Skip to content

How it works

Six steps. Four checks. One quiz.

Behind every Witan question is a small team of specialists working in the background. Here's exactly what each of them does — and why your questions come out better than what a single AI call could produce.

Scroll to meet the team

The team

Six specialists, one quiz

When you click Create quiz, these six pass the work to each other in order. They never wait for you — and you never wait for them.

  1. Step 1

    The Librarian

    Finds the right passages in your material

    She looks at what the next question should be about — say "Photosynthesis". Then she pulls exactly the paragraphs from your uploaded material that actually mention it.

    She does not haul the whole textbook over to the next desk. Just the relevant snippets. Takes a fraction of a second.

  2. Step 2

    The Drafter

    Sketches the question on a notecard

    He reads what the Librarian brought and scribbles a short notecard: "I want to ask where the light reaction happens. The correct answer is thylakoid membrane. Three plausible wrong answers: stroma, mitochondria, cytosol."

    He does not write the full sentence yet — just the skeleton. That is faster, and easier for the next two desks to check.

  3. Step 3

    The Bouncer

    Checks two things before letting the card pass

    Before anything is handed to the next desk, the Bouncer checks two things in a fraction of a second:

    • "Have we asked this before?" — he searches across every quiz you ever made (yes, the old ones too). If yes → back to the Drafter: "Try something else."
    • "Is this actually in your material?" — he double-checks the Drafter did not invent the answer but pulled it from a real passage.
  4. Step 4

    The Examiner

    Decides if the question is worth asking

    The Examiner looks the notecard over and gives a sharp verdict: keep or discard. She tosses questions that are…

    • trivial ("What is photosynthesis?" — everyone already knows)
    • vague ("Explain photosynthesis in general")
    • ambiguous (more than one answer could be right)
    • poorly distracted (the wrong options are obviously wrong)
  5. Step 5

    The Writer

    Polishes the sentence you actually read

    Only now — once every check is passed — does the Writer turn the notecard into the readable sentence. The hard thinking is already done; he just makes it sound right.

    This is the part you see typing live in front of you, character by character. Everything else happened in the background.

  6. Step 6

    The Archivist

    Files the question so it never appears twice

    The finished question lands in your quiz. At the same moment, the Archivist quietly logs it — so next time, the Bouncer knows this exact question already exists and will not let an accidental duplicate through. Not this week, not next month, not in another quiz set entirely.

What changed

One AI call vs. a real pipeline

A single AI call writes one sentence and hopes for the best. A pipeline catches the trivial, the repeated, and the made-up before they reach you.

A single AI call

The old way
  1. 1

    Read your entire material into the prompt.

  2. 2

    Ask the AI to write a question. Any question.

  3. 3

    Check that the JSON has the right shape.

  4. 4

    Ship it.

Nothing else is checked.

No "is this trivial?", no "have we asked this before?", no "did the AI invent the answer?" The first thing that comes out is what you get.

Witan's pipeline

The new way
  1. 1

    Pull only the relevant passages — not the whole book.

  2. 2

    Draft a question skeleton (not the full sentence yet).

  3. 3

    Two gates: "already asked?" and "really in your material?"

  4. 4

    Examiner verdict — keep or discard the draft.

  5. 5

    Only survivors get written into the readable sentence.

  6. 6

    Archived so it never appears again — in any quiz.

Four checks before you ever see it.

And because the steps overlap in parallel, the first question still appears as fast as before.

0×

specialists working on each question

0

quality checks every question must pass

0%

of duplicate questions blocked before they're written

The pedagogy behind it

Not just AI. Actual learning science.

Three principles inform every question we generate and every review we schedule. None of them are guesses — they come from decades of research on how humans actually retain knowledge.

Bloom's Taxonomy

Difficulty grows with you

Questions move from recall ("what is the mitochondrion?") up through apply, analyse, and eventually evaluate and create — the same six rungs Bloom defined in 1956 and educators have used ever since.

The Examiner uses these rungs to spot mismatches: a "recall" question that asks you to evaluate is rejected, and so is an "evaluate" question with a one-word answer.

Source anchoring

Every answer points to a passage

The Drafter is required to cite a real passage from your material as the source of every answer. Not paraphrase, not "based on general knowledge" — a literal quote.

The Bouncer then verifies the passage actually exists. This is the single biggest reason Witan does not hallucinate answers the way generic AI chats do.

FSRS scheduling

The forgetting curve, predicted

Every time you rate a question, FSRS — the Free Spaced Repetition Scheduler — updates its model of your memory and predicts the ideal moment to ask again. Easy answers push the next review further out. Wrong answers pull it back in.

You don't decide when to review. The scheduler does. That is the difference between studying and actually remembering.

Now you know what's happening.

Drop in some material and let the team get to work. The first question appears in about a second.