Prompt Writing for Teams: A Practical Starter Guide

The four parts of a useful prompt

A language model predicts a plausible answer from what you give it. A one-line request leaves it to guess the audience, the tone, the length and what counts as done, so it guesses the most generic version. Each of the four parts removes a guess.

Giving context without leaking data

Context is what makes a prompt good, and it is also where data leaks happen. Under GDPR your company is responsible for personal data it sends to an AI vendor, and consumer versions of chat tools may keep or use what you type. Settle which tools are approved first; our guide to an AI usage policy for staff covers that. Then teach the habit of describing rather than copying.

• Replace names, personal identity numbers, addresses and phone numbers with roles: the customer, the tenant, employee A.
• Summarise a long email thread in your own words instead of pasting it whole.
• Never paste health information, union membership or other sensitive categories, even into approved tools, unless your policy explicitly allows it.
• If the task needs the real data, it belongs in an integration built on a business agreement with a data processing agreement, not in a chat window.

Examples, formats and constraints

The fastest way to improve output is to show one example of what good looks like. A short sample of your house tone or a previous reply you were proud of tells the model more than a paragraph of adjectives. Keep examples anonymised.

• Ask for a fixed structure: headings, a table with named columns, or a numbered list with a set length.
• State the language explicitly. Models drift into English or into stiff, translated Swedish unless told otherwise.
• Ask the model to list assumptions and questions before it writes when the task is unclear.
• For facts, prices and legal points, tell it to use only the context you supplied and to say so when that context is not enough.

A shared prompt library

Once a prompt works, it should stop living in one person's chat history. A library can be a shared document, a page in your intranet or a folder of saved prompts in the tool itself. The format matters less than the fields each entry carries.

Start with five to ten prompts for tasks the team does every week, such as replies to common questions, meeting summaries, job ads or first drafts of quotes. A small, maintained library beats a large, abandoned one.

Checking and improving results

Every output is a draft until a person has checked it. Make the checking step explicit in the library entry: what must be verified, by whom, before the text leaves the company. When a prompt gives a poor result, change one part at a time, rerun it and compare, so you learn which part was missing.

• Facts, names, dates and prices checked against the source.
• Tone read aloud once: would you say this to the customer?
• No promises the company has not approved.
• Swedish spelling and terms that match your industry, not literal translations.

Decision rule: if the same prompt is used more than a few times a week and the output always goes into the same system, it is a candidate for automation rather than copy and paste. That is where an AI integration audit earns its fee, because it maps which of those tasks are worth building.

When you do not need outside help: if your team uses AI for occasional drafts and a shared prompt document already works, keep going on your own. An audit only pays off when AI use spreads into customer-facing work or systems holding personal data; if that is where you are, book a short call and we will tell you honestly which it is.

Frequently asked questions