Teaching · Practical guide
How to Create Practice Questions with AI—and Check the Answers
Use AI to draft a small question set from approved material, then solve and improve it yourself. This guide takes you from a learning outcome to three checked questions. The examples are fictional illustrations, not classroom-tested results.
Start with evidence of learning
Choose one observable outcome before requesting questions. For this exercise: given a total and a subset, calculate the percentage and explain what the denominator represents. Learners already know division and multiplication. The purpose is low-stakes practice, not a protected examination.
Write a short source card in your own words: percentage = subset ÷ total × 100, where total is greater than zero. The subset and total must count the same kind of thing over the same period. If every item has exactly one of two outcomes, their counts add to the total and their percentages add to 100%. These are the complete rules for our fictional example.
Ask for one interpretation question, one calculation question and one error-analysis question. This varies the evidence learners provide; simply changing numbers three times would practise a narrower skill. If your outcome is performing a procedure, include a demonstration rather than relying only on multiple-choice answers.
Give the model boundaries and an output contract
Paste only material you are allowed to share. Use synthetic data rather than student records, confidential reports or unpublished exam questions. Follow your institution’s approved-tool policy. UNESCO’s guidance stresses human agency, privacy and educational validation; it does not endorse a particular chatbot or this exercise.
Use the reusable prompt below. Request learner questions separately from the answer key, a source-rule reference for each answer, and an explicit “insufficient information” response when the supplied material cannot support an item. These instructions help define the task but do not guarantee compliance.
Keep the first batch to three items so you can inspect every answer. Tell the tool to avoid extra domain facts: our example is arithmetic, so it does not require claims about quality standards, failure causes or acceptable performance.
Worked example: three questions and checked answers
Fictional source data: a training workshop processes 300 test records. Exactly 24 fail a validation check and the remaining records pass. Every record is counted once. “Fail” describes the check outcome; it does not tell us the cause.
Question 1—interpretation: To calculate the percentage of records that fail, which quantity is the denominator? A: the 24 failed records; B: the 300 total records; C: the 276 passed records. Answer: B. The question asks for failed records as a share of all records, so the reference total is 300.
Question 2—calculation: What percentage fails? A: 24%; B: 92%; C: 8%. Answer: C, because 24 ÷ 300 × 100 = 8%. Option A confuses the count with a percentage; B is the percentage that passes. These are plausible error routes, not evidence that any particular learner holds a misconception.
Question 3—error analysis: A draft answer says, “24 records failed, so failure is 24% and passing is 76%.” Rewrite it and explain the mistake. Example answer: “24 of 300 records failed, giving 8% failure. The other 276 passed, giving 92% passing. The draft treated the failed-record count as a percentage instead of dividing by the total.”
Check: 24 + 276 = 300 and 8% + 92% = 100%. The complement calculation is valid here because the scenario defines two exhaustive, non-overlapping outcomes. Do not assume it for data that also contain pending or unclassified records.
Audit the questions before sharing
Carnegie Mellon’s Eberly Center recommends aligning questions with learning objectives, using clear wording and giving objective items one best answer. Its guidance also recommends plausible wrong options and avoiding clues from grammar or option length. Use those principles when reviewing the draft.
First hide the generated key and solve each item independently. Check arithmetic with a calculator, then compare the key and explanation. Inspect every option under the stated assumptions; an answer key can be wrong even when the question reads smoothly.
For Question 2, do not add “0.08 as a fraction of the total” as another option while keeping a vague stem such as “What is the failure rate?” Both 0.08 and 8% describe the same ratio. Asking specifically for a percentage and keeping all options in percent removes that ambiguity.
Check for unfamiliar vocabulary, unnecessary reading load and missing assumptions. For the short answer, define the evidence you expect: correct denominator, correct calculation and explanation of the count-versus-percentage error. Let a colleague review high-stakes items through your normal assessment process.
Turn answers into useful practice
Give learners the questions before the key and ask them to show one line of reasoning. Then reveal the explanation and invite them to correct their own work. A selected option alone does not show the full reasoning behind it.
If a learner selects 24%, ask them to write the ratio before converting it to percent. If they select 92%, ask which outcome the question names. These are suggested follow-up questions for this illustration, not diagnoses of a learner’s ability.
Keep an item log with the outcome, source card, draft, correction and final answer. Label any unverified item for revision. Do not claim the set is reliable, validated or effective simply because an AI tool generated it or approved it on a second pass.
Try it: a 15-minute question-writing activity
Minutes 0–3: write an outcome and source card for a familiar topic, using public or fictional material. You can reuse our percentage rule with 240 test records, of which 18 fail.
Minutes 3–7: request three questions with the template. Minutes 7–12: solve all questions without looking at the key, then check explanations and ambiguous options. Minutes 12–15: revise one weak question and record why the revision is better.
For the alternative data, the checked results are 18 ÷ 240 × 100 = 7.5% failing; 222 records pass and 222 ÷ 240 × 100 = 92.5%. Explain the two-outcome assumption if you use the complement.
Self-review: does each question test the intended outcome? Is each answer independently supported by the source card or calculation? Have you removed ambiguity and protected private or examination material? Keep the revised set and your checks; save your adapted prompt in My DYAP.
Reusable prompt
Select and copy this template, or customise your own using the free builder.
Create three low-stakes practice questions from the SOURCE CARD below. Learners: [level and prerequisites]. Outcome: [observable skill]. Use one interpretation question, one calculation or application question, and one error-analysis question. Use only the supplied rules and fictional data; if something is missing, say what is needed instead of inventing it. For each multiple-choice item, provide three plausible, non-overlapping options with exactly one best answer; vary its position. Use plain language and no trick wording. Return (1) learner questions without answers, (2) a separate key with worked reasoning, the source rule used, and why each distractor is wrong, (3) assumptions and checks for the teacher. Do not claim validation or classroom testing. I will solve and review every item before use. SOURCE CARD: [paste approved non-sensitive rules and data].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.