ChatGPT Prompts for Quiz Writing (With a Free Review Checklist)
Copyable ChatGPT prompts for quiz writing, a real output and what went wrong with it, and a free Excel prompt builder and review sheet.
A ChatGPT prompt for quiz writing works best with four things: who the learners are, the learning objective, the source material, and the question type.
- Give the chatbot the approved source and tell it to use that material. Check the output against it: this instruction does not prevent every invented fact.
- In a quiz prompt, ask for realistic wrong options, exactly one correct answer, no negative stems, and one sentence of feedback for each option.
- A chatbot can put correct answers in the same few positions even when asked not to, so count them and shuffle the options yourself.
- Review every question for source accuracy, objective alignment, one defensible answer and plausible alternatives. Spreadsheet pattern flags cover only part of that review.
Generate engaging quizzes in seconds
ChatGPT can draft quiz questions, an answer key and feedback from your source material. Give it the learning objective too, so the questions have a clear purpose. Then check the result: a plausible question can still have a wrong answer or test something you did not intend.
Below are copyable prompts, an actual Claude output from the main prompt and a review of what went wrong. The free Excel builder helps you prepare a prompt and record your review; its formulas flag selected patterns in multiple-choice questions.
I work in marketing at Coursebox, and I use AI chatbots heavily in my own work. So I tested these prompts the way I'd test any prompt: by running them and checking what came back. Coursebox also makes an AI quiz generator that writes questions from your course or document and puts them straight into the course.
Among those requests, ten asked for plausible wrong answers and one ruled out “all of the above”. The analysis covers written requests in the 180 days to 20 October 2026, rather than the full assessment plans people may have held elsewhere.
The worked example is a fictional complaints procedure for new contact centre agents. I ran its five-question prompt in Claude, then reviewed the output. One question did not test the intended application of the procedure, and the correct answers were concentrated on one letter.
Download the blank builder to write your own prompt and review what the chatbot sends back. The example file has the complaints quiz: the prompt, the real questions it produced, and the review that flagged them. It also has 10 more prompts and a 10-point review checklist.
Open the example workbook preview at full size
Can ChatGPT write a quiz?
Yes. Given the source and the objective, ChatGPT, Claude, Gemini and Copilot will draft multiple-choice, true or false, scenario and short-answer questions, with an answer key and feedback.
A 2024 review of 23 studies on ChatGPT-written multiple-choice questions found they save time but vary in quality, and recommends checking them before use (Kıyak and Emekli, 2024). The studies are from medical education, but the advice holds for any quiz.
A chatbot can quiz you one question at a time for practice, or draft questions you can review and put into a course. This guide covers the second task: preparing questions, checking the answer key and deciding whether they assess the intended learning.
The prompt: questions from your own material
This is the prompt the builder assembles. Replace the words in capitals, then paste your procedure, policy or lesson at the end.
Write a quiz on a procedure, policy or lesson, using nothing but that material.
- The source material, pasted in full.
- The learning objective.
You are an experienced workplace trainer writing quiz questions for LEARNERS. Write NUMBER multiple-choice questions that check this learning objective: OBJECTIVE Use only the source material below. Do not add facts that are not in it. If the source doesn't cover something, leave it out. For each question: - Write a short scenario or a direct question (the stem). Ask about applying the source to a real situation, not recalling a definition. - Give 4 options (A to D) with exactly one correct answer. - Make the 3 wrong options mistakes a learner might really make. Keep all options about the same length and in the same grammatical form. - Don't use "all of the above", "none of the above", or negative stems such as "Which is NOT". - Spread the correct answers across A, B, C and D. - After the options, give the correct letter and one sentence of feedback for each option, explaining why it is right or wrong, with the part of the source it comes from. Use Australian English. Source material: """ SOURCE """
- LEARNERS
- : Who the quiz is for, in the words you'd use at work (for example, new contact centre agents)
- NUMBER
- : How many questions (20 or fewer per run) (for example, 5)
- OBJECTIVE
- : The learning objective the questions check (for example, given a customer complaint, new agents will apply the complaints procedure)
- SOURCE
- : The text the questions must come from (for example, your procedure, policy or lesson notes)
Example output (written by hand, not a live run)
Run in Claude (claude-sonnet-5) on 21 October 2026, with the complaints procedure as the source. The first two of five questions, without the feedback lines: Q1. A customer rings about a faulty order and asks for a $150 credit. What should you do? A. Tell the customer you'll check with your team leader before promising anything. B. Approve the $150 credit yourself, as it is within your own approval limit. C. Offer a $150 refund instead, since refunds need separate approval. D. Log the complaint but wait 5 business days before approving the credit. Correct answer: B Q2. A customer says their complaint is also being sent to the industry ombudsman. What's your first step? A. Continue handling it yourself since it's still within your approval limit. B. Wait and see if the customer follows through before taking any action. C. Escalate to a team leader straight away because of the ombudsman mention. D. Log it in the CRM and revisit it if it takes longer than 5 business days. Correct answer: C
- Every marked answer is correct according to the source.
- Every fact in the questions and options is in the source.
- Each question tests the objective, not a definition.
- The correct answers are spread across A, B, C and D.
Change "multiple-choice questions" to "scenario-based multiple-choice questions" or "true or false statements", or use the builder, which rewrites the instructions for each type and level.
Each line in the prompt is there for a reason:
| Line | What to check in the output |
|---|---|
| Use only the source material | Plausible facts the chatbot fills in when your source is thin |
| Ask about applying, not recalling | A quiz of definitions when the job needs decisions |
| Exactly one correct answer | Two defensible answers, which punish the learners who think hardest |
| Wrong options are real mistakes | Silly options that give the answer away |
| Same length and form | The longest, most detailed option being the right one |
| No all or none of the above, no negative stems | Questions answered by test-taking tricks, or misread under pressure |
| Spread the answers | Every answer landing on B |
| Feedback for each option | Learners who get it wrong and never find out why |
How to write a quiz with ChatGPT
The steps follow the complaints quiz: five questions for new contact centre agents, from a six-step complaints procedure.
Step 1: Start from the objective and the source
Write the objective first: "given a customer complaint, new agents will apply the complaints procedure, including what they can approve themselves and when to escalate." It decides what every question is for. There's more on writing one in these SMART learning objectives examples.
Find the approved procedure, policy or lesson the questions must come from. If important information is missing, get it from the document owner before asking learners about it. Telling the model to use only the source helps state the task; it does not guarantee it will follow that instruction.
Step 2: Fill in the builder
Enter the learners, objective, number of questions, type, level, feedback preference and source. The builder checks supported inputs and assembles the prompt. Its Question review sheet supports four-option multiple-choice questions; review true-or-false and short-answer outputs manually.
The inputs for the complaints quiz
Example: Five questions for new contact centre agents, from an example complaints procedure.
- 1. Source: The fictional six-step procedure is in the downloadable example.
- 2. Review: The objective calls for applying the procedure, rather than only recalling fields.
Open this workbook preview at full size
Step 3: Run it and paste the questions back
Paste the prompt into ChatGPT, Claude, Gemini or Copilot. I ran it in Claude (claude-sonnet-5) on 21 October 2026, then copied each question into the review sheet: the stem, the four options and the correct letter.
Step 4: Make the two judgements yourself
Start with two judgements recorded on the sheet: does the marked answer match the source, and does the question test the objective? All five marked answers matched this example procedure. You must also check the other checklist items yourself, including whether another option could be defended and whether the wrong answers are plausible.
But question 3 asked "which of these is a mandatory field you must include?" That's recall, even though the prompt asked about applying the procedure. I marked it as not testing the objective.
Step 5: Read the review
Formula prompts look for missing or duplicate options, selected “all/none” phrases, selected negative wording, a disproportionately long correct option and concentrated answer positions. They do not judge meaning. Review the full checklist even when a row says OK.
A question that needs rewriting
Example: Claude (claude-sonnet-5), 21 October 2026, from the builder's prompt.
- 1. Question 3: The human reviewer marked the recall question as misaligned.
- 2. OK: No tested pattern flag; finish the full human checklist.
- 3. Answer positions: The complete example contains A 0, B 3, C 1 and D 1.
Open this workbook preview at full size
The bunching wasn't a one-off. An earlier run with almost the same prompt put the correct answers on B, C, B, C, B. In that run, one correct answer was nearly twice as long as the wrong options, and the review flagged it.
Step 6: Fix what's flagged
I asked the chatbot to turn question 3 into an application question. It added a sentence about a customer calling about a billing error, then asked the same recall question. So I rewrote it myself:
"A customer says: 'You've charged me twice for October. I want the extra $45 back.' What do you enter in the CRM description field?" The answer is the customer's own words; the three wrong options are summaries an agent might write instead.
Swapping the options in question 1 makes its answer A, giving A 1, B 2, C 1 and D 1. Update the key and feedback whenever you move an option. After the question edits and recorded checks, the status reads “No automated flags; complete checklist”; that is a prompt to finish the review, not approval.
The two recorded runs concentrated correct answers on B despite the instruction. You can ask for a revision, but check the positions yourself and update the answer key if you reorder options. A perfectly even count is not required for every short quiz.
Step 7: Add the questions to your course
Paste the final questions into your quiz tool or course. If you're building a short check inside a module, this guide to knowledge checks covers where to place them and how many to use.
RP Global Training Network: review before publication
RP Global Training Network describes using Coursebox for a ten-question assessment at the end of each lesson, with the trainer reviewing content against the source before publication. The useful practice here is the source review; the case does not validate this prompt or a universal ten-question length. Read the case study.
Coursebox's AI quiz generator writes questions from your course content or a topic, in the number and types you choose, and you can edit every question. You can shuffle the answer order for each question. It doesn't write an explanation for each multiple-choice option, so add feedback yourself where learners need it.
More ChatGPT prompts for quizzes
The workbook library contains ten additional prompts. These three cover scenario questions, improving wrong answers and reviewing a draft.
Test whether people can use a rule, not recite it.
Write NUMBER scenario-based multiple-choice questions for LEARNERS. Each starts with a workplace situation of 2 or 3 sentences and asks what the person should do. Use only the source below; do not add facts. Give 4 options with one correct answer, make the wrong options mistakes a new person might really make, keep the options about the same length, and give the correct letter with one sentence of feedback per option. Source: """SOURCE"""
- NUMBER
- : How many questions (for example, 5)
- LEARNERS
- : Who the quiz is for (for example, new warehouse staff)
- SOURCE
- : The text the questions must come from (for example, your procedure)
- Each scenario needs the source to answer, not common sense alone.
Fix wrong options that are obviously wrong.
Here is a multiple-choice question for LEARNERS: QUESTION. Rewrite only the three wrong options so each is a mistake a learner might really make, based on the source below. Keep them about the same length and grammatical form as the correct answer. Don't change the stem or the correct answer. Source: """SOURCE"""
- LEARNERS
- : Who the quiz is for (for example, new contact centre agents)
- QUESTION
- : One question with its options and answer (for example, your question)
- SOURCE
- : The text the question comes from (for example, your procedure)
- The correct answer and the stem are unchanged.
Get a second opinion before your own review.
Review these questions against the source. For each one, tell me: is the marked answer correct according to the source; could any other option also be defended; does anything in the stem or options give the answer away; does it test this objective: OBJECTIVE. Don't rewrite anything yet. Questions: QUESTIONS. Source: """SOURCE"""
- OBJECTIVE
- : The learning objective (for example, given a customer complaint, apply the complaints procedure)
- QUESTIONS
- : The questions with their options and answers (for example, your questions)
- SOURCE
- : The text the questions come from (for example, your procedure)
- Treat its answers as a second opinion. You still check the answer key against the source.
The other seven cover feedback for each option, a second version of each question for a question bank, true or false refreshers, questions against objectives, moving recall questions up a level, and a short-answer marking guide.
The review checklist
| Check | Why | Who reviews |
|---|---|---|
| The answer key is right | Chatbots sometimes mark the wrong letter | You |
| Every fact is in the source | When the source is thin, chatbots fill the gap | You |
| It tests the objective | Recall questions are easier to write, so they crowd in | You |
| Only one answer can be defended | A second defensible option punishes careful learners | You |
| Wrong options are real mistakes | Silly options make the answer obvious | You |
| No all or none of the above | Learners can answer without knowing the content | Template prompt and you |
| No negative stems | "Which is NOT" gets misread under time pressure | Template prompt and you |
| The answer isn't the longest option | Test-wise learners pick the most detailed option | Template prompt and you |
| Answers spread across A to D | Chatbots can repeat the same positions | Template prompt and you |
| No two options are the same | Duplicates leave only three real choices | Template prompt and you |
The University of Waterloo's teaching centre gives the same advice for any multiple-choice question: avoid "all of the above" and "none of the above", and make the wrong options plausible and similar in length to the answer (University of Waterloo).
The levels behind "tests the objective" are in this guide to Bloom's taxonomy in course design.
Common mistakes with AI quiz prompts
| Mistake | What happens | What to do |
|---|---|---|
| No source material | Questions on facts your organisation doesn't use | Paste the procedure or policy, and say "use only this" |
| No objective | A quiz of whatever was easiest to ask | Name the one thing the quiz checks |
| Asking for more than you can review | Questions go out unchecked | Ask for 20 or fewer per run, and review each batch |
| Trusting the answer key | Wrong answers marked right | Check every answer against the source |
| Asking the chatbot to fix its own patterns | The same bunching or recall question comes back | Shuffle and rewrite yourself |
Where to go next
- To compare tools, see this round-up of the best AI quiz generators.
- To decide what kind of quiz, read about formative and summative assessment.
- For prompts beyond quizzes, see the best ChatGPT prompts for course creators.
- To mark open answers, build a rubric.
The builder is version 1.1, reviewed on 22 October 2026 and tested in Excel for Windows. The recorded examples were run in Claude on 21 October 2026. Other models and spreadsheet applications were not tested here; review their outputs independently.
Writing quizzes with ChatGPT
Give the chatbot your learners, the objective, the source and the question type, and tell it to use nothing else. Then check every answer against the source, make sure each question tests the objective, and shuffle the options yourself before learners see them.
To have the questions written from your course and added to it, try Coursebox's AI quiz generator.
Frequently Asked Questions
Yes. Give it the source material, the learning objective, the number and type of questions, and ask for an answer key. Check every question before learners see it.
One that names the learners and the objective, pastes the source and says "use only this", asks for one correct answer and realistic wrong options, rules out "all of the above" and negative stems, and asks for feedback on each option.
They can repeat the same few positions even when asked to spread them. In both of our runs, B came up most. Shuffle the options yourself and update the key.
The workbook supports batches of 1 to 20 questions. Choose a smaller batch if that is what you can review carefully; 20 is a file limit, not a proven ideal batch size.
You can ask a chatbot to quiz you interactively for practice. Drafting a quiz for other learners also requires reviewed questions, an accurate answer key, suitable feedback and checks in the delivery tool.
Parshad Desai
Parshad is a full stack technical marketer based in our Perth office. He began his Coursebox career as a marketing intern. After impressing the founders with his incredible work, he later transitioned into a full time position coordinating digital marketing. His focus lies in managing our website, product analytics, and email marketing. Recently he has taken a keen interest in applying AI to digital marketing and growth.
