What makes a useful AI work instruction?
A prompt is the task description you give an AI system. Describe the goal, audience, required information and output format as you would when briefing a new colleague. Include approved examples of your own tone where useful.
AI can turn notes into a structured draft. A better instruction cannot supply missing facts: say whether the system should ask a question or flag the gap.
A copyable template for recurring office tasks
Copy the template below and replace the placeholders with information you are permitted to use in your chosen AI tool.
“Task: [specific result]. Audience and purpose: [who needs it and why]. Source material: [insert approved information]. Format, language and tone: [e.g. short email, English, neutral]. Boundaries: use only facts in the source material. Do not invent names, numbers, dates or commitments. List missing or conflicting information separately under ‘Internal questions’. Output: the draft first, then the internal questions.”
The template makes it easier to compare the instruction with the result. You still need to check the draft against the source material.
Filled in: turning call notes into an email draft
Fictional example; the output below illustrates a desired result, not a recorded model run.
- Source material: Ms Keller asks about 20 F4 replacement filters. She wants to know whether delivery by Friday is possible. Stock has not yet been checked. No price is available. Next internal step: check availability and delivery options with the warehouse.
- Instruction: Use these notes to write a short, neutral email in English to Ms Keller. Acknowledge her enquiry about 20 F4 replacement filters and say that we are checking the delivery options. Use only the supplied facts. Do not confirm a delivery date or quote a price. Output the email first, followed by a separate “Internal questions” section listing the missing information.
- Expected email draft: “Dear Ms Keller, thank you for your enquiry about 20 F4 replacement filters. We are checking whether delivery by Friday is possible. Kind regards”
- Expected internal questions: “Are 20 F4 replacement filters available? Is delivery by Friday possible? What is the price?” These questions do not belong in the customer email.
- Check before sending: are the name, quantity and filter type correct? Does Friday remain a customer request rather than a commitment? Has a price been added? After checking, add the sender’s name and copy only the email draft.
Test templates and improve them deliberately
If the draft says “We will deliver on Friday”, a request has become an unsupported commitment. Correct that statement before using the text. For the next attempt, make the instruction more specific: “Friday is only the requested delivery date. State explicitly that the delivery options are still being checked.” Generate another draft and compare it with the same notes. Improved wording does not confirm that delivery is possible.
Then test the saved template with an ordinary case, one with missing details and one with conflicting information. Change only one element at a time and keep the inputs for comparison. This also lets the team check which errors occur after a model or template change.
- Do names, numbers and commitments match the input?
- Are all requested points present?
- Does missing or conflicting information remain visible?
- Are internal questions separate from the recipient’s text?
Requirements and limits
For recurring work, save the template together with a permitted sample input and a professionally checked sample answer. Decide who will review the output before use. If essential facts are missing or contradictory, clarify them with the responsible person first. Adding more text to the prompt does not replace that conversation.