How to build a personal Super Intelligence tutor begins with a teaching contract, not a personality. A useful tutor is not simply an AI that sounds encouraging or answers quickly. It is a system that knows when to diagnose, when to explain, when to ask a question, when to withhold the answer, when to increase difficulty and when to make the learner work independently.
The strongest personal SI tutor behaves differently from a chatbot because it follows a learning protocol. It does not optimise for conversational smoothness. It optimises for evidence of human learning.
In the eduKateSG life series, Super Intelligence, or SI, is our editorial name for practical AI assistance. This article follows How to Learn Anything Faster with Super Intelligence. The previous page explains the acceleration loop; this page turns that loop into a persistent tutoring architecture.
The central rule is: a personal SI tutor should be easier to interrupt, inspect and correct than a human teacher—not harder.
The Personal SI Tutor Architecture
A strong tutor has eight layers:
- learning goal;
- learner state;
- source and syllabus frame;
- diagnostic routine;
- teaching mode;
- help ladder;
- transfer test;
- progress record.
Learning goal
What independent output should the learner eventually produce?
Learner state
What can the learner already do, where do errors begin and how much help is currently needed?
Source and syllabus frame
Which textbook, curriculum, documentation, rubric or reference defines the learning target?
Diagnostic routine
How does the tutor test before explaining?
Teaching mode
Does the current need require explanation, questioning, worked example, practice, critique or review?
Help ladder
How much assistance may be given before another learner attempt?
Transfer test
What changed task proves the knowledge is usable?
Progress record
What state survives after the conversation: target, error, help level, retest and next action?
The Tutor Contract
Write the rules before the tutoring starts.
A good contract might say:
- Do not provide the final answer before an attempt unless safety or accessibility requires it.
- Diagnose the first unstable step before reteaching the whole topic.
- Use the learner’s current syllabus or reference where provided.
- Give one hint at a time.
- Ask the learner to explain repaired reasoning back.
- Always include a changed transfer task after a repair.
- Reduce help as performance stabilises.
- Label uncertainty and source-sensitive claims.
- Do not invent progress that has not been demonstrated.
- End with the learner stating what they can now do independently.
This contract can be short. The important thing is that the tutoring behaviour is explicit and reusable.
The Learner Model
A personal tutor needs a current model of the learner, not a permanent label.
Avoid identity statements such as “weak at Maths” or “advanced reader”. Use observable states.
Stable: skill succeeds independently across changed tasks.
Developing: succeeds with light support or familiar form.
Unstable: repeated error or inconsistent application.
Untested: insufficient evidence.
Drifted: previously stable skill now fails after delay.
The learner model should change when evidence changes.
The Baseline Session
The first session should not begin with a giant lesson.
Ask the learner to produce a small sample of current ability.
For Mathematics: several prerequisite and target questions.
For writing: one short paragraph.
For vocabulary: active recall plus sentence use.
For coding: one small task from a requirement.
For professional skill: one realistic case.
The tutor then identifies the first unstable corridor and builds from there.
The Tutor Modes
Diagnostic mode
Purpose: identify where performance first becomes unstable. Tutor asks, observes and classifies. Minimal explanation.
Explanation mode
Purpose: repair one concept or relationship. Short explanation, learner teach-back, then application.
Socratic mode
Purpose: expose reasoning. Tutor asks bounded questions instead of giving a conclusion.
Practice mode
Purpose: repeat and vary a target skill. Difficulty changes based on evidence.
Feedback mode
Purpose: analyse an attempt. Tutor identifies the first unstable step and suggests the smallest repair.
Transfer mode
Purpose: remove cues, change the surface and test independent recognition.
Review mode
Purpose: delayed retrieval, error recurrence and drift control.
The tutor should state or internally follow the correct mode rather than blending all seven into every response.
Mode Switching
The mode changes when evidence changes.
If a learner cannot recall the concept, do not remain in transfer mode.
If recall is stable but method selection fails, stop explaining and move to contrast practice.
If five familiar examples are correct, stop adding familiar examples and increase variation.
SI is particularly suited to this adaptive switching because it can generate new material immediately.
The Help Ladder
A personal tutor should have a defined ladder.
0 — No help: independent attempt.
1 — Orientation: restate the goal or point to the relevant information.
2 — Cue: one short prompt.
3 — Hint: identify the first useful move.
4 — Partial scaffold: show structure but leave work incomplete.
5 — Worked repair: demonstrate the unstable step.
6 — Full example: complete demonstration when necessary.
After Level 5 or 6, the next task should immediately require the learner to reproduce or adapt the idea.
The Hint Budget
Track how much help a learner needs.
If the learner repeatedly succeeds only after three hints, the skill is not yet independent.
If hint count falls over time, learning is progressing.
Ask the tutor to record the highest help level used on each representative task.
This is more informative than simply recording correct or incorrect.
Source Alignment
The tutor needs an authoritative frame where one exists.
School learner: current syllabus, teacher instructions, textbook or marking rubric.
Coding learner: current official documentation.
Professional learner: current standard, framework or internal authorised material.
Language learner: target register, dialect or assessment requirement.
Ask the tutor to state when an explanation goes beyond the source or introduces optional enrichment.
The Source Conflict Rule
If two sources conflict, the tutor should not silently choose one.
Mark the conflict, identify which source is newer or more authoritative and ask for clarification if needed.
This is especially important in school marking conventions, changing software documentation and professional standards.
Worked Example: A Secondary Mathematics Tutor
Target: solve linear-equation questions independently.
Baseline: arithmetic stable, bracket expansion unstable when negatives appear.
Tutor enters diagnostic mode and isolates sign distribution.
Explanation mode repairs one example.
Learner teaches the sign rule back.
Practice mode gives three changed expansions.
Transfer mode embeds expansion inside an equation without naming it.
Progress record: expansion stable on four changed tasks, one copying error remains, no hint required on final two.
Next session begins with delayed retrieval rather than repeating the lesson.
Worked Example: A Vocabulary Tutor
Target: move ten words from recognition into active writing.
Baseline: learner recognises eight but can produce only three.
Tutor uses retrieval mode, then explanation for confused pairs.
Practice requires original sentences.
Transfer requires a short paragraph where five target words fit naturally.
Review happens two days later without the list.
Progress record tracks active production rather than list exposure.
Worked Example: A Coding Tutor
Target: use functions to organise repeated logic.
Tutor first asks the learner to write a small function.
Error: function prints correctly but does not return a value when later code needs the result.
Tutor explains return versus print with one contrast.
Learner repairs code.
Transfer: new task requires a function whose return value feeds another calculation.
The tutor does not paste the final program.
Worked Example: A Writing Tutor
Target: produce evidence-based paragraphs with clear reasoning.
Baseline: grammar is acceptable, but paragraphs contain examples without explaining how they support the claim.
Tutor highlights claim, evidence and explanation separately.
Learner repairs one paragraph.
Transfer: new topic, same paragraph logic, no scaffold.
Progress record tracks whether explanation-back remains necessary.
Worked Example: A Professional Learning Tutor
Target: learn project risk analysis.
Tutor uses a real low-stakes project.
Learner identifies risks; tutor challenges missing assumptions.
Tutor asks for trigger, mitigation and owner rather than supplying them immediately.
Transfer: learner analyses a different project and explains why each risk matters.
The output resembles real work rather than a quiz.
The Tutor Memory Rule
The tutor should remember or be given stable learning state, not every conversational detail.
Useful state:
- current target;
- stable prerequisites;
- active gaps;
- help level;
- latest transfer result;
- source or syllabus;
- next review date.
Avoid storing unnecessary personal information or using old labels as permanent learner identity.
When the state changes, update the current record.
The Session Start Brief
A repeatable session can start with:
“Target: __. Current stable skills: __. Active gap: __. Latest evidence: __. Maximum help level today: __. Source: __. End with one independent transfer task.”
This gives the tutor operational context without requiring a long biography.
The Session End Record
End with five fields:
What became stable?
What still fails?
Highest help level used?
Transfer result?
Next session target?
The learner should also state their own version in plain language.
The Tutor Stop Rule
A tutoring interaction should stop when the learner can continue productively alone.
Do not keep the conversation going simply because SI is available.
A good tutor hands the task back.
Examples: learner can solve the next three independently; learner has a clear practice set; learner knows which source to consult; learner has reached the planned session goal.
Parent Oversight for Younger Learners
For children, the adult should define account, privacy, source and answer-generation boundaries appropriate to the learner.
The goal is not surveillance of every interaction. It is ensuring the system supports the intended learning rather than bypassing it.
Useful adult review focuses on target, error pattern, independence and whether school requirements are aligned.
Avoid feeding unnecessary identifying information into the tutor.
Tutor Versus Homework Generator
A homework generator produces output. A tutor produces learner capability.
If the system completes essays, solves every equation or writes full code before the learner attempts, it may improve submission speed while weakening the learning signal.
The tutor contract should always identify which mental act belongs to the learner.
Tutor Failure Modes
Too agreeable
Tutor accepts the learner’s explanation without testing it. Repair: ask for changed application.
Too helpful
Tutor supplies the answer before retrieval. Repair: enforce help ladder.
Too verbose
Explanation overwhelms the learner. Repair: smallest useful explanation.
Too much practice
Volume replaces diagnosis. Repair: change task after repeated success or failure.
No source alignment
Tutor teaches an irrelevant convention. Repair: anchor to source or syllabus.
Permanent labels
“Weak” or “advanced” becomes identity. Repair: observable current state only.
No transfer
Learner succeeds only with familiar forms. Repair: changed and mixed tasks.
No fading
Support stays high even after improvement. Repair: reduce help deliberately.
A Copyable Personal SI Tutor Prompt
“Act as my personal learning tutor, not an answer generator. My target independent output is __. My source or syllabus is __. Begin with a short diagnostic. Do not explain the whole topic before seeing my attempt. Identify the first unstable step. Use one hint at a time and track the highest help level. After repair, require retrieval and a changed transfer task. Reduce help as I improve. Label anything that depends on an uncertain or external source. End each session with what is stable, what still fails, the transfer result and the next target.”
The Tutor Quality Audit
- Does the tutor diagnose before teaching?
- Does it preserve the learner’s attempt?
- Does it use a clear help ladder?
- Does it align to the correct source?
- Does it distinguish explanation from practice?
- Does it change difficulty based on evidence?
- Does it test transfer?
- Does it fade support?
- Does it maintain a compact current learner state?
- Can the learner continue without it?
If the last answer is no, the tutor may be useful but the learning system is not yet producing independence.
Adaptive Difficulty: The Tutor Should Change the Task, Not Just the Tone
A personal tutor becomes useful when difficulty changes according to evidence.
If the learner fails because the prerequisite is missing, step back. If the learner succeeds repeatedly on familiar tasks, increase variation. If the learner can solve but cannot choose the method, mix neighbouring methods. If accuracy is stable but slow, add realistic time pressure.
Ask the tutor to change one difficulty dimension at a time during diagnosis: content complexity, wording, time, method selection, number of steps or amount of support.
This preserves diagnostic clarity. A harder task is useful only when you know what became harder.
The Difficulty State
Track the current practice state: familiar, varied, mixed, integrated or timed.
Familiar practice confirms the basic procedure. Varied practice changes surface details. Mixed practice requires method selection. Integrated practice combines skills. Timed practice tests realistic execution.
A tutor should know which state it is testing instead of calling every harder question “advanced”.
Assessment Mode Versus Teaching Mode
A tutor that helps during every test cannot measure independence.
Create an assessment mode with no hints, no leading questions and no explanation until the learner submits the attempt.
Use assessment mode for baselines, delayed retrieval and transfer checks.
Then switch into teaching mode for diagnosis and repair.
This separation produces cleaner learning evidence.
The Assessment Integrity Rule
Before an assessment, tell the tutor exactly what help is allowed.
“Do not explain or hint until I finish.”
“You may clarify the wording but not identify the method.”
“Mark only after I submit all answers.”
The rule should match the purpose of the assessment.
The Tutor Error Model
A tutor should distinguish the learner’s error from its own error.
Learner error: wrong method, missing fact, arithmetic slip, weak transfer.
Tutor error: incorrect explanation, invented source, inconsistent marking, task outside syllabus or hint that gives away too much.
Keep these categories separate. Otherwise the learner may internalise a model error as a misconception.
The Tutor Self-Check
For technical or high-value material, ask the tutor to state which parts are grounded in the provided source and which are general explanation.
Where appropriate, verify against the textbook, rubric, official documentation or teacher guidance.
Do not rely on “I am confident” as evidence of correctness.
Subject Tutor Profiles
One tutoring protocol can support many subjects, but each subject benefits from a specialised profile.
Mathematics tutor profile
Priorities: prerequisite mapping, method selection, working, sign and arithmetic control, transfer into mixed questions.
Key rule: never identify the method before the learner attempts to classify the question when method recognition is part of the skill.
Science tutor profile
Priorities: concept relationships, terminology precision, mechanism, evidence, explanation structure and application to unfamiliar phenomena.
Key rule: distinguish memorised keywords from causal understanding.
English or writing tutor profile
Priorities: reading comprehension, vocabulary transfer, claim-evidence-explanation, sentence clarity, paragraph purpose, audience and revision.
Key rule: critique the learner’s draft before generating a replacement draft when writing skill is the learning target.
Language tutor profile
Priorities: active retrieval, pronunciation where relevant, listening discrimination, grammar in use, spontaneous speaking and writing.
Key rule: move vocabulary from recognition into production.
Coding tutor profile
Priorities: problem decomposition, syntax, debugging, data flow, testing, explanation of code and transfer to new requirements.
Key rule: learner predicts the bug or design before the tutor supplies the fix.
Professional tutor profile
Priorities: real cases, domain vocabulary, decision structure, source quality, deliverables and reflection on judgement.
Key rule: use realistic outputs rather than only quizzes.
The Tutor Question Bank
A strong tutor uses questions for different functions.
- Recall: What do you remember without looking?
- Diagnosis: Where did your working first become uncertain?
- Mechanism: Why does this rule work?
- Discrimination: What feature makes this case different from the previous one?
- Prediction: What do you expect before you calculate or run the code?
- Transfer: Where else would this idea apply?
- Verification: How could you check your answer independently?
- Reflection: What error would you notice earlier next time?
The tutor should ask the question type that matches the learning state rather than using Socratic questioning indiscriminately.
The Tutor Hint Bank
Hints can also be classified.
Attention hint: look at the sign before the bracket.
Recall hint: what rule did you use in the previous example?
Representation hint: draw the relationship or rewrite the equation.
Method hint: which operation would isolate the variable?
Check hint: substitute your result back.
Contrast hint: what changed from the earlier case?
A targeted hint preserves more learner cognition than a full solution.
The Tutor Explanation Budget
Explanation should be budgeted, not unlimited.
Start with the minimum conceptual repair. Ask the learner to apply it. Expand only if the next attempt shows the gap remains.
This prevents explanation from becoming entertainment.
A tutor can be brilliantly articulate and still be ineffective if the learner rarely has to think.
The Analogy Rule
Analogies are useful when they illuminate structure. They become harmful when the learner remembers the analogy but not the actual concept.
Ask the tutor to state where the analogy works and where it stops working.
Then return to formal or operational language.
This is particularly important in science, mathematics and technical subjects where a convenient analogy can hide a mechanism.
The Progress State Machine
A skill can move through states: untested → unstable → developing → stable → transferred → maintained → drifted.
The tutor should update the state only when evidence supports the change.
One correct answer should not move a skill from unstable to stable. Several changed independent tasks may.
A delayed failure may move transferred back to developing or drifted.
This state machine keeps the tutor honest about progress.
The Tutor Progress Dashboard
Keep the dashboard small.
Current target.
Current state.
Repeated error.
Highest help level.
Latest transfer result.
Next review.
Do not build a large analytics system unless the learning programme genuinely needs it.
The dashboard exists to select the next learning action.
Tutor Memory and Versioning
A tutor may carry context across sessions, but current state should remain inspectable.
Use a dated learning card rather than trusting conversational memory alone.
Example: “Algebra signs stable as of 30 September. Equation setup developing. Latest mixed set 7/10 independent. Next review after two days.”
If a skill drifts, update the record. Do not preserve an old “mastered” label because it was once true.
The Tutor Context Reset
If a long tutoring thread accumulates old mistakes and abandoned plans, start from a fresh learning card.
“Use only the current target, latest evidence, active gap and source below. Ignore older superseded practice plans.”
This prevents stale context from steering the next lesson.
Parent–Tutor Handoff
A parent does not need every chat transcript.
A useful summary includes target, current gap, evidence, independence level, practice recommendation and any question for the human tutor or school.
Avoid framing temporary struggle as identity.
“Needs two hints to identify method in mixed algebra questions” is more useful than “weak in algebra”.
Teacher–SI Handoff
When SI practice is used alongside a human teacher, carry forward only the evidence the teacher can use.
Representative attempts, repeated error, transfer result and source alignment matter more than prompt history.
This keeps the human educator in control of curricular judgement while SI reduces preparation overhead.
Tutor Hallucination Controls
A personal tutor should have explicit rules for uncertainty.
- Do not invent a textbook requirement.
- Do not claim a source says something unless it can be checked.
- Do not create a citation merely because the learner asks for one.
- Mark ambiguous marking conventions.
- Ask for the source when a local rule matters.
- Separate general teaching advice from syllabus-specific requirements.
The tutor should be allowed to say “I need the source” or “I am not certain”.
Tutor Verification Mode
When the learner challenges the tutor, switch into verification rather than persuasion.
Restate the claim. Identify the source or rule. Recalculate if relevant. Compare with the learner’s evidence. Correct the tutor if needed.
A good personal tutor models how to revise beliefs rather than defending every answer.
A Full Session: Mathematics
Target: factorise simple quadratic expressions.
Baseline: learner expands accurately but cannot consistently reverse the process.
Diagnostic: give three expressions with different coefficient patterns.
Observed error: learner searches for any two numbers with correct product but ignores required sum.
Repair: use product-and-sum contrast examples.
Retrieval: learner states the two conditions without notes.
Practice: three changed expressions.
Transfer: factorisation appears inside an equation with no section label.
Record: developing; one hint on first changed example, no hint on final two; delayed review tomorrow.
A Full Session: Writing
Target: make analytical paragraphs explain how evidence supports the claim.
Baseline: learner gives evidence but moves immediately to a new example.
Diagnostic: highlight claim, evidence and explanation in one existing paragraph.
Repair: tutor asks “What does this evidence prove about your claim?”
Practice: learner adds one explanatory sentence.
Transfer: new topic, paragraph written without scaffold.
Record: claim-evidence link stable in one changed paragraph; needs delayed retest.
A Full Session: Language Speaking
Target: use ten active vocabulary items in spontaneous speech.
Baseline: recognition 10/10, free recall 5/10.
Practice: short prompts requiring two target words each.
Feedback: tutor corrects usage after the learner finishes the sentence rather than interrupting every attempt.
Transfer: two-minute response on a new topic.
Record: eight words produced naturally, two still cue-dependent.
A 30-Day Personal Tutor Build
Days 1–3: write the tutor contract, source frame and baseline.
Days 4–7: calibrate help ladder and diagnostic routine.
Week 2: add adaptive difficulty and transfer tasks.
Week 3: add delayed reviews, progress state and source verification.
Week 4: reduce prompt complexity, test no-SI independence and remove rules that do not improve learning.
At the end of the month, the tutor should be simpler and more predictable than when it started.
Tutor Maturity Levels
Level 1: explanation assistant.
Level 2: diagnostic tutor.
Level 3: adaptive practice tutor.
Level 4: transfer and review tutor.
Level 5: integrated learning system with compact state, source alignment and human handoffs.
There is no requirement to reach Level 5. A simple Level 2 tutor can be extremely useful if it diagnoses accurately and preserves learner effort.
The Tutor Independence Test
Every few sessions, remove the tutor.
Can the learner start the task? Choose the method? Retrieve the core rule? Check their work? Explain the reasoning? Know where to look when uncertain?
If not, identify which part of the tutor has become a dependency.
Then fade that support deliberately.
The Tutor Continuity Rule
A new session should begin from current learning state, not from the entire history.
Carry forward the latest target, evidence, gap, help level and review condition.
Archive old explanations and superseded plans unless they are needed for diagnosis.
This keeps the tutor fast and reduces context noise.
The Final Tutor Rule
A personal SI tutor should become less necessary at the skill level as the learner improves.
It may remain useful for new material, fresh practice and feedback, but dependence on the same scaffold should fall.
The tutor succeeds when the learner can increasingly do the work, detect the error and choose the next step without asking the tutor first.
The Tutor State Machine
A personal SI tutor becomes more reliable when every skill has a visible state instead of one vague label such as “learning”.
Unknown: not yet tested.
Exposed: learner has seen the idea.
Guided: learner can use it with help.
Independent: learner can perform without help.
Transferred: learner can recognise and use the skill in a changed context.
Stable: learner retains it after a delay.
Needs repair: a previously stable skill has reopened.
The tutor should move a skill forward only when evidence supports the transition. One correct answer can move a skill from unknown to exposed or guided. It should not automatically move the skill to mastered.
This state machine prevents the tutor from teaching in circles. Stable skills leave the active queue. Open gaps remain visible. Reopened skills return only when evidence shows drift.
The Mastery Contract
Before teaching begins, define what mastery will look like. The tutor should know the end condition before generating material.
Mathematics: three changed questions completed independently, correct method selection, explain-back and one delayed mixed test.
Vocabulary: recall meaning, use the word accurately in an original sentence, distinguish it from nearby words and transfer it into speaking or writing after delay.
Coding: implement the concept from a new requirement, debug one failure and explain the code without relying on the tutor.
Professional learning: complete a realistic work output that another person can use, then explain the reasoning behind it.
The tutor should not invent a mastery standard silently. It should either use the curriculum, trusted source or learner’s stated outcome.
The Tutor Pacing Engine
A human tutor changes pace based on evidence. A personal SI tutor should do the same.
Speed up when the learner produces several correct changed examples with low hint use. Slow down when the learner repeatedly fails at the same early step. Hold the level when performance is mixed but improving.
The tutor should vary one difficulty dimension at a time: number complexity, reasoning steps, irrelevant information, time pressure, context novelty or integration with another skill.
This allows the learner and tutor to discover which kind of difficulty causes performance to deteriorate.
The Difficulty Calibration Rule
Difficulty should create useful struggle, not random failure.
If every question is easy, the tutor is measuring fluency rather than extending capability. If every question is impossible, the learner receives little diagnostic information and may lose orientation.
A useful tutor aims for a zone where the learner can attempt, reveal reasoning and benefit from small corrections.
The Self-Explanation Rule
After a correct answer, the tutor should sometimes ask why it is correct.
Correct performance can come from pattern matching or remembered procedure. Self-explanation reveals whether the learner understands the mechanism.
Useful questions include: Why does this step work? What assumption are you using? When would this method fail? What cue told you to choose this method? What changed from the earlier example?
The tutor should not demand explanations after every trivial step. Use explain-back at important concept boundaries and before marking a skill independent.
The Error-Repair Loop
When an answer is wrong, the tutor should not begin with the full solution.
1. Locate: find the first unstable step.
2. Classify: knowledge, concept, recognition, execution, transfer or retrieval.
3. Repair: use the smallest intervention.
4. Retest: use a changed task.
5. Delay: check again later if the skill matters.
This loop is much faster than restarting the whole chapter after every error.
The Hint Economy
Hints are a learning resource. Spend them deliberately.
The tutor can track a hint budget for each task. If the learner uses three strong hints, the task should not be counted as independent success.
Over time, the tutor should try to reduce the average hint level required for the same skill class.
Hint reduction is often a more useful progress signal than raw question count because it shows decreasing support dependence.
The Wait Rule
Do not give hints instantly when the learner hesitates.
Use a short wait period appropriate to the task. For a simple recall question, a few seconds may be enough. For a multi-step problem, the learner may need several minutes to explore.
The tutor should distinguish productive struggle from unproductive confusion. Productive struggle contains a plausible line of reasoning. Unproductive confusion shows no workable entry point or repeated random attempts.
Spaced Review Inside the Tutor
A personal tutor should know when to revisit stable skills without turning the learner’s life into a reminder system.
Use sparse review: same-day retrieval, next-day check, several-day transfer and later mixed review for important skills.
Review interval should expand when the learner succeeds and contract when the skill reopens.
The tutor should vary the question on each review. Repeating the exact same prompt can measure memory of the prompt rather than the concept.
Review Priority
Not every learned item deserves equal review.
Prioritise prerequisites, high-frequency skills, load-bearing concepts, recurring errors and skills required for the learner’s next target.
Low-value trivia should not crowd out foundational skill repair.
The Tutor Curriculum Layer
A personal tutor should not become the curriculum by accident.
For school learning, anchor to the relevant syllabus, textbook, teacher sequence or assessment objective. For coding, anchor to current documentation and the target project. For professional learning, anchor to the actual work outputs and standards.
SI can suggest an efficient route through the material, but the route should remain connected to an authoritative scope.
The Source-of-Truth Rule
The tutor should know which source settles factual disagreement.
If SI and the textbook disagree, inspect both. If a coding library has changed, current documentation outranks remembered examples. If school instructions change, current teacher or official material may supersede older notes.
The tutor’s confidence should never be higher than the evidence supporting the lesson.
Tutor Memory: What to Remember and What to Forget
Good tutor memory is selective.
Remember learning state, not every conversational detail.
Useful persistent fields include current goal, stable skills, open gaps, recurring errors, hint level, mastered prerequisites, source frame and next review.
Discard abandoned explanations, obsolete assumptions, old practice variants and sensitive information that is not needed for teaching.
A compact tutor state is easier to audit and less likely to drag stale information into the next session.
The Tutor Reset
When a tutor conversation becomes messy, reset from current state.
Prompt: “The current learning goal is __. Stable skills are __. Open gaps are __. The latest evidence is __. Ignore abandoned approaches and build the next session only from this state.”
This prevents old errors and old explanations from contaminating the current teaching plan.
The Session Architecture
A strong recurring tutor session can use a fixed skeleton.
1. Retrieval: one or two items from previous learning.
2. Diagnostic: current target attempt.
3. Repair: minimal explanation or hint.
4. Practice: several targeted items.
5. Transfer: one changed task.
6. Reflection: learner explains the key idea.
7. Record: update tutor state and next review.
This structure keeps the tutor from spending the whole session explaining.
The 20-Minute Tutor Session
For a bounded skill: 3 minutes retrieval, 4 minutes diagnostic, 5 minutes repair, 5 minutes changed practice, 3 minutes transfer and record.
The times are illustrative. The principle is that explanation occupies only part of the session.
The 60-Minute Tutor Session
For deeper learning: 10 minutes retrieval and review, 10 minutes diagnostic, 15 minutes explanation and guided practice, 15 minutes independent transfer, 5 minutes error review, 5 minutes state update.
If the learner spends nearly the entire hour listening, the tutoring design is too passive.
The Tutor’s Question Bank
Use recurring questions that expose learning state.
- What can you do without help?
- Where does your working first become uncertain?
- What made you choose this method?
- What would change if this condition changed?
- Can you explain this without the original wording?
- Can you give a counterexample?
- What did you learn from the error?
- Can you do the same skill in a different context?
- What support can we remove now?
A good tutor asks fewer, better questions rather than filling every silence.
A Full Tutor Build: Secondary Mathematics
Goal: improve independent Secondary Mathematics performance over eight weeks.
Source frame: school syllabus, current topics, recent assessments and marked working.
Baseline: one mixed diagnostic identifies algebra expansion and method selection as the highest-leverage gaps.
Tutor policy: no full solutions before an attempt; use a six-level hint ladder; classify the first unstable step; retest with changed questions.
Weekly loop: repair one algebra corridor, run mixed recognition questions, then retest under time.
Mastery threshold: three changed algebra questions, one mixed application, one delayed retest and learner explanation of method choice.
Escalation: if the same prerequisite fails after repeated targeted repair, involve the human tutor or teacher to observe the learner directly.
The SI tutor does not replace the education system. It creates a faster feedback layer around the learner’s own work.
A Full Tutor Build: Writing
Goal: write clear argumentative paragraphs with evidence and controlled vocabulary.
Baseline: learner produces one paragraph without SI.
Tutor diagnoses claim, evidence, reasoning, structure and language separately.
Hint policy: ask questions before rewriting. “What is your claim?” “Which sentence contains evidence?” “How does the evidence support the claim?”
Practice: revise the original paragraph, then write a new paragraph on a different topic using the same structure.
Mastery: learner can plan, write and self-review without SI, then compare their review with the tutor’s feedback.
The tutor improves writing by strengthening the learner’s editing judgement, not by replacing the draft.
A Full Tutor Build: Professional Learning
Goal: learn to analyse operational bottlenecks.
Baseline: professional maps one real workflow.
Tutor asks for outcome, sequence, dependencies, delays, error points and handoffs.
Feedback focuses on missing structure rather than supplying a finished analysis.
Transfer: professional analyses a different workflow with no template.
Mastery: the professional can identify the bottleneck, justify it with evidence and propose a testable intervention.
Tutor Quality Control
A personal tutor should be audited like any other workflow.
Check whether explanations are accurate, difficulty adapts, hints are proportionate, mastery states reflect evidence, source alignment is current and the learner’s independent performance improves.
Track near misses: incorrect explanation caught by the learner, generated practice with an ambiguous answer, a skill marked stable too early.
The tutor improves when these failure modes become design rules.
The Tutor Dependency Audit
Ask once a month:
- Which skills now work without SI?
- Which learner actions still depend on prompts?
- Is the tutor being asked before the learner attempts?
- Are explanations replacing source reading?
- Can the learner identify errors independently?
- What support should be removed next?
The best tutor becomes smaller as the learner becomes stronger.
The Personal Tutor Exit Rule
For each skill, the tutor should know when to step back.
When independent performance, transfer and delayed retention are stable, move the skill to periodic review rather than active tutoring.
A tutor that never reduces support is measuring engagement, not independence.
The Tutor Curriculum Map
A tutor needs more than a list of chapters. It needs a map of how skills depend on one another.
For each subject, the tutor should know: foundational concepts, prerequisite procedures, common misconception corridors, application tasks and the final independent outputs expected.
Ask SI to build the map from the learner’s actual syllabus or trusted source rather than from a generic subject outline when local alignment matters.
The map should remain editable. A learner may demonstrate that one assumed prerequisite is already stable, allowing the tutor to skip it. Another supposedly simple skill may turn out to be the actual bottleneck.
The Current Corridor Rule
At any given time, keep only a small number of active learning corridors.
A corridor is one connected sequence of skills currently being repaired or built.
Example: negative signs → bracket expansion → equation solving.
The tutor can know the whole curriculum while focusing the session on one or two corridors.
This prevents a learner from receiving ten simultaneous improvement targets after one diagnostic session.
The Misconception Library
A personal tutor becomes more useful when it tracks recurring misconception patterns without turning them into permanent labels.
Example entries:
Fraction size: assumes larger denominator means larger fraction.
Algebra signs: distributes coefficient but not sign.
Writing: evidence listed without explaining relevance.
Vocabulary: confuses two near-synonyms in register.
Coding: confuses output printed to screen with a returned value.
For each misconception, store one contrast test and one transfer task.
If the learner later performs accurately across changed contexts, mark the misconception resolved rather than carrying it indefinitely.
The Recurrence Rule
One error may be a slip. Repeated errors with the same structure deserve a misconception hypothesis.
The tutor should not infer deep misunderstanding from a single typo or arithmetic mistake.
Ask for recurrence or a targeted diagnostic before changing the learner model.
The Tutor’s Instructional Priorities
When several gaps are present, prioritise in this order where appropriate:
- prerequisites blocking several downstream skills;
- errors causing high mark or performance loss;
- skills required soon by the learner’s real environment;
- gaps that can be repaired quickly and release confidence;
- extension work only after the core corridor is stable.
The tutor can suggest priorities, but a human teacher, curriculum deadline or learner need may legitimately override the order.
The Time-to-Need Rule
Learning priority depends partly on when the skill will be needed.
A prerequisite required tomorrow may deserve immediate triage. A broad long-term skill can be built more gradually.
Ask the tutor to distinguish urgent repair, near-term build and long-term development.
This prevents the tutor from teaching the most intellectually interesting topic while the learner’s immediate pipeline is failing elsewhere.
Exam Mode
Exam mode should approximate the real assessment conditions closely enough to produce useful evidence.
Remove hints. Avoid method labels. Use the relevant time limit. Do not correct during the attempt. Mark only after completion.
Afterwards, switch back into tutor mode and classify errors.
The purpose is not to make every session stressful. It is to obtain periodic evidence that the learner can perform without the tutoring layer.
Mock-Exam Debrief
A useful debrief separates knowledge, recognition, execution and time.
Question missed because the concept was unknown? Repair concept.
Method known after seeing the solution but not selected independently? Repair recognition.
Method selected but careless error occurred? Repair execution and checking.
Correct when untimed but incomplete under exam conditions? Repair speed, sequencing or time allocation.
The tutor should use the mock to change the learning plan rather than merely produce a score.
The Tutor Tone Rule
Tone should support learning without replacing feedback precision.
Encouragement is useful when it reinforces process: “Your method choice improved; the remaining error is in sign control.”
Generic praise can hide the learning state: “Great job!” is less useful if the learner still needed three hints.
Likewise, harsh correction is unnecessary. The tutor can be direct without being discouraging.
The strongest tone is calm, specific and evidence-based.
Motivation Should Not Become Diagnosis
If a learner avoids a task, the tutor should not immediately label them unmotivated.
Check task difficulty, prerequisite gaps, fatigue, unclear instructions, excessive length and repeated failure first.
Motivation may matter, but it should not become a catch-all explanation for poor performance.
Accessibility and Alternative Representations
A tutor can explain the same concept through different representations: text, table, diagram description, analogy, worked steps, example, contrast or verbal explanation.
Changing representation can reveal whether the difficulty lies in the concept or the presentation format.
Ask the tutor to preserve the underlying structure while changing the representation.
Do not assume one preferred representation is universally superior. The final skill may still require the learner to operate in the format used by school, work or assessment.
The Representation Transfer Test
If a learner understands a diagram, can they explain the same idea in words?
If they understand a verbal rule, can they apply it to a symbolic problem?
If they can follow code, can they write the equivalent logic?
Representation transfer is a useful sign of deeper understanding.
The Tutor’s Source-of-Truth Hierarchy
When several sources exist, define a hierarchy.
Example for school: teacher instructions → official syllabus → school notes or approved textbook → general supplementary explanation.
Example for software: current official documentation → current verified examples → secondary tutorial.
Example for workplace learning: authorised internal policy → current official standard → public guidance → general explanation.
The hierarchy prevents the tutor from choosing the most convenient source rather than the most relevant one.
When the Tutor Should Refuse to Guess
The tutor should stop and request a source when local rules, current requirements or high-consequence claims are uncertain.
“I need the current marking rubric.”
“I cannot verify that this software version supports the feature from the material provided.”
“This question depends on a current regulation or professional judgement.”
Abstention can be good tutoring because it models epistemic discipline.
The Tutor Lesson Planner
A compact lesson plan can use six fields.
Target: one independent output.
Baseline: evidence from the last session.
Repair: one active gap.
Practice: one sequence from familiar to varied.
Transfer: one changed task.
Review: one delayed check.
Do not let the tutor generate an hour of activities when twenty minutes of precise repair is enough.
The Tutor Homework Rule
Homework should target the active learning state.
If the learner needs retrieval, assign retrieval.
If the learner needs method selection, assign mixed classification.
If the learner needs transfer, assign changed application.
Generic volume is less useful than aligned practice.
SI makes it easy to generate large homework sets. The tutor should resist that temptation.
A Full Week with a Personal SI Tutor
Monday: baseline and diagnostic. Identify one active corridor.
Tuesday: short retrieval plus targeted explanation and three varied tasks.
Wednesday: no tutor or minimal tutor. Independent practice.
Thursday: mixed transfer task and error classification.
Friday: light review or no study depending on the learner’s schedule.
Weekend: delayed assessment, update state, set next corridor.
The schedule is illustrative, not universal. The important design is alternating support, independence and delayed evidence.
The 90-Day Tutor Review
After three months, review the tutor itself.
Which rules improved learning? Which prompts created unnecessary friction? Which subjects need different tutor profiles? Which progress fields were actually useful? Did hint dependence fall? Did transfer improve? Did source errors occur?
Simplify the tutor after evidence accumulates.
A mature tutor usually has fewer, clearer rules than an experimental tutor.
The Tutor Red-Flag List
- The tutor answers before the learner attempts.
- The tutor keeps giving harder questions without diagnosing.
- The tutor invents progress from conversational fluency.
- The tutor teaches outside the source frame without saying so.
- The learner cannot explain what they learned.
- The learner’s hint use never falls.
- The same error recurs without changing practice.
- The system stores more personal information than the learning task requires.
- The tutor becomes the only way the learner can start work.
A red flag should trigger a tutor-design repair, not simply more tutoring.
The Tutor Robustness Test
Test whether the tutor still works when the subject, learner or task changes.
The tutoring protocol should transfer: diagnose, scaffold, retrieve, apply, transfer, review.
What changes is the source frame, error taxonomy and independent output.
If the tutor requires an entirely new philosophy for every subject, the core architecture is probably too content-specific.
The Tutor Ownership Rule
The learner should be able to state what the tutor is doing and why.
“It is testing whether I can choose the method without help.”
“It is giving me one cue because I forgot the definition.”
“It is changing the context to test transfer.”
When the learner understands the tutoring process, they begin to internalise the process and become more capable of self-tutoring.
The final purpose of a personal SI tutor is to teach the learner not only the subject, but how to detect and repair their own learning state.
The Tutor Governance Card
For recurring use, keep one compact governance card: source frame, allowed help level, assessment rules, privacy boundary, progress fields, human escalation route and conditions for no-SI testing.
This card prevents the tutoring system from drifting toward whatever is easiest for the model in the moment.
The Learner Escalation Rule
Escalate to a teacher, tutor, parent or appropriate expert when the same misconception persists after several targeted repairs, when the source itself is unclear, when the learner’s performance changes sharply without an obvious instructional explanation, or when the task requires judgement the SI tutor should not own.
Escalation is not failure. It is part of a tutor that knows its limits.
The Self-Tutoring Handoff
As the learner improves, ask them to perform the tutor’s functions themselves: diagnose the first unstable step, choose the smallest hint, design a changed test and decide when to revisit the skill.
SI can then critique the learner’s self-diagnosis rather than leading it.
This is the final stage of tutor maturity: the system teaches the learner to become their own first-line tutor.
The Final Personal Tutor Test
Ask one question: “Is this tutor increasing the learner’s capacity to begin, reason, check, repair and continue without it?”
If yes, the tutor is building capability. If no, redesign the support before adding more sessions.
The Tutor Continuity Audit
At the end of each month, verify that the tutor’s current state still matches the learner. Remove resolved gaps, update drifted skills, retire obsolete source material and check whether old scaffolds are still being offered out of habit.
A tutoring system can become stale even when its prompts remain technically functional. The learner changes, the syllabus moves and the target output becomes more demanding.
Ask SI to compare the current learning card with the previous month and show only material changes. The audit should make the tutor simpler where possible.
The Tutor Success Standard
Success is not measured by conversation length, number of generated exercises or how human-like the tutor sounds.
Success means the learner reaches the intended level of independent performance, retains it, can transfer it and needs less support on the same skill over time.
A personal Super Intelligence tutor earns its place by increasing learner independence, not by increasing tutor usage.
Motivation Is Not a Substitute for Diagnosis
A tutor can encourage effort, but it should not label every repeated failure as a motivation problem.
If a learner avoids a task, ask whether the task is too difficult, too easy, unclear, repetitive, poorly timed, emotionally loaded or disconnected from a meaningful goal.
“Try harder” is a poor response when the learner lacks a prerequisite. “You are not motivated” is a poor diagnosis when the learner is exhausted or confused.
SI can generate possible causes, but it should test the learning mechanics before inventing psychological explanations.
The Motivation Support Rule
Use motivation support to clarify progress, reduce unnecessary friction and reconnect effort to a visible goal.
Useful tutor behaviours include showing what improved, shrinking the next task, celebrating evidence of independence, varying repetitive practice and making the mastery threshold visible.
Avoid exaggerated praise disconnected from performance. The learner should trust the tutor’s feedback because it is specific and evidence-based.
The Emotional-Load Boundary
Learning difficulty can carry frustration, embarrassment, fear of failure or conflict with adults.
SI can help the learner slow down, separate the task from identity and identify the next manageable action.
It should not diagnose mental health or replace appropriate human support.
A good tutor can say: “This error tells us which skill to repair next. It does not tell us what kind of learner you are.”
When emotional distress is persistent, severe or outside the learning task, involve an appropriate adult, teacher, counsellor or professional rather than extending the tutoring script.
Identity-Safe Tutoring
Avoid fixed labels such as “weak student”, “bad at Maths”, “not a language person” or “not technical”.
Translate them into observable states: “currently unstable on bracket expansion”, “recognises vocabulary but retrieves slowly”, “can read code but has not yet built functions independently”.
Observable states are repairable. Identity labels are not useful teaching instructions.
Exam and Assessment Mode
A personal tutor should have a mode where assistance disappears.
Assessment mode simulates the conditions under which the learner must eventually perform: no hints, no answer checking during the attempt, realistic timing and mixed question types.
After the attempt, the tutor returns to reviewer mode and classifies the errors.
Do not confuse guided practice scores with examination readiness. The tutor should keep those metrics separate.
The Assessment Integrity Rule
If a school or examination task is intended to measure the student’s own performance, SI should not supply hidden assistance that invalidates the assessment.
Use SI before the assessment for practice and after the assessment for review. During the assessment, follow the applicable rules.
The tutoring system should strengthen authentic performance rather than merely produce better-looking submitted work.
The Source Verification Layer
A tutor can teach confidently and still be wrong.
For important content, define the correctness source: school syllabus, textbook, teacher material, official documentation, standard, reference work or other trusted source.
The tutor should be able to say, “This explanation is based on the current source frame” or “This detail is uncertain and should be checked.”
For current technical subjects, refresh sources as the domain changes. A coding tutor that relies on old library behaviour can teach a learner into error.
The Tutor Fact-Check Routine
Before using generated practice or explanations for important learning, check:
- Is the answer unambiguous?
- Does the explanation match the source?
- Is the difficulty appropriate?
- Does the task genuinely test the intended skill?
- Are there hidden assumptions or multiple valid interpretations?
- Has the relevant rule or software changed?
Quality control is part of tutoring, not an optional extra.
The Parent–Tutor–SI Triangle
For children, a personal SI tutor may sit inside a wider support system.
The learner supplies attempts and experience. The human tutor or teacher supplies observation, curriculum judgement and relationship. The parent supports environment, schedule and oversight. SI accelerates practice, explanation and pattern review.
These roles should complement one another rather than compete.
SI should not report invented conclusions about the child to the parent. It should report observable learning evidence: “Three changed algebra questions still require Hint Level 2”, not “the child lacks discipline”.
A Parent Review Card
Current target: skill being developed.
Evidence: recent independent performance.
Support level: how much help is required.
Recurring error: first unstable step.
Progress: what improved.
Next test: how mastery will be checked.
Concern: anything requiring a teacher or tutor conversation.
This keeps the parent informed without turning the tutoring system into surveillance.
The Tutor and Curiosity
A strong tutor should make room for learner questions that are not on the planned path.
Curiosity can create powerful memory and conceptual links. But unlimited detours can also dissolve the lesson.
Use a curiosity queue. Capture the question, answer it briefly if it supports the current concept, or save it for a later exploration session.
The tutor should protect both structure and curiosity.
The Exploration Session
Occasionally switch from mastery mode to exploration mode.
Ask broader “why”, “what if” and connection questions. Build concept maps. Compare related ideas. Explore applications.
Then return to a production task so exploration leaves behind usable understanding.
The Tutor’s Learning Analytics
Keep analytics minimal and decision-relevant.
Useful metrics include independent accuracy, hint level, delayed retention, transfer success, error recurrence and time to stable mastery.
Do not track dozens of numbers simply because SI can.
The analytics should answer: what should we teach next, what can we stop teaching and what support can we remove?
The Tutor Drift Audit
Over time, the tutor can drift from its charter.
It may start answering too quickly, generating too much material, repeating familiar formats or marking mastery too generously.
Once a month, compare actual behaviour with the tutor charter.
Repair the tutoring policy, not only individual prompts.
The Tutor Handoff to Independent Study
The final stage of tutoring is not another tutoring mode. It is learner independence.
Move stable skills into self-study. The learner chooses practice, checks work, identifies errors and decides when to ask for help.
SI becomes a reference or reviewer rather than a constant companion.
This handoff should be celebrated as successful tutoring, not a loss of engagement.
The Final Personal Tutor Rule
Build the tutor around a disappearing support curve.
Early learning may need explanation, structure and frequent feedback. Growing competence should lead to fewer hints, harder transfer and longer independent periods. Stable mastery should lead to occasional review only.
A personal SI tutor is well designed when the learner increasingly carries the reasoning, the checking and the next-step decisions themselves.
The Tutor Minimum-State Rule
Keep only the learner state required to select the next useful teaching action. A target, active gap, evidence, help level, source and review condition are usually more valuable than a long conversational archive.
Minimal state makes the tutor easier to correct. If the learner improves, one card changes. If the syllabus changes, one source frame changes. If a misconception is repaired, it leaves the active queue.
This also reduces the risk that old descriptions become permanent learner identity. A tutor should know the current learning state, not construct a biography around every past error.
The smaller the accurate state, the easier it is to keep the tutoring system aligned with reality.
Frequently Asked Questions
Can I really build a personal AI tutor?
Yes. The essential step is defining tutoring rules, sources, diagnostics, help limits, transfer tests and progress records rather than relying on conversational style alone.
Should the tutor remember everything about me?
No. Keep task-relevant learning state such as target, gaps, help level and latest evidence. Avoid unnecessary personal data.
Should a tutor give answers?
Sometimes, after attempts and scaffolding when a worked repair is genuinely needed. It should then require the learner to reproduce or transfer the skill.
How do I know the tutor is working?
Independent performance improves, hint use falls, transfer strengthens and errors become less frequent or more precise.
Can parents use this with children?
Yes, with age-appropriate oversight, privacy boundaries, source alignment and limits on direct answer generation.
Can it replace a human tutor?
It can handle many explanations, practice and feedback tasks. Human teachers and tutors add observation, judgement, relationship, curriculum context and responsibility that may remain important depending on the learner and task.
What if the tutor is wrong?
Check important material against trusted sources, teachers, textbooks or current documentation. The tutor should label uncertainty rather than bluff.
What comes next?
The next article shows how to ask SI to explain difficult ideas at multiple levels without losing precision.
Helpful Reading
- How to Learn Anything Faster with Super Intelligence
- How to Give Super Intelligence Better Context
- How to Create Reusable Super Intelligence Prompts for Your Life
- How to Leverage Your Life with Super Intelligence
Build a Tutor That Gives Learning Back
A personal SI tutor should be highly available without becoming highly possessive of the task.
It should diagnose quickly, explain precisely, generate useful practice, fade help and hand the cognition back to the learner.
The best personal tutor is not the one that can answer everything. It is the one that helps the learner need fewer answers over time.