The mentor is not a content database — it is a teacher. Its behaviour is grounded in established learning science, and it adapts to the learner in real time.
Every reply follows these rules (from
templates/mentor_system_prompt.md):
- Teach, don't dump — guide with a hint before giving the answer.
- One concept at a time — don't advance until the current idea is solid.
- Check understanding — retrieval questions, not "does that make sense?".
- Adapt — raise difficulty when the learner succeeds; shrink the step when they struggle.
- Concrete before abstract — example first, then the principle.
- Fade the scaffolds — do less as the learner gains competence.
- Practice over passivity — every session produces something.
- Specific feedback — name what was right, the exact gap, and the next step.
The design draws on:
- Zone of proximal development — teach just beyond what the learner can do alone, within reach with guidance.
- Cognitive load theory — one new element at a time; worked examples early; scaffolds removed as fluency grows.
- Retrieval practice & the testing effect — pulling knowledge from memory strengthens it far more than re-reading.
- Spacing and interleaving — revisit and mix, rather than block.
- Formative feedback — where am I going, how am I going, where to next.
- Metacognition — predict, self-explain, summarise.
Full notes: references/pedagogy.md.
Phase 1 Intake — a few questions to find the real starting point
Phase 2 Roadmap — a short, agreed sequence of milestones
Phase 3 Teach loop — activate → explain + example → check → exercise → feedback
Phase 4 Consolidate— learner summarises, gets one piece of independent practice
A learner profile stores what was demonstrated, what was shaky, known misconceptions, and open questions:
{
"name": "rust",
"field": "Rust",
"level": "intermediate",
"goal": "ship idiomatic Rust",
"sessions": 3,
"mastered": ["ownership"],
"shaky": ["lifetimes", "borrowing"],
"misconceptions": ["Rc and Arc are interchangeable"],
"open_questions": ["when to use Rc vs Box"]
}When history exists, it is injected into the system prompt so the mentor stops re-teaching what you know and deliberately revisits what is weak.
mentor review asks the model to assess the latest transcript and return strict
JSON: mastered / shaky / misconceptions / open questions / level / next lesson.
It merges the result into the profile, promoting anything newly mastered out of
shaky. Nothing is invented — items require evidence from the transcript.
mentor review rust # show proposed changes
mentor review rust --apply # save them
mentor review rust --dry-run # show the review prompt, call nothingmentor cards --generate turns a session into flashcards; mentor quiz reviews
due cards with an SM-2-lite scheduler. Each card tracks an interval (days),
ease factor, repetition count, and due date. Grades:
| Grade | Effect |
|---|---|
again |
relearn in ~10 minutes; ease down |
hard |
shorter next interval; small ease penalty |
good |
standard interval growth |
easy |
longer interval; ease up |
The mentor distinguishes established fact, expert consensus, opinion, and uncertainty; never invents sources; and for health, law, finance, or safety, frames guidance as educational and flags when a licensed professional is required.