teach

A Claude Code skill that teaches you a topic over multiple sessions — and can teach a whole book faithfully.

Forked from the teach skill in mattpocock/skills, extended for teaching from complete books: when you upload a textbook or nonfiction book, the skill reads the whole book, distills its reusable methods through RIA-TV++, proves coverage and method quality, and only then designs the mission-bound course.

Install

Clone this repository into your skills directory — the repo root is the skill:

git clone git@github.com:Sebastianhayashi/skills.git ~/.claude/skills/teach

What it does

The upstream skill turns a directory into a stateful teaching workspace: a MISSION.md grounding why you're learning, numbered HTML lessons, durable reference documents, and learning records used to pitch each lesson at your zone of proximal development.

This fork keeps that philosophy and makes book teardown plus teaching one continuous pipeline:

• Whole-book reading gate — before any book-derived lesson, the skill classifies the source (fiction is declined, excerpts are not treated as books), inventories the entire book into a traceable reading ledger, and must reach 100% coverage. Unreadable ranges stay blocked, never silently skipped. See BOOK-READING-FORMAT.md.
• RIA-TV++ distillation gate — after full reading and before curriculum design, five independent passes extract frameworks, principles, cases, counterexamples, and glossary terms; triple verification retains only non-trivial reusable methods; every verified method becomes an atomic R/I/A1/A2/E/B unit, is linked to neighboring units, and passes trigger, non-trigger, cross-unit, and boundary pressure tests. Rejections stay auditable and meaningful rejected content still receives a curriculum disposition. See RIA-DISTILLATION-FORMAT.md.
• Auditable course design — after the overview and verified RIA list are confirmed, the curriculum blueprint gives every meaningful source unit and every verified RIA unit an explicit teach / reference / defer destination; silent drops are forbidden. The blueprint is bound to the mission by SHA-256 and to the RIA revision, so either kind of drift locks teaching. See CURRICULUM-BLUEPRINT-FORMAT.md.
• Per-lesson publication contracts — every book-derived lesson starts as a brief pinned to stable source anchors: one capability slice, a complete original case with its reasoning chain, a transfer example, a boundary, practice with feedback, and explicit exclusions. See LESSON-BRIEF-FORMAT.md.
• A generated course home — each workspace gets a static index.html (built by scripts/build-course-index.py) linking mission, lessons, references, and reading-gate status, with a shared stylesheet. See COURSE-INDEX-FORMAT.md.
• Structural checking — scripts/check-book-course.py validates briefs and lessons against the fidelity gates before a lesson may be published.

Delta against upstream

New files: BOOK-READING-FORMAT.md, RIA-DISTILLATION-FORMAT.md, CURRICULUM-BLUEPRINT-FORMAT.md, LESSON-BRIEF-FORMAT.md, COURSE-INDEX-FORMAT.md, assets/course.css, scripts/build-course-index.py, scripts/check-book-course.py.

Modified files: SKILL.md (rewritten around the reading gate, fidelity rules, course home, and brief-gated lesson design), agents/openai.yaml (description and default prompt).

Unchanged files: MISSION-FORMAT.md, RESOURCES-FORMAT.md, LEARNING-RECORD-FORMAT.md, GLOSSARY-FORMAT.md.

License

MIT — see LICENSE. Original work by Matt Pocock, modifications by Sebastianhayashi.