Chiron — Gerador Universal de Aulas com IA
gyasis/chiron
Gerador LLM universal de aulas interativas para aprendizado solo
Ver no GitHubSkill Claude Code que gera aulas interativas auto-contidas em HTML a partir de qualquer fonte (repositório, PDF, CSV, paper). Suporta especialização em três domínios: código, medicina e linguagem. Utiliza arquitetura Q8 onde o agente Claude Code pai executa todas as etapas de LLM, com validação por Zod schemas e orquestração de tentativas.
Instalação
npx convez-ai gyasis/chiron1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Criar aulas interativas a partir de repositórios de código, PDFs ou papers
- Gerar conteúdo pedagogicamente especializado em code, medicina ou linguagem
- Produzir lições HTML self-contained com quiz, spaced repetition e logging de fraquezas
Veredito

Chiron é um projeto ambicioso e bem-arquitetado para geração de aulas multi-domínio com IA, usando padrões robustos (pipeline 5-stage, validação Zod, SQLite per-lesson). Requer familiaridade com Claude Code e design pedagógico; documentação detalhada suporta extensão e customização por domínio.
README
A tradução para o português deste README ainda está na fila.
Chiron — Universal Lesson Generator
"A centaur. A teacher. The tutor of Achilles, Asclepius, Jason, and Heracles."
TL;DR
Chiron is a universal LLM-powered lesson generator for solo learners across code, medicine, and language. Drop in any source — repo, PDF, vocab CSV, research paper — and get a self-contained interactive lesson HTML.
Fresh clone? Read SETUP.md first. It covers prerequisites, install, build, and verification — including how a brand-new Claude Code session bootstraps the skill.
Foundational facts
- Three co-equal domains: code, medicine, and language. Not a code-tutorial tool with the others bolted on. Specialized paths per domain are encouraged where pedagogy demands it.
- Single learner = Gyasi. No multi-user features, no auth. SQLite per-lesson at
<lesson-output-dir>/.chiron-state.db. - AI multi-persona content (peer learners + tutor + native speaker) inside one solo-learner's lesson — content-layer feature, NOT a multi-user system.
Documentation hierarchy
| Doc | Purpose |
|---|---|
| skill/README.md | Skill how-to — quick-start, supported domains, drop-in extension |
| skill/SKILL.md | Skill descriptor — top-level skill manifest |
| prd/chiron_design_v1_2026-04-28.md | Comprehensive design PRD (post-debate) |
| prd/universal_lesson_generator_2026-04-28.md | Original design PRD |
| specs/001-chiron-v1/ | Spec + design artifacts: spec.md, plan.md, data-model.md, tasks.md, contracts/ |
| memory-bank/ | projectbrief.md, systemPatterns.md, techContext.md, activeContext.md |
| CLAUDE.md | Project directives — Claude Code agent guardrails |
Architecture summary
Pipeline. Five-stage Ingest → Generate → Validate → Build → Assemble flow. Each stage has typed inputs/outputs, validators, and retry semantics. See specs/001-chiron-v1/contracts/pipeline-stages.md.
LLM architecture. Per Q8 of the design PRD, the parent Claude Code agent runs all text-LLM steps in its own context — Chiron's skill structures prompts, validates outputs against Zod schemas, and orchestrates retries. There is no separate text-LLM gateway process. Vision (diagram OCR, figure description) routes through mcp__gemini-mcp__interpret_image.
Output. A single self-contained lesson.html per lesson — vendored libs (MathJax + mhchem, Mermaid, optional Pyodide, ChalkAI for reactive-math, an abstract MoleculeRenderer) are inlined per FR-037, so opening the HTML in a browser works with no build step. Runtime state — quiz attempts, mastery, SR scheduling, weakness log, bookmarks — lives in .chiron-state.db (SQLite, per-lesson, single learner).
Heritage repos
Audited locally in ~/dev/audits/:
~/dev/audits/codebase-to-course/— HTML rendering shell (reused ~85% as-is)~/dev/audits/classbuild/— 5-stage pipeline + typed-schema pedagogy~/dev/audits/ai-course-generator/— concept DAG validator + LLM-as-advisor
Per-domain assessment formats
| Domain | Primary | Secondary | SR card type |
|---|---|---|---|
| Code | MCQ + true/false + spot-the-bug | drag-and-drop matching | Concept flashcards |
| Medicine | MCQ paired with USMLE/AMBOSS-style clinical vignette | agreement-matrix (always/sometimes/never) | Disease/drug/mechanism cards |
| Language (DE/IT) | Fill-blank + matching | sentence reordering, dictation | Vocabulary + sentence cloze cards |
Mode A vs Mode B
| Mode | Trigger | Output |
|---|---|---|
| A — Course-style | "turn this textbook chapter into a lesson" | Multi-chapter scroll-snap HTML site, Coursera-style |
| B — Case-study-style | "make this incident a teaching moment" | 3-act lecture (Evidence → 2 Lectures → Synthesis) — see ~/.claude/skills/case-study.md |
Open PRDs affecting Chiron
~/dev/prd/scratch/chiron_tts_provider_selection_2026-04-29.md— TTS provider selection (Gemini API primary, ElevenLabs fallback) for the Italian native-speaker persona. Lives in the global PRD librarian, not the project tree.
License
TBD — likely MIT. No LICENSE file committed yet.
Status / next steps
WIP — see specs/001-chiron-v1/tasks.md for current progress. Wave 12 in flight.
Why "Chiron"?
Chiron was unique among centaurs: civilized, learned, and the prototype tutor. He taught medicine to Asclepius, strategy and music to Achilles, navigation to Jason, and astronomy to Heracles. One teacher, many subjects, deep mastery.