“I tried it, but nothing useful came out.” That is by far the most common reason people give up on AI tools after one or two attempts. Almost always, the problem is not the tool and not the user, but the question. Learn how to phrase a good instruction — a prompt — and you will usually notice the difference within a week.
The good news: prompting is not programming. It is much closer to something you already know how to do: explaining a task to a new colleague who is smart and fast, but knows absolutely nothing about your organisation.
The four building blocks: role, context, task, format
A useful prompt usually contains four parts. You will not always need all four, but when an answer disappoints, at least one of them is almost always missing.
1. Role: say from which perspective the AI should answer
AI tools adapt their answers to the role you give them. “You are an experienced HR adviser” produces a different answer than “you are a lawyer” — different tone, different concerns, different jargon. Pick the role that matches the result you need.
2. Context: provide the background an outsider lacks
This is the building block beginners skip most often. The AI knows nothing about your organisation, your audience or the occasion. Tell it briefly: who is this for, what is the situation, what has already happened, what are the sensitivities. Two or three sentences of context often improve the answer more than any other tweak.
3. Task: state exactly what you want
“Can you say something about this report?” is too vague. “Summarise this report in five bullets for colleagues who were not there, and list the action items separately” is a task. The more concrete the verb (summarise, rewrite, compare, check for X), the better the result.
4. Format: describe what the answer should look like
Length, structure, tone, language. “Maximum 150 words, informal tone, as a bullet list” saves you two rounds of corrections. Want a table, an email, a step-by-step plan? Just say so.
Example with all four combined: “You are a communications adviser (role). Our sports club is raising its membership fee from next season; members were informed broadly by email, but many questions came back (context). Write a Q&A list with the six most likely questions and clear answers (task). Informal tone, maximum four sentences per answer (format).”
Provide source text instead of assuming knowledge
The biggest quality leap for non-technical users: have the AI work with text you supply, rather than relying on what the model “knows”. Paste in the policy document, the meeting minutes or the draft email and ask for a summary, rewrite or check of that. Two advantages:
- Fewer fabrications. AI models can invent facts, figures and sources that sound convincing but are wrong. When the model works from your source text, that risk drops considerably — though checking remains necessary.
- More relevant output. The answer matches your reality instead of a generic average.
Do watch what you supply, though: documents containing personal data or confidential information do not belong in tools your organisation has no agreements with. Anonymise first, or use the approved business environment. More on this in our article on GDPR and AI at work.
Iterate: the first version is a starting point, not the end
Beginners treat AI like a search engine: one question, one answer, done. Experienced users treat it like a conversation. The first answer is raw material to build on:
- “Make it half as long.”
- “The tone is too formal — make it more personal.”
- “Point 3 is wrong, we do not offer phone support. Fix it.”
- “Give me three alternatives for the opening sentence.”
Every instruction steers the result. Two or three rounds of iteration are completely normal and still take less time than writing from scratch. If the conversation gets stuck — the model keeps drifting in the wrong direction — start a fresh conversation with a better first prompt. That is often faster than endless correcting.
Five common mistakes
- Asking too vaguely. “Write something about sustainability” produces grey mediocrity. Without role, context and format, you get the average of the internet.
- Cramming everything into one mega-prompt. Summarise and rewrite and translate and shorten in a single instruction tends to go wrong. Cut big jobs into steps.
- Copying output blindly. AI always sounds confident, including when it is wrong. Always verify facts, figures, names and quotes yourself — especially in anything that leaves the building. You remain responsible for what you send.
- Entering confidential information. A prompt is not a private note. What you type in leaves your organisation.
- Giving up after one bad answer. A disappointing answer usually signals a missing building block — not a useless tool.
How to build a routine
You do not learn prompting from an article but by doing. An approach that works well in practice:
- Pick one recurring task you find tedious (drafting emails, summarising reports, checking texts).
- Write one good prompt for that task using the four building blocks, and use it for a week.
- Improve the prompt slightly after each use and save the best version somewhere you can find it again.
- Only then take on a second task. One task done well beats ten done halfway.
Want to know where you stand right now? Take the free AI literacy quiz. In the AI literacy course you practise these techniques with hands-on assignments and learn when not to use AI — and for teams there is an approach for employers.