The Best ChatGPT Prompts for Writing Quiz Questions From a Document

The best ChatGPT prompt for quiz questions names five things: the source, the number of questions, the exact format, the cognitive level, and the output layout. Prompts that skip any of those five come back as definition recall trivia. The prompts below are grouped by the job they do, and each one explains which specific phrase is carrying the weight, so you can adapt them instead of copying them blind.

All of these assume you have attached the source document first. Prompting from memory produces questions about the topic in general, not about the chapter your class actually read, and the difference shows up the moment a student challenges an item.

The five parts of a prompt that works

Prompt 1: the standard chapter test

Write 20 multiple choice questions from the attached chapter, four options each with exactly one correct answer. Distribute them evenly across every major section. After the questions, list the answer key separately with the section each item came from.

The phrase doing the work is "distribute them evenly across every major section." Without it, the model over samples whatever it retrieved first, which on a long document usually means the opening pages. The request for the source section on the key is not decoration either, it is how you audit coverage in ten seconds instead of rereading the whole quiz.

Prompt 2: questions above recall

Write 10 questions at the apply and analyze levels of Bloom's taxonomy. No question may be answerable by matching a phrase from the text. Each stem must present a short scenario a practitioner would actually face.

"No question may be answerable by matching a phrase from the text" is the single most useful constraint in this list. It blocks the model's default behavior, which is to lift a definition sentence and turn it into a stem. The scenario requirement forces transfer, which is what you were trying to test in the first place.

Prompt 3: fixing the distractors

Rewrite the wrong answers so each one reflects a specific misunderstanding a learner might actually hold. Remove any option that is obviously absurd. Keep all four options within a similar length and level of detail.

Run this as a second pass on questions you already have. Weak distractors are the most common defect in AI written multiple choice, and the length rule matters more than people expect: models tend to write the correct answer longer because it carries the qualifiers, which hands test wise students a free pattern to exploit.

Prompt 4: the coverage audit

List every major section of the attached document, then state how many of the questions you just wrote came from each section.

Not a generation prompt, a verification one, and the one most people never run. It exposes the gap between what the quiz appears to cover and what it actually covers. If six sections have zero questions, you know exactly what to ask for next instead of guessing.

Prompt 5: a second version of the same test

Write a parallel form of the quiz above: same section coverage, same difficulty, same question count, different scenarios and different numbers. Provide a separate answer key labeled Form B.

Ask for a parallel form, not a shuffle. A shuffled test is the same test, which does nothing about a student who saw the original. The trade offs between the two approaches are in how to make two versions of the same test.

Prompt 6: matching a real exam format

Write questions in the format of the attached exam: same number of options, same stem length, same balance of question types. Use the attached study material as the source of content.

Two attachments, two different roles. This is how you get practice questions that feel like the real thing rather than generic items, and it works well for certification and licensing prep where the format itself is part of the difficulty.

Prompt 7: short answer with a scoring note

Write 8 short answer questions. For each, provide a model answer and a one line scoring note describing what a response must contain to earn full credit and what earns partial credit.

The scoring note is the difference between a key you can hand to a TA and a key only you can grade from. Anyone else marking your test needs to know whether a partially correct answer counts, and without that line the scores will not be consistent between graders.

Prompt 8: true or false that is not trivial

Write 12 true or false statements from the attached document. Half must be false. Every false statement must be false because of one precise detail, not because it is nonsense. No absolute words like always or never.

Absolutes are the giveaway that makes true or false items free points, because a statement containing "always" is almost always false. Banning them in the prompt is faster than editing them out afterward.

Prompt 9: pulling out what is actually testable

Before writing questions, list the ten facts, procedures, or distinctions in this document that a reader must understand to use it correctly. Then write two questions on each.

Splitting the work into an outline step and a writing step produces noticeably better coverage than asking for questions directly, because the model commits to what matters before it starts generating. It also gives you something to argue with. If item four on that list is not actually important, say so and regenerate.

Prompt 10: the review pass

Review each question you wrote against the source. Flag any item where the source is ambiguous, where two options could be defended as correct, or where the answer depends on information not in the document.

Models are better at critiquing their own output than at getting it right the first time. This catches a real share of the items that would otherwise reach a student, though it does not replace your own read through.

What the prompts cannot fix

No prompt turns chat output into a document. You still end up splitting questions from answers by hand, cleaning formatting, adding answer space, and retyping everything into whatever platform the class or team actually uses. That is the ceiling on the prompting approach, and it does not move with better wording. The comparison against uploading the file directly, along with the shortened prompt set, sits on the ChatGPT quiz generator page, and the Google equivalent is on the Gemini quiz generator page.

One more thing that helps with prompts 2 and 6. Good scenario stems need plausible situations, and that is a separate creative problem from writing the question itself. Spending five minutes to generate a list of realistic situations your learners face, then feeding those into the question prompt, produces noticeably better applied items than asking the model to invent both at once.

Frequently asked questions

What is the best prompt to make a quiz in ChatGPT?

The one that names the source, the count, the format, the cognitive level, and the output layout in a single instruction. For example: write 20 multiple choice questions from the attached chapter, four options each with one correct answer, evenly spread across all sections, then list the key separately with the section for each item.

Why are ChatGPT's quiz questions so easy?

Because recall is the default. Unless you ban phrase matching and specify an applied level, the model lifts definitions from the text and converts them into stems, which produces items a student can answer by pattern matching without understanding anything.

How do I stop the wrong answers from being obvious?

Run a dedicated distractor pass. Ask for every wrong option to reflect a specific misunderstanding, ban absurd options, and require similar length across all options. Doing this as a second prompt works better than folding it into the first one.

Can ChatGPT tell me if its own questions are wrong?

Partly. Asking it to flag ambiguous items and items with two defensible answers catches a meaningful share of problems. It will not catch everything, so read the flagged items and the numeric ones yourself against the source.

Do these prompts work in Gemini and Claude too?

Yes. The five part structure is not model specific, and the constraints that improve quality, banning phrase matching, requiring scenario stems, controlling distractor length, work the same way across assistants. Only the file limits differ.