VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

How to Create Practice Questions with Super Intelligence

Three secondary students review open books and notes together in a classroom.
Three students studying together with open books at a classroom table.

How to create practice questions with Super Intelligence is not the same as asking for “20 questions on this topic”. A useful practice question has a job. It may diagnose a prerequisite, build fluency, distinguish two concepts, test retrieval, expose a misconception, train method selection, integrate skills or prove transfer.

SI makes question generation abundant. That abundance only helps if the questions are aligned to the capability being learned and if the answers are checked.

In the eduKateSG life series, Super Intelligence, or SI, is our editorial name for practical AI assistance. This article follows How to Use Super Intelligence to Find Gaps in Your Knowledge. Gap detection identifies what is missing; question design creates the evidence and practice needed to repair it.

The central rule is: generate questions from a learning purpose, not from a topic name alone.


The Practice-Question Architecture

Every useful question should define six things:

  • target skill;
  • question function;
  • difficulty state;
  • allowed support;
  • evidence expected;
  • answer or marking logic.

Target skill

What mental move should the learner perform?

Question function

Diagnose, build, retrieve, discriminate, transfer, integrate or assess.

Difficulty state

Familiar, varied, mixed, integrated or timed.

Allowed support

No help, cue, hint, formula sheet, reference or open book.

Evidence expected

Final answer, working, explanation, code, paragraph, classification or decision.

Answer logic

Correct answer, acceptable variants, rubric, common errors and verification method.

Question Function Comes First

Diagnostic question

Separates competing gap hypotheses with the smallest useful task.

Build question

Practises one target skill with controlled difficulty.

Retrieval question

Requires recall without the answer visible.

Discrimination question

Asks the learner to distinguish neighbouring concepts or methods.

Transfer question

Changes the surface while preserving the underlying skill.

Integration question

Combines several stable skills in one task.

Assessment question

Measures independent performance under defined conditions.

One topic can require all seven question types at different stages.


Diagnostic Questions Should Be Small

A diagnostic should change the learning plan.

If the learner may have either a concept gap or time-pressure gap, use an untimed task before a long practice set.

If the learner may confuse two vocabulary words, ask for one contrast rather than ten sentences.

If a programmer may misunderstand return values, ask for a tiny function where print and return produce different downstream behaviour.

The best diagnostic is often short because it isolates the uncertain mechanism.

The One-Variable Rule

During diagnosis, change one meaningful dimension at a time.

Do not simultaneously increase numerical complexity, wording difficulty, time pressure and method selection if you are trying to locate one gap.

Once the skill is understood, later transfer tasks can combine several changes.


Build Questions: From Familiar to Varied

A strong practice sequence moves through levels.

Level 1 — Familiar: same structure as the worked example.
Level 2 — Varied: different numbers, wording or representation.
Level 3 — Contrast: similar-looking cases requiring different responses.
Level 4 — Mixed: target appears among other methods.
Level 5 — Integrated: target combines with other skills.
Level 6 — Timed or realistic: performance under the conditions that matter.

SI can generate the ladder rapidly and adjust after each answer.

Difficulty Calibration

Difficulty should come from the intended skill, not accidental ambiguity.

A question becomes harder in legitimate ways: more steps, less cueing, deeper concept, method choice, unfamiliar context, tighter time or integration.

It becomes badly harder when wording is confusing, information is missing or the marking logic is inconsistent.

Ask SI to explain what makes each question harder.


Worked Example: Mathematics Question Ladder

Target: expand brackets with negative coefficients.

Diagnostic: expand -3(x + 2).

Familiar build: -4(a + 5).

Varied: 2(3x – 4).

Contrast: compare -2(x – 3) and 2(-x – 3).

Mixed: simplify 5x – 2(x + 4).

Integrated: solve 3(x – 2) = 2x + 5.

Transfer: word or geometry context requiring construction and expansion.

The sequence trains the underlying distribution skill without leaving it trapped inside one worksheet style.

Worked Example: Vocabulary Question Ladder

Target: use “reluctant” accurately.

Recognition: choose the closest meaning.

Recall: define without options.

Discrimination: compare reluctant, hesitant and unwilling.

Production: write an original sentence.

Register: decide whether the word fits formal and informal contexts.

Transfer: use it naturally in a new paragraph without a word bank.


Worked Example: Science Questions

Target: explain photosynthesis as a process using matter and energy.

Retrieval: state inputs and outputs.

Mechanism: explain the role of light energy.

Misconception probe: where does most new plant mass come from?

Transfer: predict what happens when light becomes limiting while water remains available.

Integration: connect photosynthesis to a food chain and respiration.

The question set tests more than keyword memory.

Worked Example: Coding Questions

Target: understand function return values.

Prediction: what does this function return?

Contrast: one function prints, one returns.

Debugging: downstream calculation fails because result is None or equivalent.

Build: write a function whose output feeds another function.

Transfer: refactor repeated logic into a function with a meaningful return value.


Worked Example: Writing Questions

Target: explain how evidence supports a claim.

Recognition: identify claim, evidence and explanation in a paragraph.

Diagnostic: learner receives claim + evidence and writes one explanation sentence.

Contrast: compare a sentence that repeats evidence with one that explains significance.

Production: write a paragraph on a new topic.

Transfer: revise an existing paragraph where the evidence is correct but reasoning is weak.

Multiple-Choice Questions

Multiple choice is useful for recognition, discrimination and diagnostic contrast when distractors represent meaningful errors.

Bad distractors are random.

Good distractors correspond to misconceptions or near-neighbour concepts.

Ask SI to explain why each distractor is plausible and what error it diagnoses.

Do not use multiple choice as the only evidence for a skill that ultimately requires free production.


Open-Ended Questions

Open questions reveal generation, organisation and reasoning.

They are useful when the learner must explain, construct, write, code or justify.

A good open question should still have a clear expected output and marking logic.

“Discuss this topic” is vague. “Explain two mechanisms, compare their effects and support each with one example” is more assessable.

Short-Answer Questions

Short-answer formats are efficient for retrieval and precise concept checks.

Ask one thing at a time when diagnosing.

Use several short answers before a long essay if you need to locate which concept is missing.


Question Banks Need Metadata

A useful question bank stores more than the question text.

Skill.
Function.
Difficulty.
Prerequisite.
Common error.
Answer/rubric.
Last used.
Transfer level.

This lets SI select questions intelligently rather than sampling randomly.

Avoid Repetition by Surface Only

Changing names and numbers does not always create meaningful variation.

Ask which cognitive demand changes.

If nothing changes except numbers, the learner may simply be drilling execution.

That can be useful for fluency, but it is not transfer.


Answer Generation and Verification

SI-generated questions are only useful if the answer is correct.

For Mathematics, recalculate independently where stakes matter.

For science and technical topics, ground questions in trusted sources.

For writing, use a rubric rather than one “model answer”.

For coding, run or test code where practical.

For current regulations or software behaviour, verify against current authoritative sources.

The Hidden-Answer Rule

Do not place the answer immediately below the question when retrieval is the goal.

Keep answers separated, collapsed, delayed or in a different section.

The learner should commit to an attempt first.

SI can wait for the learner’s response before revealing feedback.


Rubrics for Generated Questions

For complex outputs, define the rubric before the learner answers.

Writing: claim, evidence, explanation, organisation, language.

Science: concept accuracy, mechanism, key terminology, application.

Coding: correctness, structure, tests, readability, handling of edge cases.

Professional case: evidence, assumptions, trade-offs, recommendation, risk.

A rubric makes feedback less arbitrary.

The Rubric Fidelity Check

Ensure the question actually elicits what the rubric scores.

Do not score creativity if the prompt gives no room for choice. Do not score evaluation if the task only asks for recall.

Question and rubric must point to the same capability.


Adaptive Practice Generation

Use the learner’s latest response to choose the next question.

Correct but slow → similar skill with slightly more fluency pressure.

Correct with hint → same skill with reduced support.

Wrong due to misconception → contrast question.

Correct on familiar form → changed surface.

Correct on changed form → mixed or integrated task.

This adaptive loop is one of SI’s strongest advantages over static worksheets.

Do Not Let Adaptation Become Random Difficulty

The next question should have a reason.

Ask SI to state the transition: “I am increasing method-selection demand because execution is now stable.”

This keeps the practice path inspectable.


A Copyable Practice-Question Prompt

“Create practice questions for this exact skill, not the whole topic. State the target skill and question function first. Begin with one diagnostic. Then create a progression from familiar to varied, contrast, mixed, integrated and transfer tasks as appropriate. Change one difficulty dimension at a time during diagnosis. For every question, provide the answer or rubric separately, common error it can reveal and what should happen next depending on the learner’s response. Align to this source or syllabus: __. Do not expose answers before the learner attempts.”

Practice-Question Quality Audit

  • Does each question have a clear target skill?
  • Is its function explicit?
  • Is the difficulty caused by the intended cognition rather than bad wording?
  • Does the answer exist and verify cleanly?
  • Are distractors meaningful?
  • Does the rubric match the prompt?
  • Is there variation beyond changing numbers?
  • Does the sequence include transfer?
  • Are source-sensitive facts grounded?
  • Are answers hidden until the learner attempts?


Question Mutation: Change the Task Without Losing the Skill

Question mutation is one of the most powerful uses of SI.

Start with one verified question and create controlled variants.

Change the numbers. Change the wording. Change the representation. Remove a cue. Add irrelevant information. Embed the skill in another topic. Add a realistic time limit.

The important rule is to preserve the target skill while changing the surface.

Ask SI to state which feature changed and which feature remained invariant.

Near Mutation and Far Mutation

Near mutation changes the surface slightly. Far mutation changes the context or representation substantially.

A Mathematics near mutation changes coefficients. A far mutation hides the same algebra inside a geometry or rate problem.

A vocabulary near mutation changes the sentence. A far mutation requires the word in a different genre or speaking context.

Use near mutation first when the learner is still stabilising. Use far mutation to test transfer.


Error-Driven Question Generation

The learner’s wrong answer can become the blueprint for the next question.

If the error is sign distribution, generate a contrast set where signs are the only changing feature.

If the error is method selection, generate several questions that require classification before solving.

If the error is vocabulary register, generate sentence contexts where only one near-synonym fits naturally.

If the error is code flow, generate a tiny program where the same misconception produces a visible failure.

Practice becomes faster because the next question is derived from evidence rather than from a generic bank.

The Error-to-Question Mapping

Knowledge error: retrieval question.
Concept error: mechanism or contrast question.
Recognition error: classification question.
Execution error: focused procedural question.
Transfer error: changed context.
Boundary error: counterexample.
Integration error: multi-skill question.

This mapping turns question generation into a repair system.


Distractor Design for Multiple Choice

A distractor should represent a plausible wrong model, not random noise.

For Mathematics, distractors can reflect sign errors, denominator errors or using the wrong operation.

For science, distractors can reflect common causal reversals or keyword-only thinking.

For vocabulary, distractors can be near-synonyms with wrong register or connotation.

For coding, distractors can reflect scope, return-value or type misconceptions.

Ask SI to label the misconception represented by every distractor.

The Distractor Quality Test

A bad distractor is obviously impossible.

A good distractor is attractive to a learner holding a specific misconception and unattractive to a learner with the correct model.

If everyone can reject it without using the target skill, it is not doing useful diagnostic work.


Partial Credit and Mark Schemes

Complex questions should distinguish final-answer accuracy from method quality.

For Mathematics, allocate marks to setup, method, execution and answer where the assessment style permits.

For writing, separate claim, evidence, explanation and language.

For science, separate concept, mechanism, terminology and application.

For professional cases, separate evidence, assumptions, trade-offs and recommendation.

SI can draft the mark scheme, but align it to the real rubric or syllabus where one exists.

Mark Scheme Before Question

For important assessments, sometimes build the expected evidence first.

What must a good answer demonstrate?

Then write a question that actually elicits that evidence.

This reduces mismatch between prompt and marking.


Practice Questions Versus Assessment Questions

Practice questions can be scaffolded, targeted and repetitive.

Assessment questions should measure independent performance and minimise teaching inside the prompt.

Do not use a heavily scaffolded practice question as proof of mastery.

Likewise, do not use every assessment question as practice. Learners need some clean measurement points.

Formative and Summative Question Modes

Formative: question exists to improve learning. Feedback can be immediate and adaptive.

Summative: question exists to measure current independent performance. Help is restricted until completion.

SI should know which mode it is operating in.


Spaced Question Scheduling

Good question generation includes timing.

A question answered correctly today can return later in a changed form.

Use same-day retrieval, next-day variation, later mixed review and periodic transfer where the skill matters.

SI can schedule or select questions based on last success, help level and error recurrence.

Do not repeat every item on a rigid interval. Higher-priority or weaker skills deserve more frequent retrieval.

The Question Return Rule

When a question returns, change enough that the learner must retrieve the skill rather than remember the answer.

Change numbers, wording, order, context or output.

The returning item should test the concept again, not the learner’s memory of the exact question.


Interleaving Questions

Once several skills are stable, mix them so the learner must select the method.

Mathematics: ratio, algebra and geometry in one set.

Science: concepts from several chapters inside unfamiliar scenarios.

Vocabulary: current and older words across different writing tasks.

Coding: functions, loops and data structures inside one small program.

Interleaving questions are harder because they remove chapter labels. That difficulty is often desirable once execution is stable.

The Interleaving Boundary

Do not interleave several unstable skills too early.

If every component is weak, the mixed task produces noise rather than useful evidence.

Stabilise enough components first, then test selection and integration.


Worked Question Bank: Secondary Mathematics

Diagnostic: two prerequisite questions for the target skill.

Build: three familiar but non-identical questions.

Contrast: two similar questions requiring different method choices.

Mixed: target among three other current skills.

Transfer: unfamiliar wording or context.

Timed: one short set under realistic pressure.

Review: one old error reappears after delay.

SI can generate the bank from the learner’s active corridor rather than the whole syllabus.

Worked Question Bank: Science

Use retrieval, diagram interpretation, causal explanation, misconception probe, unfamiliar application and evidence evaluation.

Do not let all questions become keyword recall.

A strong science bank moves from naming to mechanism to transfer.

Worked Question Bank: Vocabulary and English

Use recognition, active recall, synonym contrast, register choice, sentence production, paragraph integration and delayed writing transfer.

Question generation should reflect the final output—speaking, writing, reading or listening—not only word-list memory.

Worked Question Bank: Coding

Use prediction, trace execution, debugging, code completion, requirement decomposition, implementation, test writing and refactoring.

Questions should increasingly remove scaffolds and require the learner to make design choices.


Question Provenance

For curriculum-aligned or professional practice, record where the question came from.

Official past paper, teacher-created, textbook-derived, SI-generated from a verified template or fully synthetic.

Provenance matters because official assessment style may differ from generic practice.

SI-generated questions are excellent for variation, but representative assessment should include authentic source-aligned material where available.

Authenticity Checks

Ask whether the generated question matches the real environment’s vocabulary, permitted methods, difficulty, format and marking expectations.

A question can test the right concept and still be poor exam preparation if its style is unlike the actual assessment.


The Bad-Question Audit

  • Does the question have more than one plausible interpretation?
  • Is required information missing?
  • Does wording create difficulty unrelated to the target?
  • Is the answer actually unique where it should be?
  • Does the rubric reward something the prompt did not ask for?
  • Is the question outside the stated syllabus or source?
  • Does a distractor become correct under a reasonable interpretation?
  • Is the model answer wrong or overconfident?

Discard or repair questions that fail the audit. Abundant generation means there is little reason to keep a flawed item.

Question Versioning

If a question is corrected, keep a simple version note when the bank matters.

“v2: fixed ambiguous wording.” “v3: updated answer after source check.”

This prevents old bad versions from re-entering a shared question bank.


A 30-Day Practice-Bank Build

Week 1: create diagnostic and familiar build items for one active corridor.

Week 2: add contrast and varied items based on observed errors.

Week 3: add mixed, transfer and timed tasks.

Week 4: audit answer quality, remove weak items and tag the most diagnostic questions for reuse.

At the end of the month, the bank should be smaller and more purposeful than a raw dump of generated questions.

The Question Bank Continuity Card

Active skill: __.
Diagnostic items: __.
Build items: __.
Transfer items: __.
Common distractors: __.
Verified answers: __.
Next difficulty change: __.

This card is enough for SI to continue adaptive practice without regenerating the whole bank.

The Final Question-Design Rule

Do not judge a practice question by how clever it looks.

Judge it by what evidence it produces about the learner and what useful next action follows from the answer.

The best generated question is the one that makes the learner’s current capability more visible.


Adaptive Question Sequencing

A question bank becomes much more useful when the next item depends on the previous response.

Correct and fast → increase variation or integration.

Correct but cue-dependent → repeat the skill with less support.

Wrong due to procedure → isolate the procedure.

Wrong due to method selection → use contrast and classification.

Wrong due to misconception → generate a counterexample or diagnostic contrast.

Correct on familiar form but wrong on changed form → stay in transfer practice.

SI can implement this branching instantly, but the branch logic should remain visible enough that the learner or teacher can inspect it.

The Next-Question Rule

The next question should answer one of three needs: confirm, repair or advance.

Confirm: was the previous result representative?

Repair: what smaller task targets the observed gap?

Advance: what harder or more varied task is justified by the current evidence?

This rule prevents random question jumps that feel adaptive but do not form a learning sequence.


Question Chains

A question chain is a short sequence where each answer determines the next cognitive step.

Example for Mathematics: classify method → solve → verify → explain why the method fit → solve a changed problem.

Example for science: predict → explain mechanism → compare with evidence → revise explanation → apply to new scenario.

Example for coding: predict output → run mentally → identify bug → propose fix → write test.

Chains are useful because they test reasoning structure rather than isolated answers.

The Chain Length Rule

Keep diagnostic chains short enough that the first error remains identifiable.

Long multi-step chains are better for integration after component skills are stable.

Do not use a ten-step case to diagnose a two-step misconception if a simpler item will do.


Feedback-Linked Questions

A strong practice system connects feedback to another task rather than stopping at explanation.

Wrong answer → identify error → explain or cue → learner retries a changed item.

The retry matters because it proves whether the feedback changed capability.

Ask SI to generate the repair question before giving a long post-mortem.

The Immediate Retry Versus Delayed Retry

Use an immediate retry to check whether the repair was understood.

Use a delayed retry to check whether the learning remained available after the explanation disappeared.

Both are useful. Immediate success alone is weak evidence of retention.


Question Difficulty Has Several Dimensions

Difficulty is not one slider.

Knowledge load: how many facts or concepts must be retrieved?

Selection load: does the learner know which method applies?

Execution load: how many steps are required?

Representation load: must the learner translate between words, symbols, diagrams or code?

Integration load: how many skills must work together?

Time load: how much performance pressure exists?

Noise load: how much irrelevant information must be ignored?

Ask SI to change the intended dimension deliberately.

The Difficulty Attribution Rule

After a learner fails a harder question, ask which difficulty dimension caused the failure.

If the difficulty came from confusing wording rather than the target skill, repair the question, not the learner.


Open-Book Practice Questions

Not all learning requires closed-book recall.

Open-book questions can test source navigation, interpretation, synthesis and application.

Examples: use documentation to solve a coding task, use a formula sheet to choose the correct equation, use a policy document to answer a scenario, use notes to compare two theories.

The target skill determines whether access to references is appropriate.

Closed-Book Practice Questions

Closed-book questions are useful when retrieval itself is part of the target.

Use them for vocabulary, core formulas, definitions, method recall and exam conditions where external reference is unavailable.

Do not make everything closed book simply because memorisation is easy to measure.


Oral Practice Questions

Oral questions expose retrieval speed and explanation fluency.

They are useful for language, oral examinations, concept explanation and professional communication.

Ask for short spontaneous answers before allowing notes.

If the learner can write accurately but cannot explain orally, the gap may be output-specific rather than conceptual.

Diagram and Visual Questions

Where the subject uses diagrams, charts or spatial relationships, question banks should include them.

Ask the learner to interpret, label, predict, transform or explain the visual rather than only recall text.

SI can describe or generate structured visual tasks where supported, but important diagrams should be checked for correctness and clarity.


Question Sets for Misconceptions

A misconception set should include three stages.

Prediction: reveal the learner’s current model.

Contrast: show where the wrong model and correct model diverge.

Transfer: require the corrected model in a new case.

This is more efficient than flooding the learner with routine questions after a conceptual error.

Question Sets for Fluency

Fluency practice can use higher repetition because execution speed matters.

But monitor accuracy. Faster repetition of the wrong procedure strengthens the wrong habit.

Use short sets, immediate error correction and periodic transfer.


Question Sets for Method Selection

Do not ask the learner to solve every question first.

Give ten questions and ask only: which method would you choose, and why?

This isolates recognition from execution.

Then solve only the items where classification was uncertain or wrong.

This can save substantial practice time while targeting the real gap.

Question Sets for Verification

Give an answer and ask the learner to check whether it is correct.

Mathematics: substitute or estimate.

Science: check consistency with mechanism or units.

Coding: write a test case.

Writing: check whether evidence actually supports the claim.

Verification questions build error-detection capability, which is often neglected in ordinary worksheets.


Question Sets for Recovery

Give a flawed solution and ask the learner to find the first error, explain its effect and repair the work.

This is valuable because real performance often involves recovering from mistakes rather than producing perfect work from the first step.

SI can create flawed examples representing known learner errors.

Clearly label them as intentionally flawed so incorrect material is not mistaken for a model solution.


Exam-Style Question Generation

If the learner is preparing for an examination, practice should reflect the real assessment without pretending SI-generated items are official past-paper questions.

Use official specifications, past papers and rubrics to define question style, command words, mark allocation and difficulty.

Ask SI to create original questions matching those structural features.

Label the generated items as synthetic practice.

Command Words Matter

Define, describe, explain, compare, evaluate, justify and calculate require different outputs.

Generated questions should use command words consistently with the marking logic.

SI can help build a command-word bank aligned to the learner’s assessment system.


Professional Practice Questions

Professional learning benefits from cases rather than trivia.

Ask the learner to analyse incomplete information, identify assumptions, compare options, state risks and recommend a next action.

Then vary the case so the same framework must adapt.

SI can generate scenario diversity rapidly, but source-sensitive requirements should be grounded in current professional references.

Case Question Rubrics

Score evidence use, reasoning, assumptions, trade-offs, recommendation and uncertainty rather than expecting one identical answer.

This makes practice closer to real judgement.


Question Quality and Learner Trust

A few bad generated questions can reduce learner trust in the entire system.

Use smaller verified sets for core learning rather than massive unreviewed banks.

Allow the learner to challenge a question.

If wording is ambiguous or the answer key is wrong, correct the bank and do not count the learner’s response as a gap.

The Question Challenge Protocol

If the learner disputes the answer, pause marking.

Restate the question. Check the source or calculation. Identify whether multiple interpretations are reasonable. Correct the item if needed.

A good practice system teaches learners that answer keys can also be examined.


The Practice-Question State Machine

Questions can have states: draft, verified, active, diagnostic, transfer, retired or flawed.

Do not keep using a flawed item because it already exists.

Promote useful questions that reliably reveal the target skill.

Retire items that are memorised or no longer add information.

The Question Reuse Rule

Reuse structure more often than exact text.

A strong diagnostic pattern can generate many fresh variants.

This preserves question quality while reducing answer memorisation.


The 15-Minute Adaptive Practice Session

Minute 1–2: one retrieval or diagnostic item.

Minute 3–6: two targeted build questions.

Minute 7–10: one contrast or mixed question.

Minute 11–13: one transfer item.

Minute 14–15: record error and next question type.

The exact timings vary. The principle is that every session produces evidence and changes the next session.

The Question Ownership Rule

Advanced learners should eventually be able to design some of their own questions.

Ask them to create one diagnostic, one transfer and one boundary question for the concept.

Then let SI critique whether the questions really test the intended skill.

Question design becomes another form of understanding.

The Final Practice-Question Rule

Generate fewer questions with clearer jobs.

Know what each item is supposed to reveal, how the answer will be checked and what the next step is after the response.

Abundant question generation becomes educational leverage only when every question remains connected to a learning decision.


Question Generation Should Follow the Learning State

The same topic needs different questions at different stages. A learner who does not know the definition needs retrieval and concept questions. A learner who knows the definition but misuses the idea needs application. A learner who performs on familiar examples needs variation and transfer. A learner who performs accurately but slowly needs fluency or timed integration.

Ask SI to state the learner state before generating questions. This prevents a question bank from becoming a random pile of difficulty.

The Question State Machine

Use six states: diagnose → repair → build → vary → integrate → assess.

Diagnose: isolate the likely gap.
Repair: target the unstable concept or procedure.
Build: repeat the skill with controlled variation.
Vary: change representation or context.
Integrate: combine with neighbouring skills.
Assess: remove support and test under realistic conditions.

After each answer, the question state can move forward, stay in place or step backward.


Question Difficulty Has Several Dimensions

Difficulty is not one number.

Concept depth: how much understanding is required?
Number of steps: how many operations or ideas must be coordinated?
Cue level: does the question reveal the method?
Representation: familiar or changed form?
Context: routine or unfamiliar?
Integration: one skill or several?
Time: untimed or performance constrained?
Distractors: how much irrelevant information is present?

When SI says it is making a question “harder”, ask which dimension changed.

The Difficulty Isolation Rule

During diagnosis, change one or two dimensions only. During transfer, combine dimensions intentionally. This distinction matters because diagnostic clarity and realistic performance are different goals.


Designing Distractors That Reveal Misconceptions

A strong multiple-choice distractor is not merely wrong. It represents a plausible mistaken rule.

In Mathematics, one distractor can result from a sign error, another from applying the operation to only one term. In science, distractors can represent confusion between cause and effect or between related processes. In vocabulary, distractors can be near-synonyms with the wrong register or connotation. In professional learning, distractors can represent common but incomplete decision criteria.

Ask SI to label the misconception behind every distractor. If a distractor has no diagnostic meaning, remove it.

Distractor Quality Check

A distractor should be plausible to a learner with a specific misunderstanding but clearly wrong under the correct model. Avoid joke answers, obviously different lengths or wording that accidentally reveals the correct option.


Generating Short Diagnostic Sets

Before producing a full worksheet, generate three to five diagnostic items with different purposes.

Item 1: direct prerequisite.
Item 2: target skill in familiar form.
Item 3: contrast case.
Item 4: changed representation.
Item 5: integrated or transfer task.

The response pattern tells you what larger practice set, if any, is actually needed.

The Stop-Before-Worksheet Rule

If a three-question diagnostic reveals the gap, stop generating generic practice and repair the gap. Abundance should shorten the route to diagnosis, not encourage unnecessary volume.


Creating Retrieval Questions

Retrieval questions should require the learner to produce knowledge without seeing it. Useful formats include free recall, short answer, fill-the-structure, label-from-memory, explain-the-step and predict-before-reveal.

SI should avoid answer leakage through the question wording.

Example: instead of “What is the distributive property a(b+c)=…?” ask “State the property that describes multiplying a factor across every term in a bracket, then apply it to -3(x+4).”

Creating Generation Questions

Generation questions require the learner to construct an example, argument, sentence, code fragment, diagram or method. They are especially useful after recognition is stable.

Ask the learner to create a counterexample, invent a word problem using the equation, write a sentence with a target word or design a test case for a function. Generating forces deeper control than selecting from options.


Creating Contrast Questions

Contrast questions are powerful when learners confuse neighbouring ideas.

Ask: “Which method applies to A and which to B, and what feature causes the difference?” Examples include mean versus median, mass versus weight, correlation versus causation, substitution versus elimination, formal versus informal vocabulary, return versus print.

The learner should explain the decisive feature, not only name the answer.

Creating Boundary Questions

Boundary questions ask when a familiar rule stops being sufficient. They are useful for preventing overgeneralisation.

Ask for exceptions, limiting cases or conditions that invalidate the shortcut. This helps move learners from beginner rules toward expert judgement.


Creating Transfer Questions

A transfer question should preserve the underlying cognitive demand while changing enough surface features that pattern matching becomes unreliable.

Change context, representation, method cues, order, wording or output type.

For Mathematics, embed the same algebra inside geometry or a word problem. For vocabulary, require the word in a new topic and different sentence structure. For science, place the mechanism inside an unfamiliar scenario. For coding, change the requirement so the same concept must be selected without a template.

Near and Far Transfer Questions

Near transfer changes the surface modestly. Far transfer changes the context substantially. Use near transfer first when the skill is newly repaired. Use farther transfer after stable success.


Creating Integrated Questions

Integrated questions combine skills that the learner can already perform separately. The purpose is to test orchestration.

Mathematics: geometry + algebra + ratio. Writing: claim + evidence + explanation + vocabulary. Science: diagram reading + concept + causal explanation. Coding: input handling + functions + data structures + error checking.

Do not integrate unstable skills too early or the question loses diagnostic value.

The Integration Decomposition Rule

If an integrated question fails, decompose the failure and test the components independently before reteaching everything.


Creating Timed Questions

Time pressure should be added only when the underlying skill is sufficiently stable. Timed questions can test fluency, sequencing and realistic performance.

Compare untimed accuracy with timed accuracy. A large gap suggests a performance issue rather than missing knowledge. SI can generate comparable timed and untimed sets to isolate the effect.

The Timing Integrity Rule

A time target should reflect the real task or a deliberate fluency goal, not an arbitrary countdown. Unrealistic time pressure can create noise and teach rushed error patterns.


Question Sequencing Matters

The order of questions changes what they measure.

If Question 1 explains the method, Question 2 may no longer measure independent recognition. If several similar items appear together, the learner may infer the method from repetition.

For assessment, mix or separate questions so earlier items do not leak answers into later ones.

The Cue Contamination Check

Ask SI whether one question reveals the answer, vocabulary or method needed for another. This is especially important in quizzes, revision decks and mock tests.


Practice Questions for Different Output Modes

Use the output the learner eventually needs.

Listening: identify, infer, summarise.
Speaking: recall, explain, respond, defend.
Reading: locate, infer, compare, evaluate.
Writing: generate, organise, revise.
Mathematics: calculate, select, justify, transfer.
Coding: predict, write, debug, test, refactor.
Professional: diagnose, compare, decide, communicate.

Do not rely only on text questions if the target capability lives in another mode.

Oral Questions

Oral practice can reveal retrieval speed, vocabulary access and reasoning without the extra burden of writing. It is useful in language, science explanation, viva-style preparation and conceptual diagnosis.

Later, move into the actual assessment or performance mode.


Practice Questions from Source Material

When the learner has a textbook, document, lecture note or specification, generate questions from that source rather than from SI’s general memory.

Ask SI to preserve defined terminology and not introduce unsupported facts. For every factual answer, the system should be able to point back to the relevant source section where practical.

Source-Bound Question Prompt

“Use only the supplied source for factual content. Create questions that test retrieval, understanding, application and transfer. For each answer, cite the source section or quote only the minimum necessary phrase. If the source does not support an answer, mark the question unsuitable instead of filling the gap from memory.”


Question Freshness and Versioning

Questions about current software, regulations, public policy, prices or standards can become stale. Date the source set and regenerate when the underlying facts change.

When a question is corrected, keep a simple version note if it has been used repeatedly: “v2 — corrected answer key and clarified wording.” This prevents an old incorrect version from continuing to circulate.


A Full Practice Set: Secondary Mathematics

Target: choose and solve linear-equation methods under mixed conditions.

Q1 diagnostic: one direct equation to confirm procedure.
Q2 recognition: choose the first useful move but do not solve.
Q3 contrast: compare two equations and explain which is easier to simplify first.
Q4 transfer: equation embedded in a perimeter problem.
Q5 integrated: geometry diagram requiring equation construction and solving.
Q6 timed: mixed item under realistic time.

Each question reveals a different learning state. Six purposeful items can be more informative than thirty random equations.

A Full Practice Set: Vocabulary

Target: active use of ten words.

Q1 free recall from meanings. Q2 contrast near-synonyms. Q3 choose appropriate register. Q4 write one original sentence for each weak word. Q5 write a paragraph using five words naturally. Q6 delayed prompt two days later with no word list.

The set moves from memory to transfer rather than repeating definitions.

A Full Practice Set: Coding

Target: debug and design functions.

Q1 predict output. Q2 identify a return-value bug. Q3 repair without rewriting the whole function. Q4 write a new function from a requirement. Q5 create tests for edge cases. Q6 refactor repeated code into reusable functions.

The bank tests prediction, debugging, generation and design—not just syntax recognition.


Question Review after the Learner Responds

Review both the learner and the question.

Was the question aligned to the intended skill? Did unexpected wording confuse the learner? Did the answer key allow a valid alternative? Did the question actually discriminate between understanding states?

A bad question can create a false gap. The question bank itself needs quality control.

The Question Retirement Rule

Retire questions that have become memorised when the goal is transfer, contain ambiguous wording, use stale facts or no longer match the learner’s current level.

Keep strong questions as anchors and generate variants around them.


The Practice Bank Dashboard

Track only what helps select the next question.

Target skill.
Current state.
Question function.
Latest result.
Help level.
Next question type.

Do not build a massive analytics system unless the learning programme genuinely needs one.

The Final Practice-Question Rule

Every generated question should answer a hidden question from the tutor: “What do I need to learn about the learner next?” or “What capability am I trying to strengthen next?”

If the item has no clear learning purpose, abundance is not an advantage.

Generate fewer random questions and more questions that make the learner’s current state visible.


Question Calibration by Learner State

The same question can be too easy for one learner and diagnostic for another.

Use the learner state to decide what the question should do.

Untested: short baseline question.
Unstable: isolated skill with low noise.
Developing: varied forms with reduced support.
Stable: mixed and transfer questions.
Transferred: realistic integration and delayed review.
Drifted: brief diagnostic before full reteaching.

This prevents every learner from receiving the same worksheet simply because they are studying the same topic.

Question Selection Is Part of Teaching

The question chosen next communicates what the system believes about the learner’s state.

If the learner is repeatedly given recognition questions after transfer is stable, learning stalls.

If the learner receives integrated questions before the prerequisite is stable, errors become hard to interpret.

SI should justify question selection in one sentence when the sequence matters.


The Question Blueprint

Before generating an important practice item, define a blueprint.

Target skill: __.
Prerequisite assumed: __.
Question function: diagnostic/build/transfer/etc.
Difficulty dimension: __.
Allowed help: __.
Expected evidence: __.
Correct answer or rubric: __.
Common wrong model: __.
Next question if correct: __.
Next question if wrong: __.

The blueprint turns question generation into instructional design.

Blueprint Reuse

Reuse strong blueprints across new content.

A diagnostic blueprint that distinguishes recognition from execution can be applied to algebra, grammar, coding and professional procedures.

The content changes. The instructional logic remains reusable.


Question Validation Before Use

Validate high-value generated questions before using them for diagnosis or assessment.

Check correctness, clarity, source alignment, unique answer where required, intended difficulty and rubric fit.

For Mathematics, solve the question independently.

For coding, run or reason through the code and test edge cases.

For science, compare with the trusted source and inspect causal wording.

For writing, check that the task actually permits the evidence the rubric expects.

Question validation is cheaper than correcting a learner model built from a flawed item.

The Two-Answer Check

Ask whether another reasonable answer could also satisfy the prompt.

For open questions, multiple good answers may be intended. Then the rubric should express acceptable reasoning rather than one model sentence.

For closed questions, unintended second answers indicate ambiguity.

SI can search for alternative interpretations before the question is used.


Question Difficulty Auditing

After several learners or attempts, compare predicted difficulty with actual performance.

If almost everyone misses the item, the skill may be difficult—or the question may be flawed.

If almost everyone answers correctly without using the intended skill, the item may be too easy or poorly targeted.

Ask SI to review the question in light of response patterns, but keep human judgement over whether the item itself is valid.

The Response-Pattern Audit

Look at which wrong answers cluster together.

A cluster around one distractor can reveal a common misconception.

Wide random errors may suggest weak prerequisite knowledge or a confusing question.

Response patterns turn a question bank into a diagnostic instrument over time.


Questions for Retrieval Versus Questions for Thinking

Not every question should require the same kind of cognition.

Retrieval questions ask what the learner can recall.

Thinking questions ask the learner to connect, compare, infer, design, justify or evaluate.

Both matter.

A learner who can reason beautifully but cannot retrieve core facts may be inefficient. A learner who recalls perfectly but cannot apply may be brittle.

SI can balance the bank according to the real output.

The Cognition Mix

For each unit, decide what proportion of practice should involve recall, procedure, selection, explanation, transfer and integration.

The correct mix changes over time.

Early learning may contain more direct retrieval and guided procedure. Later learning should contain more selection, transfer and integrated judgement.


Question Design for Active Recall

Active recall questions should require production before recognition.

Use “What are the three stages?” before multiple choice. Use “Write the formula” before showing it. Use “Explain the term” before providing options.

SI can turn notes into recall prompts while withholding answers until the learner commits.

The next article develops this revision system in full.

Question Design for Elaboration

Elaboration asks why, how, what if and what is this connected to.

Examples: Why does this rule work? How is this concept different from its neighbour? What changes if one assumption is removed? Where else would this principle apply?

Use elaboration after basic retrieval is stable enough that the learner can reason with the idea.


Question Design for Socratic Learning

Socratic questions should uncover reasoning rather than lead the learner mechanically to the desired answer.

Ask for the claim, evidence, assumption, alternative, implication and test.

Avoid turning every question into “What do you think?” without a learning purpose.

The Socratic question sequence becomes Article 28 in this learning cluster.

Question Design for Creativity

Creativity questions can define constraints rather than one answer.

“Write three openings that create different tones while preserving the same information.”

“Design two algorithms that solve the same problem with different trade-offs.”

“Generate three hypotheses that could explain the evidence and state how each could be tested.”

The rubric should evaluate the intended creative dimension rather than forcing one model response.


Learner-Generated Questions

Ask the learner to write questions after studying a concept.

Can they create one recall question, one misconception probe, one transfer question and one boundary question?

SI can critique whether each question actually tests what the learner claims.

Question generation reveals what the learner believes is important and where their concept map may be shallow.

The Learner-Question Diagnostic

If all learner-generated questions are definition recall, they may not yet understand mechanisms or transfer.

If questions are impossible or ambiguous, the learner may not understand the boundaries of the concept.

Designing questions becomes another way to assess understanding.


Question Load and Fatigue

More questions are not always more learning.

Long banks can turn practice into endurance, reduce feedback quality and make the learner rush.

Use smaller representative sets and adapt.

If three questions already show the same misconception, stop collecting evidence and repair.

If five changed questions are independently correct, increase difficulty rather than adding twenty near-identical items.

The Question Stop Rule

Stop the set when you have enough evidence for the current decision: repair, advance, maintain or assess.

This turns question volume into a controlled resource.


Question Banks for Younger Learners

Use shorter wording, clearer source alignment and stronger adult oversight where appropriate.

Avoid overwhelming banks and hidden answer traps that turn learning into frustration.

Questions should still preserve retrieval and independent attempt.

Use age-appropriate contexts without assuming simple thinking.

Question Banks for Advanced Learners

Increase ambiguity of the real-world situation rather than ambiguity of the wording.

Advanced questions can include incomplete information, competing objectives, edge cases, source evaluation and multiple defensible answers.

The rubric should reward reasoning quality and evidence, not agreement with one phrasing.


The Question Bank Audit Cycle

Review a bank periodically.

Which questions were most diagnostic?

Which produced confusion unrelated to the skill?

Which answers or rubrics needed correction?

Which items became memorised?

Which new misconceptions need targeted items?

Retire noise. Preserve strong structures. Generate fresh variants.

The 90-Day Question System

Month 1: validate core diagnostics and build questions.

Month 2: add variation, interleaving and transfer.

Month 3: analyse response patterns, retire weak items and preserve the most informative question structures.

The bank should improve in quality while decreasing in unnecessary volume.

The Question-System Ownership Test

The learner or teacher should be able to explain why the next question was chosen.

“This is testing whether the misconception is still present.”

“This removes the method cue.”

“This is a delayed transfer check.”

If the system cannot explain the purpose, the question may be unnecessary.

The mature practice system generates fewer random questions and more deliberate evidence.

The Practice Evidence Rule

A practice set is complete when it has produced enough evidence to choose the next learning action. If the learner’s state is already clear, more questions add fatigue rather than information.

Use the smallest representative set that can justify repair, advancement, maintenance or assessment. Then change the learning state instead of continuing the same bank.

Practice quality is measured by the decisions the questions enable, not by the number of items generated.


The Question Validation Pass

Before a generated question enters a reusable bank, validate three things: the prompt, the answer and the learning function.

Prompt: is the wording complete, unambiguous and appropriate to the learner?
Answer: is the key correct, source-grounded and tolerant of valid alternatives?
Function: does the item actually test the target skill rather than a side skill?

A perfectly answered bad question is still a bad learning instrument.

The Independent-Solver Check

For important question banks, solve or review a sample independently from SI’s answer key.

This is especially valuable for Mathematics, science calculations, coding and any question where one early error can contaminate many generated variants.

If the source is current or technical, verify against the appropriate documentation or reference.

The Alternative-Answer Check

Open-ended questions may support several valid answers.

Ask SI to distinguish required elements from optional variation.

For writing, science explanation and professional cases, a rubric is often more appropriate than one model answer.

This reduces false negatives where the learner’s answer is valid but phrased differently.


Answer Keys Should Teach Less Than the Question Requires

An answer key should be sufficient to check and repair without becoming a replacement lesson for every item.

Use layers:

Layer 1: final answer or core expected elements.
Layer 2: concise working or rationale.
Layer 3: misconception note if the learner chose a common wrong path.
Layer 4: full worked explanation only when repair is needed.

This preserves retrieval and self-correction.

The Delayed Answer Rule

When retrieval is the goal, the learner should commit to an answer before feedback appears.

SI can ask, wait, record confidence and then reveal feedback.

This sequence turns a conversational interface into a genuine recall environment instead of a scrolling answer sheet.


Spaced Practice Banks

Question banks become more useful when items reappear according to learning state rather than a fixed random schedule.

Recently repaired gaps should return sooner. Stable skills can return later. High-leverage prerequisites deserve periodic maintenance. Low-priority material can appear less often.

SI can select the next questions using the gap register and previous results.

The Repetition-Without-Memorisation Rule

If the learner remembers the exact answer to a repeated question, generate a structurally equivalent variant rather than counting memorised response as fresh evidence.

Anchor questions can remain for retention checks, but transfer requires new surfaces.

Question Rotation

Rotate among retrieval, application, contrast and transfer so revision does not collapse into one response pattern.

This is particularly useful in long revision cycles where the learner can become familiar with the structure of the bank.


Exam-Mode Question Generation

When preparing for an examination, question generation should increasingly resemble the actual assessment demands.

Match format, command words, expected working, topic mixing, time pressure and marking conventions where known.

Do not invent an exam style from generic memory if official past papers or specifications exist.

Use generated questions to increase variation around the official frame, not to replace it.

The Exam Simulation Rule

Generated mock questions should be clearly labelled as generated and separated from official materials.

This protects the learner from mistaking a plausible SI question for evidence about the exact future exam.

Command-Word Practice

Different command words require different outputs: define, describe, explain, compare, evaluate, calculate, justify.

SI can generate paired questions where the content is similar but the command word changes.

The learner should explain how the required answer changes.


Teacher and Tutor Review

For formal learning, a human teacher or tutor can review a sample of generated questions rather than every item.

Ask them to check alignment, difficulty, marking logic and whether the bank overemphasises or omits any important corridor.

Then update the generation prompt from that feedback.

This turns human expertise into a reusable quality-control layer.

The Question-Generator Calibration Set

Keep a small set of approved examples showing the desired style, level and marking logic.

Ask SI to generate new items structurally consistent with the calibration set while changing surface content.

This is more reliable than repeatedly describing “make it exam standard” in vague terms.


Question Bias and Coverage

Generated banks can overproduce the easiest-to-generate question types.

For Mathematics, this may mean procedural calculation. For science, definition recall. For writing, generic essay prompts. For coding, syntax snippets.

Use a coverage matrix to ensure the bank includes retrieval, mechanism, method selection, transfer and integration where those skills matter.

The Coverage Matrix

Rows: skills or concepts. Columns: diagnostic, retrieval, application, contrast, transfer, integration and assessment.

Blank cells reveal where the bank is thin.

Do not fill every cell mechanically. Fill the cells relevant to the target capability.


A Full Practice Workflow

1. Define target skill.

2. Choose question function.

3. Generate a small diagnostic set.

4. Classify the learner response.

5. Generate the smallest appropriate repair set.

6. Validate questions and answers.

7. Increase variation only after stability.

8. Add mixed and transfer questions.

9. Retest after delay.

10. Retire memorised or low-value questions.

This workflow makes question generation evidence-responsive rather than content-volume driven.

The Practice-Question Ownership Test

Ask the learner why a question is being asked.

“This checks whether I can recognise the method.”

“This changes the context to test transfer.”

“This distractor tests the sign mistake I used to make.”

When learners understand the purpose of practice, they become better at selecting or designing their own questions.

The Final Question-Design Principle

A generated question is valuable only when it produces information or practice that changes the learner’s capability.

Question abundance should make practice more precise, not more endless.

The best question bank gets smaller as the system becomes better at choosing the right next question.

The Question Finality Rule

When a question sequence has diagnosed the gap, repaired it and confirmed transfer, stop generating more items of the same type. Move the skill to spaced maintenance or a harder integration state.

Abundant generation should shorten the route to useful evidence, not create infinite worksheets.

The strongest question system knows not only what to ask next, but when there is no educational reason to ask another question yet.

Frequently Asked Questions

Can SI create unlimited practice questions?

Yes, but unlimited volume is not the goal. Use enough questions to diagnose, build and prove transfer.

How many questions should I generate?

Generate the smallest set needed for the current learning state, then adapt from the response.

Can I trust generated answers?

Check important or technical material against reliable sources, calculations, tests or rubrics. Generated confidence is not verification.

Should all questions get harder?

Difficulty should change based on evidence. Sometimes the next question should be easier to isolate a prerequisite.

Are multiple-choice questions useful?

Yes for recognition and discrimination, especially with diagnostic distractors. Use free production for skills that require generation.

How do I create transfer questions?

Change context, representation, wording, method cues or integration while preserving the underlying target skill.

What comes next?

The next guide shows how to use SI for active recall and revision without turning review into passive rereading.


Helpful Reading

Generate Evidence, Not Just More Questions

The value of SI is not that it can produce a thousand items.

It is that it can produce the next question that matches the learner’s current state, reveals the right error and changes intelligently after the response.

A good practice question makes learning state more visible. A great sequence uses that evidence to decide what the learner should do next.