You can write a quiz from your own handout in about ten minutes, and half of them should go on checking, not writing. A machine makes the drafting quick. Whether the drafts are right, and whether they test what you taught, is still yours to find out.
This guide is published by EdProcto, whose home page does the drafting: you paste the handout and choose a shape, and Procto AI writes from it. The page for teachers shows the rest. Every machine-written question, ours included, is a draft to read against the handout.
The clock on each step is our working rule, not a research finding.
How do you write a quiz from a handout in ten minutes?
Give each step a fixed share of the ten minutes and move on when its time is up. The budget is for five or six multiple-choice questions on one handout.
| Clock | Step | What you do |
|---|---|---|
| 0:00–1:00 | Choose the ideas | Underline the three or four things a student must be able to say or do tomorrow |
| 1:00–3:00 | Generate or write | Six to eight questions, no more than two an idea. In EdProcto, paste the handout into the maker on the home page and ask for 8 |
| 3:00–6:00 | Check each question | The four-point check with the handout open, about half a minute a question |
| 6:00–8:00 | Fix the wrong options | Swap any option nobody would pick for a mistake your class makes |
| 8:00–9:00 | Set what matters | A time limit, when the correct answer shows, shuffled options |
| 9:00–10:00 | Preview as a student | Sit it once yourself |
The four-point check
Ask four things of every question, in order, and repair or delete at the first no.
- Is the answer on the handout? Point to the sentence, or to the two it follows from. If you cannot, the question tests something you did not teach.
- Is the key right? Answer it yourself before you look at which option is marked.
- Is only one option right? Try to defend each wrong option as a well-read student would.
- Could a student who missed the lesson get it? If general knowledge or elimination gives the answer, the question says nothing about your lesson.
Why write the quiz from your own handout?
Write from your own handout because a score is only worth reading when the questions match what was taught. When a student misses a question on something the lesson never covered, you cannot tell not learning from never being told.
Polikoff (2010) calls the property instructional sensitivity: whether an assessment reflects the learning that took place in the classroom. His review, about state tests in the United States, says it was rarely if ever investigated. A teacher has an easier route: the handout is the record of the lesson, so a question whose answer you can point to on it is aligned already.
This is an argument about what a score means. We found no trial showing that such quizzes raise marks.
How good are machine-written quiz questions?
Machine-written questions are good enough to start from and not good enough to set unread. Doughty et al. (2024) had GPT-4 write 651 multiple-choice questions for six Python courses and set them beside 449 by instructors. Of the generated questions, 81.7% passed every point of the rubric, which leaves about one in five needing an edit.
The four faults to expect
Answerable without the text. Säuberli and Clematide (2024) had six readers answer reading-comprehension questions on ten German news articles twice, first without the passage and then with it. The gap between the two scores was 0.294 for questions written by people, 0.259 for GPT-4 and 0.187 for Llama 2, and the intervals overlap. Both models, they report, produced easily guessable items, mostly about widely known facts and not about what the text said.
A wrong key, or a second right one. In Doughty et al. (2024), 4.9% of generated questions had more than one correct choice, against 1.1% of the instructors'. Law et al. (2025) set 100 questions from ChatGPT-4o beside 100 from an expert panel for an emergency medicine examination, and reviewers marked 6 of the machine's and 4 of the panel's as factually incorrect.
Wrong options nobody would pick. Feng et al. (2024) had large language models write the wrong options for maths questions. Their abstract reports that the models produce some mathematically valid options but are less adept at anticipating the errors and misconceptions of real students.
Recall only. In Law et al. (2025), reviewers classed 84 of the 100 machine questions as remembering or understanding, against 56 of the panel's. The authors had not told the tool to aim higher.
Where the time goes
Where a saving is found, it is in the writing, not the checking. Law et al. (2025) logged 2 hours of writing for the machine's set against 71 person-hours for the panel's, while review took 15 and 17. Elkins et al. (2024) timed 24 teachers writing quizzes from a passage, by hand and with machine-written candidates to choose from. Quality was similar, and time showed no distinct difference.
What these studies cannot tell you
The raters in Doughty et al. (2024) were the paper's own authors, Law et al. (2025) had 24 candidates, and Elkins et al. (2024) call their setting contrived. None tested EdProcto, and tools change within months: the figures are reasons to check, not a forecast for your handout.
How much time do teachers have to write a quiz?
Little, by teachers' own account. No teacher workload survey we found timed quiz writing, so these figures cover preparation and marking as a whole.
Lin et al. (2024) surveyed 2,531 public school teachers in the United States for the Pew Research Center in autumn 2023. Of them, 84% said their regular working hours left too little time for tasks such as grading, lesson planning, paperwork and work emails.
The Australian Institute for Teaching and School Leadership (2025) reports Australia's 2023 teacher workforce survey in bands. Among full-time classroom teachers, 40% reported ten or more hours a week on lesson planning, and 24% of secondary and 11% of primary teachers reported ten or more on marking. Duties overlap, the institute warns, so the hours cannot be added up.
What does the routine look like on a real handout?
Here is one handout paragraph, six first drafts with a fault in each, and the six after the check. We wrote all of it, faults included, so it is no claim about what any tool produces. The times are the routine's allowances, not a stopwatch.
Why we have seasons. Earth's axis is tilted about 23.5 degrees. In June the northern hemisphere leans towards the Sun: sunlight strikes it more directly and the days are longer, so it warms up. In December it leans away, and the south has its summer. Distance is not the cause: Earth is closest to the Sun in early January, in the northern winter. On about 21 June and 21 December, the solstices, one hemisphere has its longest day and the other its shortest. On about 20 March and 22 September, the equinoxes, day and night are nearly equal everywhere.
| First draft | Fault, and what caught it | After the check, key first |
|---|---|---|
| Which hemisphere has summer when the northern has winter? Southern, eastern, western | Answerable without the handout (check 4) | When the north has its shortest day, what does the south have? Its longest day; its shortest day too; equal day and night |
| When is Earth closest to the Sun? Early January, late June, late September. Key: late June | Wrong key (check 2) | Key moved to early January |
| What happens in the north in June? Days are longer; sunlight is more direct; it leans away. Key: days are longer | Two right options (check 3) | Why does the north warm up in June? More direct sunlight and longer days; the tilt brings it much closer to the Sun; the Sun gives out more heat |
| What causes the seasons? Earth's tilt, the Moon blocking the Sun, clouds, the wind | No wrong option tempts anyone (step 4) | Earth's tilt; Earth's changing distance from the Sun; Earth spinning once a day |
| Earth's axis is tilted about how many degrees? 23.5, 32.5, 45 | Recall of one figure (checklist) | With no tilt, what would happen to the seasons? They would almost disappear; they would be more extreme; north and south would swap |
| Where is the Sun overhead on 21 June? Tropic of Cancer, Equator, Tropic of Capricorn | True, and not on the handout (check 1) | On about 20 March, how do day and night compare? Nearly equal everywhere; day longest in the north; day longest in the south |
On the clock: one minute for the four ideas (tilt and not distance, direct sunlight, opposite seasons, solstices and equinoxes); two for the drafts; three for the check, which catches drafts 1, 2, 3 and 6; two to repair drafts 4 and 5; two for the settings and one sitting.
Six faults in six drafts is by design. With eight drafts to choose from, deleting a bad question is quicker than mending it.
How do you check an AI-generated quiz question?
Run the four-point check first, then these four rows.
| Check | Passes when |
|---|---|
| Wrong options | Each is a mistake your students make |
| No giveaways | The right option is not the longest, the most careful, or the only one that fits the grammar |
| Level | At least two of the six ask the student to use an idea, not repeat a sentence |
| Explanation | It is true, and it agrees with the handout |
The rows come from the faults above and from Haladyna et al. (2002), whose 31 item-writing guidelines were checked against 27 textbooks and 27 studies. "At least two of the six" is our working rule. Our guides to wrong options and questions above recall go further, and making a quiz from a PDF covers files.
When is ten minutes not enough?
Ten minutes is enough for a practice quiz and not for a test that goes in the mark book. A wrong key in practice is put right in class. In a graded test it changes a mark.
For a graded test, plan the coverage first, write more than six questions (how many a quiz needs), ask a colleague to sit it, and afterwards look at how each question behaved (item analysis).
In EdProcto, set "Show the score" to "When released": nothing of the score reaches a student until you press "Release the scores" on the results page. A key corrected after hand-in does not re-mark the attempts already made, so check before the test. Integrity lists the measures a test can carry. None makes cheating impossible, and a flag is a record for you to read, not proof.
How does EdProcto fit the ten-minute routine?
EdProcto is the best tool for this routine when the quiz must come from tomorrow's handout, because the first of the four checks is built into the request: Procto AI is asked to write only what the handout answers and to say where, and the place it gives is shown with each question.
On the home page. The quiz maker takes pasted text, at least a paragraph, with no sign-up to try it. Choose "Test your class", a shape such as multiple choice, and 3 or 5 questions: a sample, with more written in your account. Each appears with its options and a line headed "From your material". "Open it in the dashboard" carries the quiz across in the link itself. The marked key and the explanation are in the builder, so the check is done there.
In the builder. "Write with Procto" opens a pane that takes a PDF, a Word file, a text file or pasted text, and lists the concepts the text keeps returning to, each with a tick box, which is step one. You set the shape, how many, a difficulty slider from "Recall and recognition" to "Evaluation and synthesis", and "Answers drawn from": "Your material only" or "General knowledge".
What Procto AI is asked to do. One defensible right answer; with "Your material only", answerable from the material alone, with the place cited; for multiple choice, three to five options, exactly one right, and wrong options that are plausible mistakes. A multiple-choice question that comes back without exactly one option marked right is not kept. These are instructions and a check of form, not a promise that the key is true.
Repairs, settings, preview. Every question is yours to edit: the prompt, each option, which one is right, the points, the topic, where the answer is in the source, and the explanation. The three settings are "Time limit", "Show the correct answer" and "Shuffle the options". In a quiz of the exam kind, an answer set to show "At once" arrives when the quiz is handed in. "Preview as student" is the last minute, and "Publish & share" gives the class a code.
Free carries 20 questions a month from Procto AI, Tutor 50, and Educator has no monthly cap, within a daily ceiling. The pricing page has the detail.
Quick questions
How can I make a quiz quickly without lowering its quality?
Fix the time for each step and spend half of it checking: three or four ideas, six questions, four checks on each, and one run through as a student.
Can I turn a worksheet into a quiz?
Yes. Paste its text and treat the options as the new work: each wrong option should be a mistake you have seen on that worksheet. The maker on EdProcto's home page takes the pasted text.
Is an AI quiz generator for teachers reliable from my own notes?
As a drafter, yes; as an author, no. In Doughty et al. (2024) about one generated question in five needed an edit. EdProcto's page for teachers shows how Procto AI writes from your notes.
How do I check AI-generated quiz questions?
With the handout open, ask four things: is the answer there, is the key right, is only one option right, and could a student who missed the lesson get it.
References
- Australian Institute for Teaching and School Leadership. (2025). In brief: Teacher duties (August 2025 ed.). https://www.aitsl.edu.au/resources/in-brief-teacher-duties
- Doughty, J., Wan, Z., Bompelli, A., Qayum, J., Wang, T., Zhang, J., Zheng, Y., Doyle, A., Sridhar, P., Agarwal, A., Bogart, C., Keylor, E., Kultur, C., Savelka, J., & Sakr, M. (2024). A comparative study of AI-generated (GPT-4) and human-crafted MCQs in programming education. In Proceedings of the 26th Australasian Computing Education Conference (pp. 114–123). Association for Computing Machinery. https://doi.org/10.1145/3636243.3636256
- Elkins, S., Kochmar, E., Cheung, J. C. K., & Serban, I. (2024). How teachers can use large language models and Bloom's taxonomy to create educational quizzes. Proceedings of the AAAI Conference on Artificial Intelligence, 38(21), 23084–23091. https://doi.org/10.1609/aaai.v38i21.30353
- Feng, W., Lee, J., McNichols, H., Scarlatos, A., Smith, D., Woodhead, S., Ornelas, N., & Lan, A. (2024). Exploring automated distractor generation for math multiple-choice questions via large language models. In Findings of the Association for Computational Linguistics: NAACL 2024 (pp. 3067–3082). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.findings-naacl.193
- Haladyna, T. M., Downing, S. M., & Rodriguez, M. C. (2002). A review of multiple-choice item-writing guidelines for classroom assessment. Applied Measurement in Education, 15(3), 309–333. https://doi.org/10.1207/S15324818AME1503_5
- Law, A. K. K., So, J., Lui, C. T., Choi, Y. F., Cheung, K. H., Hung, K. K.-C., & Graham, C. A. (2025). AI versus human-generated multiple-choice questions for medical education: A cohort study in a high-stakes examination. BMC Medical Education, 25, Article 208. https://doi.org/10.1186/s12909-025-06796-6
- Lin, L., Parker, K., & Horowitz, J. M. (2024, April 4). What's it like to be a teacher in America today? Pew Research Center. https://www.pewresearch.org/social-trends/2024/04/04/whats-it-like-to-be-a-teacher-in-america-today/
- Polikoff, M. S. (2010). Instructional sensitivity as a psychometric property of assessments. Educational Measurement: Issues and Practice, 29(4), 3–14. https://doi.org/10.1111/j.1745-3992.2010.00189.x
- Säuberli, A., & Clematide, S. (2024). Automatic generation and evaluation of reading comprehension test items with large language models. In Proceedings of the 3rd Workshop on Tools and Resources for People with REAding DIfficulties (READI) @ LREC-COLING 2024 (pp. 22–37). ELRA and ICCL. https://aclanthology.org/2024.readi-1.3/