AI for TeachersLesson 21 of 57

Contents

AI for Teachers

Lesson 21 of 57. Sign in to keep your place.

Module 3 · Planning and making materials

A question bank you can trust

9 min read

The two-pass procedureGenerate thequestionsAs CSV: id,topic, type,difficulty,question.Nothing else.Open a freshchatSo it has nomemory ofwriting thequestions.Paste thequestioncolumn aloneAsk foranswers, andnothing butanswers.Reconcilethe twocolumnsEverydisagreementis a questionto read.Check thenumbersseparatelyA calculator,or code thatactuallycomputes.Record whatthe classgot rightThe seventhcolumn. Aftera year itbeats anygenerateddifficultyrating.Written in one stream, the answer is predicted from the question the model has just written, so abroken question gets an answer that pretends it works. Generated from the questions alone, in a freshchat, the two columns disagree wherever the question is bad. About eleven wrong answers in threehundred is a typical one-pass rate; the reconciliation catches most of them.
Written in one stream, the answer is predicted from the question the model has just written, so a broken question gets an answer that pretends it works. Generated from the questions alone, in a fresh chat, the two columns disagree wherever the question is bad. About eleven wrong answers in three hundred is a typical one-pass rate; the reconciliation catches most of them.

Figure

From worksheets to a bank

A worksheet is used once. A bank is used for years. The difference is structure: a question in a spreadsheet, with its answer, its topic, its type and its difficulty, can be pulled into a starter, a homework, a test or a revision pack in seconds. Departments that build one stop generating the same questions every term.

The model makes the bank fast to fill. The procedure in this lesson makes it safe to use.

The columns

Six columns are enough:

  • id: a number, never reused.
  • topic: from your syllabus list, exact wording.
  • type: recall, application, inference, multi-step, and so on. Your own labels, kept short.
  • difficulty: 1 to 3, your initial guess, corrected later by data.
  • question: the text.
  • answer: the text, or the mark scheme for longer questions.

Ask for output as CSV with exactly those columns, and the formats lesson explained why: it pastes straight into Google Sheets, Excel or LibreOffice Calc.

Generate the answers separately

This is the procedure that matters. Do not generate question and answer in one pass.

When the model writes a question and its answer in the same stream, the answer is predicted from the question it just wrote, and any error in the question’s setup flows straight into the answer with total consistency. A word problem whose numbers do not resolve gets an answer that pretends they do.

Instead: generate the questions. Paste the question column, alone, into a new chat. Ask for answers. Now you have two columns produced independently, and the disagreements are where the errors are. Eleven wrong answers in three hundred is a typical rate for a one-pass bank; a two-pass reconciliation catches most of them, because a bad question tends to produce different wrong answers on different runs, and the disagreement flags it.

For numerical questions, the calculator rule applies on top: the numbers get checked by something that computes.

Tag by what the question demands

Ask the model to label each question’s type as it writes it, and then disbelieve the labels. Its idea of “inference” is loose, and around a third of what it labels application is recall in a costume. A quick read of ten questions per type recalibrates you, and the second message fixes the rest: “Questions 12 to 30 labelled application only require the student to recall the formula. Rewrite them so the student has to decide which formula applies.”

The type column is what makes a bank useful. A test with twelve recall questions and none of anything else is a test of memory. Being able to filter by type is how you avoid setting one by accident.

Difficulty comes from the students, not the model

The model’s difficulty rating is a guess about an average class somewhere. Yours is a guess about your class. Both are wrong until students have answered the questions.

So the seventh column, added over the term, is the proportion of the class that got each question right the last time it was used. A question you rated 3 that 90 percent got right is a 1. A question rated 1 that 40 percent missed is telling you something about a misconception, and belongs in next week’s starter. After a year the bank knows your students better than any generated rating ever will. This is the single most valuable thing a question bank does and it costs a minute per test to record.

Parallel versions

For resits, for two classes that share a corridor, or for a student who missed the test, ask for a parallel version: “Same structure, same difficulty, same type, different numbers and a different context. Change the surface, keep the demand.” Then run the two-pass answer check on the new version too. Parallel versions are where the arithmetic errors hide, because the model changes the numbers and carries the old answer.

Sharing across a department

A bank in a shared spreadsheet, with the two-pass procedure written at the top as a rule, is a department resource that compounds. Add a column for who checked each question and when. Questions nobody has checked are drafts, and the sheet should say so, in a colour, so a colleague in a hurry does not print them.

The honest limit

A bank drifts from the syllabus. Topics get renamed, chapters move between years, a board changes its command words. Once a year, filter by topic and check the list against the current syllabus document, and retire what no longer fits. And a bank makes it easy to set questions without thinking about what the test is for. The type and difficulty columns are there to make the thinking faster, not to replace it.

Every lesson here is open. Your progress is recorded once you pass the AI at Work exam.

Check your understanding

A department generates 300 questions into a spreadsheet with an answer column, all in one pass. A term later they find eleven questions whose stored answer is wrong. Which change to the process would most reliably have caught those eleven before use?

Pick the one you would defend. Your answer stays private.

Next lesson · 9 minDiagrams and visuals without a designer

Stuck on something?Ask Asha, your AI tutor

How to cite: Addaly (2026) ‘A question bank you can trust’, in AI for Teachers. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-teachers/a-question-bank-you-can-trust (Accessed: 13 October 2026).