Prompting Basics: How Office Teams Get Useful Answers From AI

Key Takeaways

• A good prompt is a brief with everything needed: task, context, material, audience, format and limits.
• Weak answers usually come from missing context. Give the AI what a new colleague would need for the same job.
• Improve the first answer by saying exactly what's wrong, or let the AI ask you questions first.
• Check names, numbers and dates, and anything the AI added that wasn't in your material.
• Save prompts that work as shared templates, so the whole team benefits.

A good prompt is a clear brief: it tells the AI what to do, for whom, with which material and in what form. When the answers are weak, the tool is rarely to blame. Usually the context is missing. Give the AI what a new colleague would need for the same task, check the result before you use it, and you'll get useful drafts for most everyday office work.

What is a prompt, and why does the wording matter?

A prompt is the instruction you type into an AI assistant such as ChatGPT, Microsoft 365 Copilot, Claude or Gemini. The model only knows what's in the prompt and in any files you attach, plus the general knowledge it picked up in training. It doesn't know your client, your tone or the emails that came before. Where something is missing, it fills the gap with whatever is most likely, which usually means most generic. That's why “Write a reply to this customer” produces a polite email that says nothing. Picture a capable temp on their first day: quick and a good writer, but knowing nothing about your business. That's how to brief the AI.

What belongs in a good prompt?

Good prompts are built from a few parts. Not every prompt needs all six. The more the result matters, the more of them you should include.

1. Task.One clear verb up front: summarise, draft, compare, list, check, rewrite. One task per prompt works better than three at once.
2. Context.Who you are, what it's about and what has happened so far. For example: “I work at a fiduciary firm. The client has sent incomplete documents for the year-end accounts.”
3. Material.The text the AI should work from: the email thread, the minutes, an extract from the contract, as long as your rules allow it. Tell it to stick to that material.
4. Audience and tone.Who will read the result: a long-standing client, a tenant, the partners' meeting. Then the tone: formal, friendly, brief, and Swiss spelling if the text is in German.
5. Format and length.What the result should look like: bullet points, a table, a subject line with three short paragraphs, no more than half a page.
6. Limits.What the AI must not do: promise dates, quote prices, make things up. And what it should do when something is missing: ask, or mark the gap.

What does a weak prompt look like, and what does a good one look like?

An example from a property management office: a tenant emails to say their dishwasher has stopped working. The weak prompt is “Write a reply to this tenant.” What comes back is a friendly email with no substance or, worse, one that promises a repair date that doesn't exist. The good prompt reads something like this: “You work at a property management company. Draft a reply to the email below. Thank the tenant for letting us know, confirm that we've asked a service technician to get in touch about an appointment, and ask for the model number on the rating plate inside the door. Formal and friendly, no more than 120 words. Don't mention a date. Email: […]”

The second prompt is longer, but it contains what the AI can't know: that a technician has been booked, which detail is still missing, how you write and what must not appear in the reply. You'll barely need to change the result. If you write emails like this every day, it's worth setting the prompt up properly once and saving it as a template.

How do you get from the first answer to a good one?

The first answer is a draft. Treat working with AI as a conversation with someone who learns quickly when you tell them clearly what's missing.

• Say exactly what's wrong: “shorter”, “less formal”, “the second point is wrong, the deadline is the end of March”. “Make it better” doesn't help much.
• Let the AI ask first: “Before you write, ask me the questions you need answered for a good draft.” This helps most when you aren't sure yourself what should go in.
• Give an example: paste in an email you think works well and ask the AI to write in the same style.
• Split big tasks: summarise first, then structure, then write. That way you can check each stage.
• Start a new chat when the conversation drifts. Old instructions further up can distort later answers.

How do you check what the AI gives you?

AI writes fluently and convincingly even when it's wrong. Our article on why AI makes things up explains the reasons. For daily work, one rule of thumb is enough: the more a text matters, the more closely you check it. An internal note gets a quick read. A letter to a client or a public authority deserves a careful look.

• Always check names, numbers, amounts, dates and deadlines against the source.
• Watch for anything that wasn't in your material. If the AI adds a rule, a legal provision or a promise, ask where it came from.
• Compare summaries with the original. Missing points are harder to spot than wrong ones.
• Ask the AI to name the passages it relied on and to list where it's unsure.

What should never go into a prompt?

That depends on the tool and on your firm's rules. A business account with contractual commitments on how data is used is a different matter from a free version someone uses privately. Personal data, client data, HR files and anything covered by professional secrecy belong only in tools your firm has approved for them. When in doubt, anonymise: “Client A” instead of the name, and leave out amounts and addresses the task doesn't need. Our guide to writing an AI usage policy shows what your team's rules should cover, and AI and the Swiss FADP sets out the legal framework.

How does a team build a shared prompt library?

Good prompts are knowledge that should belong to the whole team. Rather than everyone working out alone how to summarise a client meeting, save the prompts that work as templates with placeholders like [client name] or [paste email here]. Keep them in one shared place, such as your intranet or a shared document. A good entry contains:

• What the prompt is for, and what it isn't for.
• The template itself, with clearly marked placeholders.
• An example of a good result.
• What to look out for when checking.
• Who maintains the template and when it was last reviewed.

Without someone looking after it, the library goes stale. That job suits the AI champions in your team. Go through the templates every few months, because what works well changes as the tools are updated.

Give the AI what you'd give a new colleague on their first day, and check its work the way you'd check theirs.

What's the fastest way for a team to learn prompting?

Prompting is learned by practising on your own work, and lists of tricks won't replace that. What helps most is a short, hands-on session with your own documents and the tools you've approved, followed by a few weeks in which everyone tries things deliberately and shares the prompts that work. That's how our team training is set up. To make sure the learning is still there months later, read running an AI workshop that sticks. And if everyone in your team currently uses AI differently, the page on AI in your team shows how to turn that into a shared way of working.