ai-tutor

Ship an adaptive multimodal personal tutor for a specific subject with Bayesian knowledge tracing, a curriculum graph, safety filters, and a measured two-week efficacy study. Use when you need help wi

SKILL.md

GitHub

Given a subject (K-12 algebra or intro Python), build a personal tutor with text + voice + photo-math input, Bayesian knowledge tracing learner model, curriculum-graph-driven concept selection, COPPA-aware memory, and safety filters. Run a two-week efficacy study with 10 learners.

Build plan:

  1. Curriculum graph in Neo4j: 50-150 concept nodes with prerequisite edges and attached OER content (OpenStax, Open Textbook).
  2. Learner model: Bayesian knowledge tracing with priors for guess/slip/learn-rate per concept; per-learner persisted state.
  3. Tutor policy (LangGraph over Claude Sonnet 4.7 with prompt caching): read_signal -> select_concept (graph walk) -> scaffold (Socratic) -> update_mastery.
  4. Memory: agentmemory-style persistent episodic + semantic store; COPPA-aware auto-delete after 1 year; parent-accessible deletion.
  5. Voice: LiveKit Agents worker with Whisper-v3-turbo ASR and Cartesia Sonic-2 TTS; reuse capstone 03 pipeline.
  6. Photo math: dots.ocr or PaliGemma 2 for equation recognition; feed structured input to the tutor.
  7. Safety: Llama Guard 4 input/output; age-appropriate filter blocking self-harm/adult/violence; learner-scoped memory isolation.
  8. Weekly PDF progress reports per learner.
  9. Efficacy study: 10 learners, pre-test (standardized 30-question baseline), 2 weeks of sessions (3/week), post-test; compare against non-adaptive linear cohort.

Assessment rubric:

Weight Criterion Measurement
25 Learning gain delta Pre/post-test delta in the 10-learner 2-week study
20 Socratic fidelity Rubric score on transcript samples
20 Multimodal UX Voice + photo + text coherence end to end
20 Safety + privacy posture Llama Guard 4 pass rate + COPPA-aware retention + cross-learner isolation
15 Curriculum breadth and graph quality Concept coverage + prerequisite graph consistency

Hard rejects:

  • Tutor policies that answer-dump instead of asking the next question. Socratic is a hard requirement.
  • Learner models that do not update per interaction. BKT is a floor.
  • Memory without COPPA-aware retention. Unacceptable for a K-12 audience.
  • Efficacy claims without a non-adaptive baseline cohort.

Refusal rules:

  • Refuse to deploy without Llama Guard 4 on both input and output.
  • Refuse to persist learner data without a parent-accessible deletion surface.
  • Refuse to claim "adaptive" without running the non-adaptive baseline alongside.

Output: a repo containing the curriculum graph, the BKT learner model, the LangGraph tutor policy, the multimodal input handlers, the LiveKit voice pipeline, the safety pipeline, the parental dashboard, the efficacy-study runner, the pre/post test harness, and a write-up documenting the learning gain delta versus the linear baseline with confidence intervals.