Skip to content

The six learning methods

Each method has a stance, a strength, a weakness, and a characteristic style for the AI to take during sessions. The 42-cell prompt matrix (6 methods × 7 steps) implements these stances per cycle step.

Deductive

Stance: Theory first. State the rule completely, then demonstrate with examples, then ask the learner to apply.

Strong when: the topic has clean, statable rules (formal grammar, mathematical proofs, type systems). The learner already accepts that rules govern the domain and wants to internalise them efficiently.

Weak when: the rules are fuzzy, contested, or context- dependent. Pure deductive teaching of "good taste" or "clarity" lands flat — the learner needs to see many examples before the implicit pattern crystallises.

AI's style: precise, structured, complete. Spells out the rule in plain language, demonstrates with prototypical worked examples, then asks the learner to solve a fresh instance.

Inductive

Stance: Examples first. Show three to four carefully chosen examples of the same phenomenon and let the learner derive the rule themselves. Reveal the rule only after the learner has formed a hypothesis.

Strong when: pattern recognition is exactly the cognitive skill the learner needs to develop. Language learning, music theory, chess tactics, machine learning intuition — all benefit from inductive practice.

Weak when: speed matters. The inductive route is slower than deductive when the rule is simple and unambiguous. "Always free your memory" doesn't need three examples; just state it.

AI's style: presents examples side-by-side, holds back from explaining, asks "what pattern do you see?" or "what's the next item in this series?".

Error-based

Stance: Provoke mistakes, then learn from them. Pose tasks specifically designed to trip the learner into the topic's classic pitfalls. Then explain why the trap is so tempting.

Strong when: the topic has well-known traps (subject- verb agreement in long sentences, off-by-one errors in loops, common fallacies in argumentation). The learner benefits from feeling the trap pull before they understand the corrective mechanism.

Weak when: the learner is fragile, anxious, or new. The "productive frustration" can tip into "I'm bad at this" without a careful framing. The AI's step 3 (Error) prompt in this method explicitly says "diagnose precisely without padding" — that's a teaching choice, not a personality defect.

AI's style: confrontational about the mistake, then deeply explanatory about its mechanism. "That's the classic trap X — you fell into it because Y. Here's why it's so tempting."

Dialogic

Stance: conversational exchange, low pressure. Frame tasks as invitations, not tests. Affirm what's right explicitly before delivering corrections. Let the learner co-steer.

Strong when: the learner has anxiety, fragile confidence, or has hit a wall. The relaxed tone restores agency. Also strong when the topic itself is conversational (rhetoric, debate, presentation skills).

Weak when: the learner wants direct instruction and is frustrated by "want to give it a try?" framing. Some learners read dialogic prompts as evasive.

AI's style: warm, curious, low-density. Asks "what led you there?" before correcting. Affirms partial correctness explicitly. Suggests tempo or focus shifts.

Contextual

Stance: real-world scenarios first. Set up a concrete situation where the topic is immediately needed; theory comes only after the learner has tried to act in the scenario.

Strong when: the topic is applied or domain-specific (business communication, clinical reasoning, engineering trade-offs). The learner needs to feel the situational pressure to understand which theoretical knob actually matters.

Weak when: the topic is genuinely abstract (set theory, formal logic, music theory in a vacuum). Forcing a scenario makes the lesson feel contrived.

AI's style: scene-setting. "You're at the door of a client meeting and they ask…". Asks for the learner's next concrete action. Shows consequences inside the scenario.

AI-adaptive

Stance: the AI picks per turn. Reads the profile and the session history; selects whichever of the other five methods fits this exchange. Justifies the choice in one sentence.

Strong when: the learner has a balanced profile (no dominant method) or is in a session where multiple methods might work. Also strong for advanced learners who can articulate when a method isn't landing.

Weak when: the learner wants a stable, predictable teaching style. The constant method-switching can feel jittery if not justified well per turn.

AI's style: meta-aware. Names the method it's choosing ("Let me try inductively..."), executes that method faithfully, and switches when the signal says it's not working.

How the app implements each

The six methods are not just labels. Each one drives a distinct AI personality via the 42-cell prompt matrix in plugins/.../session/prompts.py: one prompt per (method, step) pair, six methods × seven steps. A deductive Input prompt opens with the rule and asks for examples; a contextual Input prompt opens with a real-world scenario and asks how the learner would tackle it. Same step, completely different texture.

The matrix is exported verbatim to frontend/src/data/session-prompts.json for Dexie-mode parity — no drift possible between Server and Local modes.

Choosing among them

Your assessment gives you a 6-method profile. The dominant method is what new sessions start in. But:

  • The step evaluator (dual-prompt v0.5.0) may suggest staying, advancing, or — rarely — stepping back per cycle step.
  • The method-switch heuristic detects stagnation (three sessions of flat understanding + high stress) and surfaces a "want to try [other method]?" banner in both storage modes.
  • You can manually pick a method on the Session page's start button. Useful when you know the topic calls for one specific method.

Method-switching is the goal, not method-loyalty. A learner who has used five of the six methods across their Adaptive Learner history has a richer mental toolkit than one who's locked into deductive forever.