TL;DR
Paste your notes into a general chat model (ChatGPT, Claude, or Gemini) with a prompt that forces a fixed cognitive mix and a hidden answer key, or use a dedicated tool like Quizlet or Quizgecko that ingests PDFs and YouTube directly. Then re-test only the items you miss on a spacing schedule — roughly day 1, day 7, then day 16 to 35 — because the testing itself, not the reading, is what builds retention.
What you are actually trying to make
A self-test that measures understanding, not just whether you can repeat a definition. That means three things most AI-generated quizzes get wrong:
- A deliberate mix of cognitive levels, not 20 recall questions in a row.
- Multiple-choice distractors that are genuinely plausible, so you cannot pick the answer by elimination without knowing the material.
- A hidden answer key plus a retest schedule, so the quiz is a study loop rather than a one-time worksheet.
The rest of this guide is about hitting that bar reliably.
Two routes: chat model vs. dedicated tool
There is no single best tool. The right choice depends on your source material and how much control you want over question design.
| Route | Best for | Cost (as of June 2026) | Trade-off |
|---|---|---|---|
| ChatGPT / Claude / Gemini | Full control over cognitive mix, formats, and tone; pasting raw notes | Free tiers work; Plus/Pro $20/mo for longer context | You build the structure via prompt |
| Quizlet | Notes or PDFs into MCQs + flashcards with AI “Smart Grading” auto-credit | Free tier; Plus $7.99/mo or $35.99/yr (~$3/mo) | Question styles are more fixed |
| Quizgecko | Turning a PDF, webpage, or YouTube video into a quiz; short-answer grading | Free plan (3 quizzes/mo); Student $16/mo or $64/yr | Less control over Bloom mix |
| StudyGlen / Quizbot | PDF, text, and image input (OCR) with explanations, 30+ languages | Free tier available | Output quality varies by topic |
A practical rule: if your material is raw notes you typed or photographed, paste it into Claude or ChatGPT and control everything with a prompt. If your material is a PDF, slide deck, webpage, or video, a dedicated tool that ingests those formats directly saves time. Pricing on the dedicated tools shifts often, so confirm the current plan on the vendor site before you pay.
For the chat-model route, any current frontier model handles this well. As of June 2026 that means GPT-5.5 in ChatGPT, Claude Sonnet 4.6 or Opus 4.7, or Gemini 3.1 Pro. Claude (Opus 4.7 and Sonnet 4.6) and Gemini 3.1 Pro both carry a 1M-token context, so a full chapter or a stack of lecture notes fits in one prompt. ChatGPT’s in-app context is tighter on the $20 Plus tier (roughly 320 pages of text; the full 1M window is reserved for the $200 Pro tier), which is still plenty for one chapter of notes. Claude’s free tier (limited Sonnet 4.6) and ChatGPT’s free tier (GPT-5.5, tighter limits) are both enough for a single quiz session.
When AI is the right tool — and when it is not
Use AI when you have notes or readings but no instructor test bank, when you are prepping for an exam and need 50 to 200 practice items, or when you want parallel quizzes in two languages for bilingual study.
Do not lean on AI alone for:
- High-stakes credentialing that demands verified, source-cited items — medicine, law, security certifications. One wrong answer in the key misleads you for weeks.
- Math beyond intro level. Models still slip on multi-step proofs and exotic notation; check every worked answer.
- Topics that changed after the model’s training cutoff. A model that has not seen the latest standard will confidently quiz you on the old one.
In all three cases AI can still draft the questions — you just have to verify the answer key against a trusted source rather than trusting it.
What to feed the model
The quality of the quiz tracks the quality of the input. Give it:
- The notes or chapter text in full, pasted as the source, with an explicit instruction that questions must be answerable from that text alone.
- The level: high school, undergrad, professional exam, or hobbyist.
- The count and cognitive mix you want — a 50/30/20 split across recall, application, and analysis is a solid default.
- The formats: multiple choice, short answer, code-completion, essay.
- An anti-pattern list: no trick questions, no items where the answer sits in the stem, distractors must be parallel in length and grammar.
That last point matters more than it sounds. Research on multiple-choice item writing finds that a well-formed stem should pose a question answerable on its own, and that the hardest part of a good MCQ is writing distractors plausible enough to function — weak distractors are why so many machine quizzes feel too easy (UBC item-writing guidelines).
Copy-ready prompt
Replace the bracketed placeholders. This forces the cognitive mix, the hidden key, and the distractor rules in one shot.
You are a learning scientist building a self-test from notes.
Source notes:
"""
[paste notes here]
"""
Level: [e.g. undergrad biology]
Number of questions: [e.g. 20]
Format mix: [e.g. 12 multiple choice, 6 short answer, 2 essay]
Cognitive mix:
- 50% recall (Bloom: remember, understand)
- 30% application (apply concepts to a new scenario)
- 20% analysis (compare, evaluate, justify)
Rules:
- Every question must be answerable from the notes alone.
- MCQ distractors must be plausible and parallel in length and grammar.
- No item where the answer appears in the stem; no trick questions.
- Tag each question with its cognitive level.
- After the questions, insert a horizontal rule, then the answer key.
- In the key, format each line as "Q1. Answer - one-sentence explanation".
- Mark any item where the notes are ambiguous with [verify].The output should be numbered questions with type tags, a horizontal rule, then a key formatted Q1. Answer - one-sentence explanation. That layout lets you cover the key, take the quiz cold, and self-grade fast.
How to check the output before you trust it
- Spot-check 5 random items against the notes. If the model invented content not in your source, regenerate with a stricter “notes alone” constraint.
- Verify the precise facts in the answer key — numbers, dates, names. Models paraphrase, and a paraphrased number is sometimes a wrong number.
- Test the distractors. If you can pick the right answer on an MCQ without reading the stem, the distractors are too weak. Ask the model to make them “more plausible and closer to common misconceptions” and regenerate just those items.
- Take it cold. Cover the key, answer everything, then score honestly. The gap between “I felt like I knew it” and your actual score is the whole point.
The retest schedule is where learning happens
A quiz you take once and discard does almost nothing for long-term memory. The evidence on spaced retrieval practice is unusually strong: distributed testing beats massed re-reading for durable retention. There is no single magic interval — research is clear that any spacing beats none, whether you space over days, weeks, or months (retrievalpractice.org on optimal spacing). A practical starting schedule from the spaced-repetition literature is to re-test at roughly day 1, day 7, day 16, and day 35. The evidence on whether expanding intervals (each gap longer than the last) beat fixed intervals is genuinely mixed, so do not overthink it — the reliable wins come from getting a first retest in within a day or two and not skipping the later repeats.
The practical loop:
- Take the full quiz cold and score it.
- Log every missed item in a
weak-topics.mdfile. - Re-test only the missed items the next day, then about a week later, then a few weeks out.
- Feed
weak-topics.mdback into your next prompt: “weight application questions toward these three topics.” Over a few weeks you build a personalized question bank aimed at your actual gaps.
Common mistakes
- All recall, no application. Passing feels like a win, but your understanding stays shallow. Enforce the 50/30/20 mix.
- Throwaway distractors. If the wrong answers are obviously wrong, the MCQ tests nothing. Demand parallel, plausible distractors.
- Letting the model add content beyond the notes. It sounds smart and quietly ruins the test — now you are quizzing yourself on something your source never said.
- No retest schedule. One-and-done quizzes do not build retention. The spacing is the active ingredient.
FAQ
- How many questions per session? 10 to 20 for focused review of one topic; 50 or more for cumulative exam prep across several chapters.
- How often should I retest missed items? Roughly day 1, day 7, then day 16 to 35. Any spacing beats cramming; the exact numbers matter less than not skipping the repeats.
- MCQ or short answer? Short answer forces harder retrieval and gives less to guess from; MCQs are faster and easier to auto-grade. Mix both, and lean toward short answer for material you keep missing.
- Can AI grade my essays? Yes for rough feedback on structure and coverage. Always cross-check the substance against your source notes — the model will sometimes reward fluent writing that is factually off.
- Which model should I use? Any current frontier model is fine for one quiz. For long source material, Claude (Opus 4.7 / Sonnet 4.6) and Gemini 3.1 Pro both expose a 1M-token context as of June 2026, so a whole chapter fits in one prompt. ChatGPT Plus holds roughly 320 pages in-app, still enough for a single chapter.
Related
For deeper study workflows, see quiz generation prompts, flashcard prompts, and study plan prompts. To clean up your source material before quizzing, see lecture note cleanup and study notes cleanup. To turn missed items into a focused plan, see exam mistakes review.