Writing better AI prompts for work: context, constraints and checks

By Axyra Nova Editorial · 15 October 2026 · 5 min read · Data and AI

Two people can ask the same AI assistant for help and get very different results. One types "write an email to a client" and gets something generic. The other explains who the client is, what happened, what the email needs to achieve and what tone to use, and gets a draft that needs only light editing.

The difference is the prompt. This guide explains a practical approach to writing AI prompts for everyday work tasks.

Think of it as briefing a colleague

A useful mental model: you are briefing a capable new colleague who knows nothing about your business or this task. They will do exactly what you ask, as well as they can, using only the information you give them. What would they need to know?

Anthropic's documentation for Claude describes prompting in similar terms, recommending clear, direct instructions with relevant context.

The parts of a good prompt

1. Context

Explain the situation:

  • Who you are and what your business does.
  • Who the output is for.
  • Any background the task depends on.

"I run a small physiotherapy clinic. A patient emailed to complain that their appointment was cancelled at short notice because a therapist was ill."

2. The task

State clearly what you want:

"Write a reply apologising, explaining what happened, and offering two alternative appointment times."

3. Constraints

Set boundaries:

  • Length: "under 150 words".
  • Tone: "warm and professional, not overly formal".
  • Things to include or avoid: "do not offer a discount; do mention that we will call if they prefer".

4. Examples

If you have an example of the style you want, include it. Examples are often more effective than descriptions.

5. Output format

Say how you want the result:

  • "Give me three subject line options, then the email."
  • "Return a table with columns for task, owner and deadline."
  • "Use headings and bullet points."

Iterate, do not start over

The first response is a starting point. Refine it:

  • "Make it shorter and more direct."
  • "The second paragraph sounds defensive; rewrite it."
  • "Add a line about our new online booking system."

Short follow-up instructions are usually faster than rewriting the original prompt.

Ask for reasoning on complex tasks

For analysis or decisions, ask the assistant to explain its reasoning or list its assumptions before giving a conclusion. This makes errors easier to spot and helps you judge whether to trust the result.

"Before recommending a pricing structure, list the assumptions you're making about our costs and customers."

Break large tasks into steps

Complex tasks work better in stages:

  1. "Here are notes from three client calls. Summarise the main problems each client mentioned."
  2. "Group these problems into themes."
  3. "For each theme, suggest a blog post title that answers the underlying question."

Each step can be checked before moving on.

Provide source material

When the output must be accurate, give the assistant the source material rather than relying on its general knowledge:

  • Paste your actual policy text and ask for a plain-language summary.
  • Provide your product details and ask for a description.
  • Share the meeting notes and ask for action items.

This reduces invented details. Still, check that the summary reflects the source faithfully.

Templates for common tasks

Summarising

"Summarise the document below for [audience] in [length]. Focus on [what matters]. List any deadlines or actions separately. If anything is unclear, say so rather than guessing."

Drafting a reply

"I need to reply to the message below. Context: [background]. Goal: [what the reply should achieve]. Tone: [tone]. Length: [length]. Do not [things to avoid]."

Creating a checklist

"Create a checklist for [task] for [who]. Group items by [stage]. Keep each item to one line. Mark items that need a specialist."

Rewriting

"Rewrite the text below for [audience]. Keep the meaning, use plain language, keep it under [length], and keep any figures exactly as written."

Checking the output

Good prompts reduce errors but do not eliminate them. Always:

  • Verify facts, figures, names and dates.
  • Check that nothing important was dropped from a summary.
  • Make sure the tone suits the recipient.
  • Remove anything you cannot stand behind.

For legal, financial, medical or safety topics, use AI only as a starting point and rely on qualified sources. Frameworks such as the NIST AI Risk Management Framework highlight accuracy and accountability as core concerns.

Protecting information

Before pasting anything into an AI tool, consider whether it contains personal, confidential or sensitive information. Remove names and identifying details when they are not needed, and follow your organisation's policy. Our guide to AI tools for small business covers data protection in more detail.

A worked example

A freelance web developer used to ask an AI assistant to "write a proposal for a website". The results were long, generic and full of features the client had not asked for.

She began using a structured prompt instead: the client's business and goals from her discovery call notes, the agreed scope, her pricing, her usual proposal structure, a sample of a past proposal she liked, and a request to keep it under two pages and flag any assumptions. The first draft now needed only light editing, and she checked every detail against her notes before sending. Our guide to freelance client onboarding covers what to put in a proposal.

Building a prompt library

Save prompts that work well in a shared document, organised by task. Over time, your team builds a library that produces consistent results and saves time on routine work.

Learning more

Free courses on our learning platform, including Anthropic's prompt engineering tutorial, go deeper into techniques such as examples, step-by-step reasoning and structured output.

Frequently asked questions

Do longer prompts always work better?

Not always. Prompts should include the context and constraints that matter, without unnecessary detail.

Should I use the same prompt for different AI tools?

Good prompts usually transfer well, but results vary. Test and adjust for each tool.

How do I stop the AI from making things up?

Provide source material, ask it to say when it is unsure, and verify the output. No prompt removes the need to check.

Is prompt writing a skill worth learning?

Yes. Clear instructions are useful for AI tools and for working with people, and the skill improves with practice.