To write an AI prompt, describe the task, supply the facts the answer must use and specify what to do with missing information. Then ask for a format you can check. Try that structure on the customer reply below before adapting it to your own work.
Compare two instructions using the same facts
This is a fictional customer-service exercise. The source brief is identical for both instructions:
A customer asks when order A104 will arrive. The carrier status is “delayed.” There is no revised delivery date. The support team can check the carrier status again tomorrow. No refund or discount has been approved.
The first instruction says only: “Write a friendly reply to the customer.” That leaves the factual boundaries and next step unstated. For example, the following invented draft sounds helpful but fails the brief:
Sorry about the delay. Your order will arrive tomorrow, and we have applied a discount for the inconvenience.
Tomorrow is the next check, not a delivery promise. The discount is also unsupported. These are authored comparison examples, not measured outputs from an AI service.
Use this more specific instruction with the same source:
Draft a reply to the customer about order A104.
Use only these facts:
- The carrier status is delayed.
- No revised delivery date is available.
- Our team can check the carrier status again tomorrow.
- No refund or discount has been approved.
Explain what is known and the next action in plain language.
Do not promise a delivery date, refund or discount.
Write one short paragraph addressed to the customer.A reference reply that satisfies those constraints is:
I’m sorry your order A104 is delayed. The carrier has not provided a revised delivery date. We can check the carrier status again tomorrow.
This reply is written for the exercise. Your tool may use different wording, so check its claims instead of expecting an exact match.
Start with the task
Describe one action and one intended result. “Write a proposal draft for account review” is easier to assess than “help with sales.” If you need research, drafting and editing, split them into separate passes.
Add only the context that changes the answer
Include confirmed facts, the audience and the decision the output should support. Keep guesses out of the context. Label missing information so the model can leave a placeholder instead of inventing a detail.
State the boundaries
Tell the model what it must not invent, which sources it may use and which details need human approval. For a sales document, prices, customer results, legal terms and delivery dates deserve an explicit check.
Define the output
Ask for a format you can scan: named sections, a table with fixed columns or a short list. A fixed format makes omissions easier to spot and helps another person review the work.
Review the response against the brief
Check whether the response completed the stated task, used only supplied facts, followed the constraints and matched the requested format. Then inspect every name, number, date and factual claim against the source material.
For this exercise, review the answer against these checks:
| Check | Accept | Correct before use |
|---|---|---|
| Delivery status | Says delayed | Says shipped, delivered or arriving tomorrow |
| Revised date | Says unavailable or does not invent one | Adds a specific date |
| Next action | Offers another status check tomorrow | Treats tomorrow as a delivery commitment |
| Remedy | Does not promise a refund or discount | Adds compensation |
| Format | One short customer-facing paragraph | Adds internal instructions to the message |
Reuse the structure for your own task
Replace the bracketed fields. Keep facts separate from the instructions that tell the model what to do with them.
Task: [one action and the result you need]
Reader: [who will use the answer]
Source facts: [approved information only]
Unknowns: [what has not been decided or confirmed]
Boundaries: [what the answer must not claim or do]
Output: [format and required sections]
Review: [the checks you will apply before using it]Run the prompt once, inspect the answer and revise the part connected to an observed error. If the reply promises a date, make the difference between a status check and a delivery commitment explicit. Save both answers to see whether the change helped on this example.
Questions about writing prompts
Should every prompt be long?
Use enough detail to define the task and its boundaries. The delivery exercise needs a few facts and a clear prohibition on promises; adding an unrelated persona or a long company history would not help you check those requirements.
Can I ask the model to check its own answer?
You can ask it to list the source for each claim, but still compare that list with the original brief. A second generated explanation is not independent evidence.
What if a fact is missing?
Ask for a placeholder, a question or an explicit “not provided,” depending on the output. In the customer reply, saying that a delivery date is unavailable is more useful than inserting a placeholder that a customer would see.
Build your own version in the Prompt Builder. When the same prompt needs to work on several different inputs, continue with testing and revising prompts.