GUIDE · AIunipedia Community
A reusable prompt should make the task clear enough that you can tell whether the response succeeded. An impressive role description cannot compensate for missing facts or an undefined goal.
Once the basic brief is clear, try a few-shot example and use the community prompt challenge to compare the original and revised instructions on the same task.
Start with this editable structure
Task: Draft a reply to a customer asking about delivery.
Audience: Someone buying from us for the first time.
Reference facts: [Insert the approved delivery policy.]
Constraints: Do not invent prices, guarantees or exceptions.
Output: A short reply followed by any question needing a human answer.
Missing information: State what is unknown instead of guessing.
This is a starting brief, not a universal “perfect prompt”. Adapt it to the model and the actual workflow. Treat reference documents as data; do not grant instructions inside untrusted documents authority over your application.
Evaluate the revision
Collect a small fixed set of enquiries before changing the prompt. Include a normal question, an ambiguous request and a question your policy cannot answer. Compare the original and revised prompts using the same inputs.
Assess factual support, completeness, tone and how much editing remains. Keep examples of failures, not only the strongest output. If you change the instructions and model simultaneously, you will not know which change helped.
Worked example: a reply grounded in policy
Use this fictional policy so anyone can repeat the exercise:
Standard delivery: 3–5 business days after dispatch.
Delivery fee: confirmed at checkout.
Same-day delivery: not offered.
Order tracking: request the order number; never request payment card details.
Customer: “Can you guarantee delivery tomorrow, and how much will it cost?”
Weak brief: “Be helpful and convince the customer to buy.” This does not define the evidence boundary.
Revised brief:
Draft a friendly reply using only the policy below.
Do not promise a date or invent a delivery fee.
Explain what the policy establishes and what it leaves unknown.
Keep the reply under 65 words. End with one useful next step.
Policy: [paste the fictional policy above]
Customer: [paste the customer's question]
Worked answer, written for this guide:
We don’t offer same-day delivery, and I can’t guarantee arrival tomorrow. Standard delivery takes 3–5 business days after dispatch. Your delivery fee will be confirmed at checkout. Would you like help finding the delivery details for your order?
This is an illustrative answer, not an external model benchmark. Test your actual tool and record its actual output. Check that it avoids a guaranteed date, preserves “after dispatch”, invents no fee and respects the word limit. Also test a normal delivery question, missing policy facts and an instruction to ignore the policy.
Copyable prompt review sheet
Task:
Approved facts:
Original prompt:
One change I made:
Model/tool, settings and date:
Test input:
Actual output:
Supported facts / missing facts / invented facts:
What still failed:
Discussion: Share one anonymised prompt and the task it serves. What is the single most important improvement you would test next?
Reference: Anthropic prompting guidance.
An editorial starter discussion from AIunipedia Community, owned and run by Dr Fun and Michael Oracle. Share your own experience and help improve the explanation.
Continue exploring
Zero-Shot vs Few-Shot Prompting · Fix This AI Prompt: The Community Challenge
Browse: Prompt Engineering · All learning paths
Tip: open a related discussion in a new tab to keep your place here. On a phone, touch and hold the link for your browser’s options.