HomeOperational AICreative AITrainingBlogContactStart a project
← All articles

Skills3 October 2026 · 6 min read

Writing prompts at work: a starter guide

How to brief an AI model clearly, build a shared prompt library and get consistent results across a team.

A prompt is the instruction you give an AI model. At work, a good prompt works like a good brief: the kind you would give a capable new colleague who is keen to help and knows nothing yet about your business, your customers or the task.

This guide covers how to write that brief, how to turn your best prompts into a shared library, and how to get consistent results when a whole team is using AI.

Brief the model like a new colleague

Anthropic, the company behind Claude, offers a simple test in its prompting guidance: show your prompt to a colleague who has minimal context on the task and ask them to follow it. If they would be confused, the model probably will be too.

That test catches the most common gap in workplace prompts, which is missing context. "Write a reply to this customer" leaves the model guessing about tone, policy, length and what outcome you want. A colleague would ask those questions. The model will usually just make a reasonable guess, so it pays to answer them up front.

The four parts of a clear prompt

Google and Microsoft describe good prompts in very similar terms.

Google's prompting guide for Gemini in Workspace names four areas: persona, task, context and format. Microsoft's guidance for Copilot also uses four parts: the goal, the context, the expectations and the source. Both say you do not need every part every time, and Google notes that the task, with a clear verb, is the most important element.

Put together, a strong workplace prompt usually covers:

  1. The task you want done, starting with a verb such as draft, summarise, compare, rewrite or list.
  2. The context: who it is for, why you need it, and anything the model needs to know about the situation.
  3. The material the model should work from, such as a document, notes, data or an email thread.
  4. The format you want, including length, structure, tone and layout: three bullet points, a table, a 150 word email in plain British English.
  5. Optionally, a role. Asking the model to act as, say, an experienced customer service manager can help set the tone.

Here is the difference in practice.

A quick prompt:

Write a reply to this complaint.

A briefed prompt:

You are replying on behalf of [your company], a [what you do] business in [your town]. Draft a reply to the customer email below. Their order was delivered two days late. We have apologised by phone and are offering [what you are offering]. Keep it warm, brief and in plain British English, under 120 words, and sign off as "The Customer Care Team". Customer email: [paste here]

The second version takes a minute longer to write and usually saves several rounds of editing.

Give examples and keep material separate

Examples are a very effective way to show the model what you want. If you have a past email, report or product description that hits the right note, paste it in and say that it is an example of the style you want. Anthropic's guidance recommends examples that closely mirror your real task and vary enough that the model picks up the pattern you intend.

When a prompt contains several kinds of content, label each part clearly. Headings such as "Instructions", "Background" and "Customer email" help the model tell your instructions apart from the material it is working on. Anthropic's documentation suggests wrapping each kind of content in its own tags, such as <instructions> and <context>, for the same reason.

Phrase instructions positively and refine as you go

Tell the model what you want it to do, as Microsoft's guidance for Copilot recommends. An instruction such as "Write in short sentences with everyday words" gives the model a clear target.

Treat the first answer as a draft. Follow up in the same conversation: "Shorter", "More formal", "Add a line about our returns policy", "Turn this into a table". Each refinement teaches you something about what the original prompt was missing, and that is how good prompts for your library are found.

Build a shared prompt library

Once someone finds a prompt that reliably produces good results, the whole team should be able to use it. A prompt library is a shared collection of tested prompts for your recurring tasks.

For each entry, record:

  • A name and purpose, such as "Weekly sales summary for the Monday meeting".
  • The prompt itself, with placeholders in square brackets for the parts that change each time, such as [paste figures here] or [customer name].
  • Which tool it is for, since prompts can behave differently in different tools.
  • An example of good output, so people know what to expect.
  • An owner, who keeps it up to date.
  • The date it was last checked.

Start small. Five well tested prompts for your most common tasks make an excellent first library, and a shared document or spreadsheet is enough to hold them.

Some tools also have sharing built in. In Microsoft 365 Copilot, for example, you can save prompts in Copilot Prompt Gallery and share them with a team in Microsoft Teams, where colleagues find them under a team prompts tab. Check what your own tools offer, and what your administrator has switched on.

Getting consistent results across a team

Teams get consistent results when everyone starts from the same prompts and the same review step. A few habits help:

  1. Keep a one page house style document covering your tone of voice, spelling (British English), terms you always use and terms you avoid. Attach or paste it into prompts for anything customer facing.
  2. Encourage people to start from the library prompt for a task, then adapt it.
  3. Agree which outputs need a second pair of eyes before they leave the business, such as anything with prices, commitments or legal wording.
  4. When someone improves a library prompt, they tell the owner, who tests it and updates the entry.
  5. Review the library regularly. AI tools update often, and a quarterly check keeps prompts working well.

Keep data safe while you prompt

Prompting safely means following your data rules. Use the AI tools your organisation has approved, and keep to your policy on what may go into them. Where you can get the same result without real names, account numbers or other personal details, leave them out or use placeholders. Always read and check the output before you rely on it or send it to a customer, especially facts, figures and anything that commits the business.

What to do next

Pick the three tasks your team asks AI to help with most often. Write a briefed prompt for each using the structure above, test it a few times, and put the results in a shared document. That is the start of your library.

If you would like your team to build these skills together on their own work, our AI training page has details of the workplace sessions we are preparing, and you can register your interest there.

Sources

Read next

More from the blog.

Want to put this into practice in your business?