell the assistant what perspective to take. "You are an experienced bookkeeper explaining things to a first-time business owner" produces a different answer from "You are an accountant writing for other accountants". The role sets vocabulary, depth and assumptions.

2. Task: what you actually want

Use a clear verb: summarise, compare, draft, critique, list, rewrite. "Help me with my email" is a topic. "Rewrite this email so it is polite but firm about the overdue invoice" is a task.

3. Context: what the model cannot know

The assistant does not know your audience, your previous decisions or your constraints unless you say so. Include the facts it needs: who will read the output, what they already know, what has been tried, and why the task matters. Paste in the source material rather than describing it.

4. Format: what the output should look like

Specify length, structure and style: "three bullet points", "a table with columns for cost, effort and risk", "under 150 words", "plain English, no jargon". Format instructions are the quickest way to get output you can use without heavy editing.

5. Constraints: what to avoid

State the boundaries. "Do not invent statistics." "If you are unsure, say so." "Use only the information in the document below." Constraints are especially important for factual work, where a confident wrong answer is worse than no answer.

6. Examples: show, do not just tell

If you have a sample of the tone or structure you want, include it. One good example often does more than a paragraph of description.

Before and after

Weak prompt: "Write a LinkedIn post about our new feature."

Stronger prompt: "You are a product marketer at a small accounting software company. Draft a LinkedIn post announcing that our app now reconciles bank transactions automatically. The audience is owners of businesses with fewer than ten staff who currently reconcile by hand. Keep it under 120 words, open with the problem rather than the product, avoid hype words like 'revolutionary', and end with a question that invites comments. Here are the release notes: [paste]."

The second version takes thirty seconds longer to write and usually saves several rounds of revision.

Iterate instead of starting over

Treat the first answer as a draft. Specific feedback works better than regenerating:

• "Make the second paragraph shorter and more concrete."
• "The tone is too formal for this audience; make it conversational."
• "List the assumptions you made, then revise using these corrections."

Asking the assistant to critique its own output ("What is the weakest part of this answer and why?") is a useful step before you ask for a rewrite.

Prompts for factual and research tasks

For anything where accuracy matters, add three instructions:

1. Provide the source material and ask the model to answer only from it.
2. Ask it to quote or point to the passage that supports each claim.
3. Tell it to say "not stated in the source" rather than guess.

Then check the important claims yourself. Our guide to AI hallucinations and how to check AI output explains why this matters and how to do it quickly.

Build a small prompt library

When a prompt works well for a recurring task, such as meeting summaries, customer reply drafts or first-pass proofreading, save it with placeholders for the parts that change. A shared library helps a team get consistent results and makes good practice easy to follow.

What not to put in a prompt

Before pasting anything, check your organisation's rules on confidential information. Customer personal data, passwords, unreleased financial results and anything covered by a confidentiality agreement may not be allowed in external tools. If your team does not yet have rules, our article on writing an AI use policy for small teams is a good place to start.

Quick reference

• Role: who is answering
• Task: one clear verb
• Context: audience, background, source material
• Format: length, structure, tone
• Constraints: what to avoid, what to do when unsure
• Examples: a sample of what good looks like