For US teachers, instructors, and training teams
The prompts below are the ones that actually produce usable questions. If you would rather skip the prompt writing and get a numbered quiz plus a matching answer key as two files, upload the document here instead.
Upload your file and generate the quiz
Click to upload or drag and drop
PDF, DOCX, PPTX, TXT, JPG, JPEG, PNG, HEIC, ODP, ODT, BMP, or TIFF
Up to 10 files at once, 50 MB each, merged into one quiz within the free page limit
Uploading...
The short answer
ChatGPT is a capable quiz question writer and a poor quiz builder. Attach your source, name the count, the format, the difficulty, and the output layout, and it will write solid stems. What it will not do is give you a numbered student copy and a separate answer key as files, keep coverage even across a 200 page document, or produce something other people can take and be scored on. Those are the four gaps, and they are the reason a purpose built generator finishes the job in one upload instead of six rounds of prompting.
Every one of these names the source, the count, the format, the level, and the output shape. That combination is what separates a usable question set from a page of recall trivia. Attach the file first, then paste the prompt.
More on the manual workflow in how to make a quiz with ChatGPT.
None of these are reasons to avoid ChatGPT. They are the points at which the work stops being question writing and starts being document assembly, which chat was never built for.
Attach a 180 page training manual and ask for 30 questions. What comes back will look complete and will usually lean on whatever the model retrieved first. Ask it to list the section each question came from and the gaps become obvious in about ten seconds.
You get questions and answers in one stream. Splitting them into a clean student copy and a key, then keeping the numbering aligned after you cut two questions, is manual work every single time. See the answer key generator for the version that stays in sync.
Copying chat output into Word gives you stray markdown, inconsistent numbering, and no answer space. For a paper exam that has to look like an exam, the formatting pass takes longer than the generation did. A printable test maker skips it.
A chat transcript is not a quiz your class can sit and be scored on. Getting there means retyping everything into Forms, Canvas, or a quiz platform, which is where most people abandon the workflow after the first attempt.
An honest split. ChatGPT wins on flexibility and on anything conversational. A generator wins the moment the output has to be a document somebody else uses.
For a wider look at using a chat assistant to revise, read ChatGPT for studying.
Step 1
The chapter, manual, handout, or slide deck the quiz is meant to cover. Scanned pages are fine.
Step 2
Multiple choice, true or false, short answer, or a mix. This is the part you would otherwise spell out in a prompt.
Step 3
Read each question against the source. Distractors are where AI written items need a human most, whichever tool wrote them.
Step 4
Send a link that scores automatically, or download the student copy and the answer key as separate files.
One quiz a week times thirty weeks. The prompt is quick. The reformatting into a printable test with a key, thirty times over, is not.
A knowledge check has to be scored and recorded to count as evidence. A chat transcript is not documentation, which matters for compliance training.
Course length source material means coverage matters more than wording, and coverage is exactly what a chat window handles worst.
Conversational quizzing works here, and it is genuinely useful. For repeat practice from a stable set, see notes to quiz.
Yes. ChatGPT writes quiz questions well when you give it the source text and tell it the format, the count, and the difficulty. What it does not do is produce a numbered student copy and a separate answer key as downloadable files. You get questions in the chat window, and the formatting work is still yours.
It can, with two caveats. On a long PDF the model retrieves passages rather than reading the file end to end, so coverage skews toward whatever your prompt emphasized. And on a scanned PDF with no text layer, quality depends entirely on how legible the page images are. Ask it which sections it drew from before you trust the coverage.
The prompt that works names five things: the source, the question count, the format, the difficulty, and the output layout. For example, write 20 multiple choice questions from the attached chapter, four options each, one correct, at the level of an undergraduate midterm, then list the answer key separately at the end. Vague prompts produce recall trivia.
It can quiz you conversationally, one question at a time, if you ask it to. It cannot produce a shareable interactive quiz that other people take and that scores them for you. For that you need either a quiz platform or a generator that exports a link or a graded file.
The free tier will make quizzes, but file uploads are capped per day and long documents are the first thing to hit that ceiling. If you are generating quizzes weekly from full chapters or training manuals, the practical question is not price, it is how much of your time each quiz costs after the model hands back raw text.
It can format questions in a way that is easier to paste into a Form, but Google Forms has no bulk import for a block of text, so you still add each question by hand or run an add on script. The volume of manual work is the reason most people stop after one quiz. The detail is in PDF to Google Form.
It can, and the failure is subtle. Questions usually look right and the wrong ones are wrong in plausible ways, most often in the distractors and in numbers or dates the source only implies. Read every item against the passage it came from. Five minutes of review catches nearly all of it.
Same source material, no prompt engineering, and a numbered quiz with a matching answer key at the end of it.