Prompting · Practical guide
How to Improve an AI Answer: Critique, Revise and Verify
A polished answer can still miss the task or contain a mistake. Use this practical review workflow to find specific weaknesses, request a targeted revision and independently check the result.
Set the criteria before asking for a review
For this DYAP exercise, a useful answer must satisfy four checks: answer the question, use correct evidence, explain assumptions and suit the reader. Write these down before judging the draft. A longer response does not automatically meet these checks.
Clear instructions, an explicit output format and relevant examples can help guide a model. Anthropic’s prompting documentation explains these techniques; the classroom workflow below is an original illustrative exercise.
Worked example: identify the actual mistake
Fictional draft: “Of 200 packets sent, 12 were lost. Packet loss is 12%, so delivery is 88%. This proves the network is unsuitable.” The task was to calculate loss and explain what the measurement tells us.
The draft confuses a count with a percentage and makes an unsupported suitability judgement. Calculate loss independently: 12 ÷ 200 × 100 = 6%. Received packets = 188; delivery = 94%, assuming every sent packet is counted as received or lost. Deciding suitability also requires the application’s requirements and measurement conditions.
Ask for a specific revision
Request a table with the original claim, the problem, the proposed correction and the evidence needed. Then ask for a revised answer under 80 words. Keep the original draft so you can compare what changed.
Example revised answer: “In this measurement, 12 of 200 packets were lost, giving 6% packet loss. The remaining 188 packets represent 94% delivery. These figures assume each sent packet is classified as received or lost. Whether this is acceptable depends on the application and test conditions.”
Verify the revision yourself
Recalculate both percentages using a calculator, check the counts and confirm that the assumptions match the scenario. If a revision adds a source or a factual claim, inspect that source directly. Asking the same model to approve its answer is not independent evidence.
Use a stop rule: accept the draft when all four criteria are met and the important claims have been checked. If evidence is missing, label the uncertainty or remove the claim instead of repeating “make it better”.
Try a 15-minute practice task
Spend three minutes drafting a fictional explanation, four minutes identifying two specific weaknesses, four minutes requesting a revision and four minutes checking it. Avoid personal or confidential information.
Record the original error, the correction and how you checked it. You succeed when you can explain why the revised answer is better using evidence. Save your adapted prompt in My DYAP and try it on a second example.
Reusable prompt
Select and copy this template, or customise your own using the free builder.
Review the draft below for a beginner audience. Check task alignment, factual accuracy, assumptions and clarity. Return a table: original claim | problem | suggested correction | evidence I must check. Then write a revised answer under 80 words. Preserve supported facts, identify uncertainty and do not invent sources. Task: [insert task]. Draft: [insert non-sensitive draft]. I will independently verify the final answer.Open the prompt builder →
Pause · Think · Check
Test your understanding
Choose one answer per question, then check your reasoning. This is practice; scores are not saved.