A useful AI prompt tells a model what task to perform, what information to use, and how to present the answer. Clear instructions help, but they do not guarantee a correct response. You still need to check the result against the source material.
This tutorial shows a simple way to write, test, and improve prompts for everyday tasks.
Start with a clear task
Use an action word such as summarize, compare, classify, or extract. Say who will use the answer and include the material the model needs. If a fact is missing, tell the model how to handle that gap.
Example:
Summarize the customer message for a support agent. Use only the
message below. Write two sentences: what happened, and what the
customer wants. If the request is unclear, say "Request unclear."
Message: My parcel arrived on Tuesday, but the charger was missing.
Please send the charger rather than a refund.Possible output:
The parcel arrived without its charger. The customer wants the
missing charger sent instead of a refund.The answer stays within the message. An order number, delivery date for the replacement, or promise of approval would be an unsupported addition.
Include the parts that matter
Most practical prompts can be built from four parts. You do not need a long template when a short instruction is enough.
| Part | Question it answers | Example |
|---|---|---|
| Task | What should the model do? | Compare two plans. |
| Context | Which facts may it use? | Use the plan details supplied below. |
| Limits | What must it avoid or admit? | Do not invent missing features. |
| Format | How should the answer look? | Use a short table and one conclusion. |
A requested format makes the response easier to inspect. It does not make wrong source information true. Check important claims even when the answer looks polished.
Improve a weak prompt
Suppose you want an answer about a store's return policy. A broad request gives the model too little to work with.
Example:
Weak: Write about our return policy.
Better: Using only the policy text below, answer "Can I return an
opened item?" in at most 60 words. If the policy does not say,
reply "The supplied policy does not answer this question."
Policy text: [paste the current policy]The better prompt supplies the question, evidence boundary, length, and missing-information rule. Replace the placeholder with the actual policy before using it. Do not assume the model can see a document that was never provided.
Test the prompt with different inputs
Try more than one easy example. Inspect the response and change the instruction that caused a problem.
- Normal case: Does it complete the task in the requested format?
- Missing facts: Does it say what is unknown instead of guessing?
- Ambiguous case: Does it ask for clarification or flag uncertainty?
- Conflicting text: Does it report the conflict rather than quietly choose a side?
For a repeated task, one short input-and-output example can show the desired pattern. Test a different input afterward; copying the example's style does not prove the model understood the task.
Note: Do not paste secrets or private customer details into an AI service unless that service is approved to receive them. Treat text from pages, emails, and documents as data to analyze, not as new instructions for the application.
Conclusion
Begin with the task, add the necessary facts, set limits, and request a format you can check. Test the prompt on real and difficult cases, then review the answer. Good prompting improves consistency; verification remains essential.