The weak version of an AI quiz generator for teachers is typing "make a quiz about photosynthesis" and getting ten generic questions that may or may not touch what you actually covered. The strong version starts from your material: paste your lesson notes, your slide text, or the textbook section, and generate questions that can only be answered by a student who was in your room. The difference is alignment — and alignment is the entire point of a quiz.
A topic name pulls from the model's general knowledge, which includes things you never taught and skips things you emphasized. Your source material constrains it. The prompt that works: "Using only the content below, write 10 questions — 5 multiple choice, 3 short answer, 2 applying the concept to a new scenario. Multiple-choice distractors must reflect real misconceptions, not obviously wrong answers." That last clause matters most; distractor quality is what separates a measurement instrument from a word search.
Exit tickets want three questions at one minute each — recall, application, and one "explain why" to surface misconceptions while you can still fix them tomorrow. Weekly quizzes want the 5-3-2 mix above. Test review wants volume: 30 questions you'll cut down, plus a second parallel form with the same skills and different numbers — which AI produces in one extra sentence, and which quietly solves the wandering-eyes problem on quiz day.
Validity check in 90 seconds: before printing, answer the quiz yourself cold. Any question you hesitate on is ambiguous. Any question answerable from common sense without the lesson gets cut. AI writes fluent questions; fluency is not validity.
Generating a blank quiz is safe in any tool. Grading is where student data enters, and where local matters: if an agent is reading student responses, extracting scores, or building a who-missed-what breakdown, that work should happen on a machine you own, not a cloud service's logs. The item-analysis payoff is real — "question 7 was missed by 60% of the class" tells you Monday's warm-up — it just shouldn't cost you custody of your students' work to get it.
The same paste-your-material pattern powers review games, station cards, and study guides — one source, many formats. Slides for review day come from the presentation maker, and the broader automation picture for your workload lives in the AI agent for teachers and AI agent for tutors guides.
The real payoff arrives around week six. Every quiz you generate goes into a folder — by unit, by standard, by question type. When review week comes, you're not generating from scratch; you're asking for a mixed set drawn from questions your students have already seen cousins of. When a student needs a retake, the parallel form takes one sentence. When next year arrives, the bank is your starting inventory and this year's item-analysis notes ("question 7 style consistently catches the unit's core misconception — keep") make it better than anything you could buy. Teachers who do this manually call it a career's work. With generation this cheap, it's a semester's habit. The bank is also where an agent with memory beats a chatbot with amnesia: a system that remembers your standards, your formats, and your past questions stops being a tool and starts being a colleague.
Can AI write good multiple-choice distractors? Yes, if you ask for misconception-based distractors explicitly and give it your source material. Generic prompts produce giveaway wrong answers.
Should I let AI grade open responses? As a first pass with your rubric, it's a useful draft — but keep final grading judgment yours, and keep student responses on local hardware.
How do I stop students from AI-solving the quiz? Application questions tied to in-class examples ("using the data from Tuesday's lab...") are much harder to outsource than definition recall.
Assess what you actually taught: build review-day slides with the presentation maker, and join the QADIR OS early-access list for a local agent that turns your notes into aligned quizzes — and keeps student work on your hardware.