A Claude skill for long-running, multi-session learning — where you write the textbook.
You bring a subject (a topic list, a textbook index, or just a name). The AI completes the picture, calibrates to your level, builds a syllabus, and teaches Socratically — sketching each concept's core, then making you do the thinking through predictions, edge cases, and connections. At natural stopping points it proposes a checkpoint: it recaps, you write the entry in your own words, it gives honest feedback, you revise, you commit. Your committed text goes into the document verbatim and is never touched by the AI again — not even typos. Over weeks, the document becomes a textbook of the subject written entirely in your own voice, in the order you actually learned it.
Everything lives in one self-contained HTML file — the single source of truth and your save state:
- Syllabus — a guide, not a rail; tangents get grafted in where they happened
- Concept map — a Mermaid graph of topic names and links, grown as you learn
- Textbook — your committed entries, in learning order, with git-style commit lines; optionally illustrated with AI-made figures (inline SVG, Mermaid, or self-contained interactive widgets) in clearly labeled blocks below your text
- Session log — append-only, a few lines per session
- AI state (collapsed) — a handoff block one AI instance writes for the next: position, open threads, misconceptions to watch, and a learner profile that refines forever and never resets
- AI instructions (collapsed) — the full resume protocol, baked into the file
Because the instructions travel inside the file, the document is self-bootstrapping: upload it to any Claude conversation and say "resume my learning" — it works even without this skill installed. Share a document with a friend and it carries everything they need.
The file is also directly editable: an Edit/Save toolbar lets you edit your own entries in the browser. Save writes the file back in place on desktop Chrome/Edge (File System Access API) and downloads a fresh copy everywhere else. Saving is idempotent — Mermaid sources and widget runtime state are never baked into the file.
Concept map & textbook — a Mermaid graph that grows as you learn, above your committed entries in learning order with git-style commit lines.
Figures — optional AI-made illustrations (inline SVG, Mermaid, or interactive widgets) in clearly labeled blocks below your text.
Session log — an append-only record, a few lines per session, with the AI state and instructions tucked into collapsed blocks below.
Download the repo to get the full skill folder. Or download just the latest the SKILL.md from the Releases section.
claude.ai (Pro plan or higher): Settings → Capabilities → enable Code execution and file creation, then Customize → Skills → Upload skill → select the skill file
Claude Code: Place learning-companion/ folder from the repoo in your ~/.claude/skills/learning-companion/ (personal) or .claude/skills/learning-companion/ (per project).
- Start: "Start a learning session — here are my topics: …" (naming the skill on the first run guarantees it triggers)
- During: tangents are expected; "just tell me" is always honored; checkpoints are pauses, not tests
- After each checkpoint in claude.ai: download the updated document — the file is your save state, the chat is disposable. (In Claude Code the file is edited in place.)
- Resume: new chat, attach the document, "resume my learning". Attach the companion
<subject>-transcript.mdtoo if you want one continuous conversation log.
- The learner writes everything that gets committed. Rewriting in your own words is where the learning happens; the AI's recap is a refresher, never a model answer to transcribe.
- Committed text is immutable to the AI. Errors are corrected by new entries; the record of your understanding at each point in time is preserved. (You can edit your own entries freely — the rule binds the AI, not you.)
- One file, no sync. State, content, instructions, and editor all travel together; document and protocol can never drift apart.
- Tangents are features. They're tracked, checkpointed, and grafted into the syllabus and map rather than evaporating.
SKILL.md— the skill (setup protocol, teaching, checkpoints, persistence)assets/learning-doc-template.html— the document template (format v1.0)