Quiz generator from PDF · what happens on take #2
A PDF quiz is a memory test for the wording, until take #2 looks different.
Every other guide for this topic stops at the upload step. The interesting question is what arrives the second, fifth, and fifteenth time you sit down with the same PDF. If the stem opens with the same five words, you stop learning the material on take two and start memorizing the phrasing.
This page is the walkthrough of one quiz from one anatomy PDF, traced across four sessions over three weeks. The fact under the question stays the same. Nothing else does.
The cue you don't notice you're using
You upload an anatomy lecture PDF on Sunday night. The quiz comes back in about a minute. One of the questions you miss is about the interventricular septum. You read the explain panel, see the quote from page 47 of your PDF, and move on.
On Tuesday you come back to the same deck. Most quiz tools serve you the identical question, sometimes with the answer letter shuffled. That is the moment your brain stops learning the anatomy and starts learning the phrasing. The first three words of the stem become a retrieval cue for the answer letter. You feel like you're getting it; you're actually pattern-matching on a string.
The walkthrough below is what take #2, take #5, and take #15 of the same quiz look like inside Studyly. Each take pulls from the same PDF page. Nothing else holds steady.
One question, four sessions, three weeks
Take 1 — Sunday night
First quiz from the lecture PDF. Question opens with "Which structure separates the right and left ventricles?" You miss it because septomarginal trabecula sounded right. Explain pulls page 47 of the source.
Q1 (take #1) Which structure separates the right and left ventricles? A) Tricuspid valve B) Septomarginal trabecula C) Interventricular septum ← right answer D) Moderator band
Take 2 — Tuesday night
Same fact, different stem. The question now reads "The wall dividing the right and left ventricular cavities is the:" The right answer has moved from C to B. The distractor set has rotated; septomarginal trabecula is gone.
Q1 (take #2, two days later) The wall dividing the right and left ventricular cavities is the: A) Membranous interatrial wall B) Interventricular septum ← right answer C) Crista terminalis D) Tricuspid annulus
Take 5 — the following Monday
Now framed as a clinical scenario. A patient with a defect; what failed to form. The fact under the question is unchanged. The verb in the stem has flipped from passive description to active diagnosis. Pattern-matching on the first three words gets you nothing.
Q1 (take #5, eight days later) A patient has a defect that allows mixing of blood between the right and left ventricles. The structure that has failed to form properly is the: A) Pulmonary valve B) Mitral valve C) Foramen ovale D) Interventricular septum ← right answer
Take 15 — three weeks in
By now the fact rotates between MCQ and fill-in-blank. If you actually retained the term, this is trivial. If you were memorizing question wording, this is the take where it becomes obvious that you weren't learning.
Q1 (take #15, three weeks in) Fill in the blank. The _____ is the muscular and membranous wall that divides the right and left ventricles of the heart. Answer: interventricular septum
Anchor fact · the thing the other guides skip
On every revisit, the first three words of the stem change.
Take #1 opens with "Which structure separates..." Take #2 opens with "The wall dividing..." Take #5 opens with "A patient has a defect..." Take #15 isn't even a question; it's a fill-in-blank. The retrieval cue you would normally hang the answer on stops existing.
The fact under the question (interventricular septum, separates the ventricles, page 47 of the source PDF) stays locked. The surface form rotates between four templates: direct question, definitional blank, clinical scenario, inverse-select. The distractor pool rotates against the last two takes for the same fact.
What runs on every revisit
The block below is the rephrase pipeline that fires when you reopen a deck and a question is due for review. It's the same path for MCQ and free-response. Reading it top to bottom is the cleanest way to see why pattern-matching across takes fails.
The same fact, two takes
The toggle below holds the underlying fact constant and shows the stem you actually see on take #1 versus take #5. The right answer is the same word in both. Almost nothing else is.
Take #1 vs take #5 of the same fact
Q: Which structure separates the right and left ventricles? A) Tricuspid valve B) Septomarginal trabecula C) Interventricular septum (right answer) D) Moderator band
- Direct question template
- Right answer is C
- Distractor pool: tricuspid valve, septomarginal trabecula, moderator band
The held-out eval, in numbers
Three source documents (PDFs, including a slide deck and a textbook chapter) were held out. Each tool generated a quiz from the same three documents. Every output was graded on the four-criterion rubric: factual correctness, clarity, distractor quality, and question-type coverage. Same documents, same rubric, same graders. The same rubric is what gates the rephrased questions on every revisit, which is why the surface forms can change without the quality drifting.
Higher is better. The 23.5-point gap between the top of this list and the bottom is roughly the difference between a quiz where most questions are usable and one where you spend half your study time editing the model's mistakes back into shape.
A typical PDF quiz tool vs a re-takeable one
The first row is the same in both products. The differences start at take #2.
| Feature | Typical PDF quiz tool | Studyly |
|---|---|---|
| Take #1 of the quiz | Generated from your PDF. | Generated from your PDF. |
| Take #2 | Identical questions, sometimes shuffled answer order. | Stem reworded, distractors reshuffled, distractor pool can rotate. |
| Take #5 | Same questions you've now memorized by their first three words. | Surface forms have rotated several times; you're forced to recognize the fact, not the phrasing. |
| Pattern-matching defense | None beyond shuffling A/B/C/D positions. | Stem rewrite blocks first-words pattern-matching; distractor rotation blocks 'X is always wrong' pattern-matching. |
| Quality gate on revisit | No quality re-check on rephrased output. | Same four-criterion rubric runs at revisit time. Failed rephrases regenerate. |
| Source grounding on the fifth take | Explanation paraphrased, drift from the original PDF possible. | Explanation still quotes the original PDF page verbatim, regardless of which surface form you saw. |
What it takes to actually get re-takeable quizzes from a PDF
Four steps, but only the third is unusual. The first two match almost any quiz tool. The fourth is the part that makes the daily habit stick over weeks: a tree grows per deck, and the tree only moves forward when you answer correctly across rewordings, not when you remember the question.
- 1
Drop the PDF in
Lecture deck, textbook chapter, paper. The first quiz appears in about 60 seconds.
- 2
Take the quiz
Answer; wrong answers explain themselves with a verbatim quote from the page in your PDF.
- 3
Come back tomorrow
The same facts, rewritten. Different stems, different distractors, same source.
- 4
Tree grows when you understand
Right answers across rewordings move the tree forward. Memorized phrasings don't.
When the rephrasing isn't worth it
A few honest cases where the re-takeable behavior is overkill or beside the point.
- You're cramming the night before. If you're going to take the quiz exactly once before the exam, the rephrasing on take #2 doesn't help you. Studyly still works for this; it's just that the unique part of the product is wasted on a one-shot session.
- Computational problems. If your PDF is a calculus problem set or a physics textbook chapter that wants step-by-step solutions, you want a math problem-solver, not a quiz tool. Studyly handles concept questions, not the mechanics of working through an integral.
- You only need an Anki export. The .apkg export is a frozen snapshot. Anki has its own scheduling, but it doesn't auto-rephrase. If portability is the goal, you can export, but you lose the part of the product that makes take #5 different from take #1.
Try it on tomorrow's lecture PDF
Drop a PDF in. Come back in two days. See if take #2 looks different.
Free tier on app.jungleai.com, no credit card. Email gate sends a one-click access link.
Common questions about generating a re-takeable quiz from a PDF
Why does it matter what happens on take #2 of a PDF quiz?
Because a quiz that doesn't change between sittings is a memory test for the question wording, not the material. If the stem opens with 'Which structure separates...' both times, you don't have to remember the anatomy to get it right; you just have to remember that 'C' was the answer last Tuesday. The first three words become the answer key, which is why most students stop learning anything from a PDF quiz after the second take.
What changes on a Studyly revisit, exactly?
Three things. The stem is reworded to a different surface form (different opening, different sentence shape, sometimes inverted from question to fill-in-blank). The distractors are reshuffled, so 'C' is no longer always the right answer. And the distractor set itself can rotate from a wider pool, so you can't memorize 'septomarginal trabecula is always wrong.' The fact under the question stays the same; nothing else does.
Does the underlying source PDF stay the same?
Yes. The PDF you uploaded is the source of truth for every revisit. The fact each question is testing is locked. What rotates is the surface form of the question, the distractors, and the order. When you tap 'explain my mistake' on any revisit, the response still pulls a verbatim quote from a specific page of your original PDF.
How is this different from just shuffling answer order?
Most quiz tools shuffle the position of A/B/C/D between takes. That defeats positional pattern-matching but does nothing about the stem. After two takes you've memorized 'the question that starts with which structure separates is asking about the septum.' Studyly changes the stem itself, so the cue you'd pattern-match on stops existing.
How many distinct surface forms does one underlying fact have?
Practical answer: enough that on quiz #15 from the same PDF, you've usually seen at least 8 different stems for the highest-frequency facts. The model rephrases on demand at revisit time rather than pre-generating a fixed pool, so the surface forms aren't capped. You hit diminishing returns somewhere around take #20 on a small PDF, where the underlying fact set becomes the limit.
Does rephrasing ever change what the question is testing?
It shouldn't, and the rubric gate exists to catch it when it tries. Every reworded question goes through the same four-criterion check (factual correctness, clarity, distractor quality, type coverage) before it's served. If the rephrase shifts the meaning, the question is regenerated. The check costs a couple hundred milliseconds; it's the difference between drilling a fact and drilling a moving target.
Can I export a re-takeable quiz to Anki?
Yes, but with a caveat: Anki doesn't auto-rephrase. The .apkg export is a snapshot of the questions as they were on the day you exported. If you want the rephrasing behavior, drilling inside Studyly is what you want; the Anki export is for portability when you already know the wording isn't going to be the test.
What's the eval score, and what does it actually measure?
Studyly scores 81.3 out of 100 on a held-out three-document eval, against Unattle 78.0, Gauntlet 68.0, and Turbolearn 57.8. The eval grades on factual correctness, clarity, distractor quality, and question-type coverage. It does not directly grade rephrasing quality across revisits, but the same rubric runs at revisit time as a gate, so a rephrased question that fails the rubric never ships to you.