Prompting: how to write a good prompt for your work

A good prompt has five building blocks: a role, context, a task, an example and a format. You don't always need all five. But if something is missing from the answer, it's almost always down to one of these five.

1. Role: who should the AI be?

Give the AI a role that fits the task. 'You are an experienced inside sales employee at a wholesaler of installation materials.' With that, the model picks a tone, a field and a level. Without a role you get the voice of a helpful, general assistant, and that rarely sounds like you.

2. Context: what does the AI need to know?

The building block that gets forgotten most often, and the one that makes the most difference. Who is it for? What happened? What does the reader already know? What matters to you? 'The customer is a regular buyer, the delivery is three days late because of a mistake by our carrier, and we'd really like to keep him.'

3. Task: what needs to happen?

One clear instruction, with a verb. Write, summarize, compare, convert, give three options. If you ask for more than one thing, list them one below the other, in the order you want them.

4. Example: what does good look like?

An example says more than ten adjectives. 'Friendly but businesslike' can mean anything. Paste an email you wrote earlier and liked, and say: in this tone. Or give the structure you want. Do take the names out first.

5. Format: how do you want it back?

Length, structure and tone. '120 words maximum.' 'A table with three columns.' 'Conclusion first, then the details.' 'Address the customer formally.' The more concrete, the less you rewrite afterwards.

Putting it all together

You are a customer service employee at a wholesaler of installation materials. A regular customer received his delivery three days late, because of a mistake by our carrier. He's irritated but businesslike, and we want to keep him. Write a reply to his email (below) in which you apologize, briefly explain what went wrong and offer to cover the shipping costs of his next order. Use the tone of the sample email below. 120 words maximum, formal, no exclamation marks.

That seems like a lot of work for one email. It's half a minute of typing. And it saves you three rounds of 'no, shorter', 'no, less over the top' and 'no, we don't give discounts'.

The first answer is a first draft, not a finished product. You get the best results by following up in the same conversation: shorter, more formal, a different opening, three variants. Within one conversation, the AI remembers what you said earlier.

Follow-up questions that almost always work:

• 'Make it half as long.'
• 'Give three variants of the opening sentence.'
• 'What is a reader who sees this for the first time missing?'
• 'Rewrite it, but for the managing director instead of the buyer.'
• 'What are the weak points of this text?'

And the handiest of all goes at the start: 'Ask me the questions you need first, before you begin.' Then the AI asks for the context you forgot. Ideal when you're not sure where to start.

Is the conversation getting long and is the answer losing the thread? Then start a new conversation, with your instruction and the best version so far. That works better than endless correcting.

The most common mistake is too little context. After that come: wanting everything in one prompt, not giving an example, using the first answer straight away and treating AI as a source of facts. The good news: you can unlearn them in a week.

• Too short. 'Write a job ad.' The AI doesn't know which role, which company, which tone and for whom. So you get a job ad that fits every company in the Netherlands, and therefore none of them.
• Everything at once. Summarizing a report, turning it into an email, translating that and adding a LinkedIn post. Do it in steps, so you can correct each step.
• No example. 'In our house style' tells the AI nothing if it has never seen that house style.
• Using the first answer. Ask follow-up questions. See above.
• Using AI as a source of facts. A language model can convincingly make up a legal article, a figure or a source. Use AI to write, organize and summarize what you provide, and always check facts.
• Asking for what it can't know. 'What is our delivery time for this item?' The model doesn't know that, unless you tell it. Otherwise it guesses.
• Not checking what goes out. You're responsible for what you send, not the AI.

Passwords, citizen service numbers (BSN), bank details of individuals and medical or other sensitive personal data never go in a prompt. Customer data and confidential business information only go in a business account with a data processing agreement, according to your agreements, and preferably without names.

The biggest risk is in the free version and in personal accounts. There, as an organization, you have no agreements about what happens to your data. One habit prevents most problems: take the names out. 'Jan de Vries from Utrecht is complaining about his delivery of 12 March' becomes 'Customer A is complaining about a late delivery'. The answer doesn't get any worse.

In short:

• Always fine: public information, your own texts without names, general questions.
• Only in the business tool, according to the agreements: internal documents, customer emails, quotes and figures.
• Never: passwords and login details, citizen service numbers (BSN), bank details of individuals, medical and other special categories of personal data.

You write a good prompt once and use it a hundred times. Keep the prompts that work in a shared document or a shared folder, per task, with the parts you fill in yourself between brackets. That's a prompt library.

How to start:

1. 01Pick the five tasks that come up most often for each team. Customer emails, quotes, minutes, reports.
2. 02Write one strong prompt per task, with the five building blocks and a good example.
3. 03Test it on real work from this week. If it works for three different cases, it can go into the library.
4. 04Appoint an owner. Someone who adds new prompts and clears out old ones.

Our freebies page has free prompts to get you started. Want the whole team to learn it in one go, on your own work? A half-day workshop at your location costs € 4,500, excluding VAT. The forms of training we offer are on AI training for your team.

If you paste the same explanation, the same files and the same example every time, it's time for your own assistant. If you push the same work through the chat dozens of times a week and retype the result by hand into another system, it's time for a workflow.

Prompting delivers personal gains. Every employee gets a little faster. But the work itself doesn't change: someone opens the chat, pastes, reads, copies and pastes back. Two signs that you're ready for the next step:

• You keep pasting the same context. The price list, the house style, three sample quotes. Put them in a custom GPT, Claude Project or Copilot agent once, and nobody has to paste them again. That's build layer 2, and your team can do it itself. Read build your own AI assistant.
• You copy between systems. The email comes in, you paste it into the chat and retype the answer into the ERP. A workflow can do that automatically, with a human who checks. That's build layer 3, and we build that.