Prompt engineering means wording instructions to a language model so that the result is reliably usable; six techniques are enough for most business tasks: role, context, a clear task with format, examples, rules and iteration. It is not about magic spells but about complete, checkable assignments, the kind you would give a new colleague.
Which six techniques work in business?
The table shows the techniques with their effect and a wording to adapt.
| Technique | Effect | Example wording |
|---|---|---|
| 1. Give a role | Sets the professional view and tone | "You are a customer service agent for a roofing firm." |
| 2. Supply context | Prevents guessing | "Here is the enquiry and our price list: …" |
| 3. Specify task and format | Makes the result checkable | "Write a reply in formal address, at most 120 words, with three appointment suggestions." |
| 4. Show examples | Fixes style and structure | "This is what a good reply looks like: …" |
| 5. Set rules | Limits errors | "Do not invent prices. If something is missing, ask." |
| 6. Iterate | Improves step by step | "Shorter, friendlier, answer first, then the reasoning." |
The providers describe these basics in their guides, for example Anthropic and OpenAI.
What does a usable template look like?
A template puts the techniques into a fixed order. You can adapt it for recurring tasks:
Role: You are a customer service agent for a roofing firm.
Task: Answer the following customer enquiry.
Context: [insert enquiry] and [insert relevant prices or services]
Format: Formal address, at most 120 words, answer first, then an appointment suggestion.
Rules: Do not invent prices or promises. If details are missing, ask a question.Adapt the square brackets and add only what the task needs. Longer instructions are not automatically better: every line should answer a question the model would otherwise have to guess.
File templates with purpose and date and test them with real examples. A prompt that works on three cases need not work on the fourth. How to compare several models on the same tasks is described in How to evaluate a new AI model.
What can prompts not do?
A good prompt improves the likelihood of usable answers but does not guarantee them. The model can still invent facts, ignore instructions or lose details in long texts. So check figures, names and quotations before you pass results on. On dealing with errors, read AI hallucinations: what they are and how to protect your business.
Also beware of third-party text in the prompt. If pasted emails or web pages contain instructions of their own, the model may follow them instead of yours (prompt injection). Separate instruction and third-party text clearly and check results before they trigger actions.
How do you bring prompts into daily work?
Collect proven prompts in a shared library, with purpose, date, responsible person and an example result. Decide which data may go in. Review important prompts when you switch models. If you want to know more about everyday use, find ideas in AI for small businesses: which tasks can AI actually take over?.
- Goal and checkable result of the task formulated
- Role, context, format and rules named in the template
- One or two examples of the desired answer added
- Instruction for unclear cases included
- Tested with several real examples and improved
- Prompt filed with purpose, date and owner
- Data protection: no confidential data without approval
Conclusion: clear assignments, checked results
Prompt engineering is craftsmanship in wording, not secret knowledge. Begin with the template, test with real cases and check the results. If you want to turn recurring tasks into a fixed process, see our AI automation service or describe your task.
