A study buddy that refuses to let the model make things up

What it does

It’s an Agent Skills package (Python, SKILL.md-based) that runs inside agents like Claude Code, Codex, WorkBuddy or Copilot CLI. Feed it a PDF, EPUB, DOCX, slides or a pile of files, and it produces a .learnkb/ knowledge base, a self-contained interactive HTML learning page, and Markdown — all derived from the same extracted data. Two depths: a quick overview of the core argument, or a systematic study that reads to the end.

The interesting bit

The whole design is built around epistemic hygiene: material-backed facts, model-added content, uncovered areas and externally verified items are kept explicitly separate, and both depths run coverage audits with source hashes. When a material contains no methodology, it says so instead of inventing one — a rarer virtue than it should be.

Key highlights

• Fine-grained source mapping: PDF pages, PPT slides, EPUB chapters, with hover tooltips showing where each claim came from
• An accumulating method library (methods.json) with versioned cards, stable IDs, and no silent overwriting of what the source actually said
• Six modules per page: framework cards, guided content, glossary, action rules, self-check quizzes, local notes
• Bilingual Chinese/English throughout, following your explicit instruction rather than guessing from locale
• Offline by default; notes, profiles and wrong-answer records stay in the local browser

Caveats

• Pure chat environments without file access or Python can only use the prompting workflow — no extraction, coverage validation, or HTML rendering
• The README is dense to the point of exhaustion; expect a learning curve on the terminology (“unit rereads”, “dependency fingerprints”)

Verdict

Worth a look if you use an agentic CLI and want study material you can actually audit rather than trust. If you just want a chatbot summary of a PDF, this is considerably more machinery than you need.