Claude Skills for Intelligent Textbooks

Automating intelligent textbook generation

A ~3,500-hour authoring task, done in ~10 hours

Built for Claude — Portable Everywhere

  • Originally written for Anthropic's Claude Code skill system
  • Same SKILL.md + CLAUDE.md pattern now runs in any AI IDE that supports "skills"
  • Verified working in OpenAI Codex, Google Gemini, Cursor, Perplexity, and Hermes
  • Just copy CLAUDE.md → AGENTS.md

One Pipeline, Start to Finish

  • init-textbook scaffolds a brand-new project — mkdocs.yml, docs/ tree, starter pages
  • 12-step pipeline: description → Bloom's outcomes → 200 concepts → dependency DAG → taxonomy → learning graph → chapter structure → content → MicroSims → glossary/FAQ/quiz → QA → deploy
  • book-installer installs supporting infrastructure and runs quality checks along the way
  • Ends with mkdocs gh-deploy

Learning Graphs: the Backbone

  • A 200-concept DAG with Bloom's-Taxonomy tagging drives sequencing for the whole book
  • learning-graph-generator builds it in five steps:
    • Enumerate concepts → map dependencies (CSV DAG)
    • Validate with analyze-graph.py (no cycles, no orphan nodes)
    • Categorize into 12 taxonomy groups
    • Convert to vis-network JSON with csv-to-json.py
  • Nothing gets written in a chapter before its prerequisites exist

Token Efficiency: Scripts, Not Re-Derivation

  • Repetitive, deterministic work is pushed into bundled Python utilities, not redone in-context every time
  • Example: microsim-generator's batch utilities (parsing specs, scaffolding files, fixing iframes, updating nav) save ~430K tokens per batch run
  • Frees the model's attention for the part that actually needs judgment — writing the visualization logic itself

Mascots: Part of Book-Installer, Not a Standalone Skill

  • Lives inside book-installer as the learning-mascot capability
  • Designs a pedagogical-agent persona (name, appearance, voice, catchphrase)
  • Generates AI image prompts for consistent mascot poses
  • Implements it via custom CSS admonitions with the image floated in the body
  • Goal: engagement, wayfinding, and brand identity — not just decoration

Supplementary Content Tour

  • glossary-generator — ISO 11179-compliant definitions from the concept list (200 terms in this site)
  • faq-generator — FAQ set from content + learning graph + glossary, once ~30% of chapters exist (64 questions)
  • quiz-generator — Bloom's-aligned multiple-choice questions per chapter (520 questions)

Mini Graphic Novels for Historical Context

  • story-generator — illustrated narratives about scientists, mathematicians, and historical figures
  • Default: 12 panels + 1 cover (configurable 4–16 via --panels N)
  • Each panel: narrative paragraph + a detailed image-generation prompt
  • Can auto-generate every panel image (Gemini / gpt-image-1) at ~$0.039/image — about $0.51 for a full 13-image story

Getting Started

  1. Symlink skills into ~/.claude/skills/ via scripts/install-claude-skills.sh
  2. Run init-textbook to scaffold a new project
  3. Run course-description-analyzer, then learning-graph-generator
  4. Follow the 12-step pipeline chapter by chapter

What's Next: 10 Hours → 4 Hours

  • This is a roadmap, not a shipped feature
  • Push more repetitive work out of model context into scripts — the same shift already made for MicroSims
  • Tighten the handoffs between the 12 pipeline steps so less manual re-prompting is needed between stages

Presenter note: one caveat — skills that rely on Claude's image-understanding (e.g. reviewing a MicroSim screenshot for layout errors) are Claude-specific and don't port cleanly to IDEs without equivalent vision tooling.