Textbook Learn

textbook-learn

Turn a book / PDF into a single, offline-openable learning page that teaches it like a course: a map of the whole book and a reading path, then a few chapters broken down deeply, each ending with a worked example (例题精讲) and an active-recall quiz — with a progress tracker that remembers what you read and how you scored. Left = table of contents, right = the chapter content, with the book's own figures downloaded locally.

This is the family's third sibling, after github-project-learn and domain-learn. The output shell is the same; what differs is the source (one book, not a repo or the open web), the research (parsing the book's structure + sourcing every claim back to a page), the pedagogy (a chapter course, not a beginner→frontier roadmap), and the interactivity (self-grading quizzes + worked-example reveals + a read/score tracker, instead of a parameter demo).

When to use

The description covers triggering. In short: a book title or a PDF/EPUB file/link + any "I want to learn / work through / study this" intent. Examples: "帮我学《动手学深度学习》", "把这本 PDF 做成逐章学习页 + 测验", "make me a study guide for this textbook", "work through SICP".

Output shape

A folder (default -learn/) containing:

• index.html — self-contained (inline CSS+JS), double-click to open, fully offline
• assets/ — the book's real figures/diagrams (and cover), downloaded

Workflow (high level)

1. Identify the book & confirm scope. Pin down the exact edition; ask the user **which

chapters to go deep on** (a whole book is too much — pick 3–6 representative chapters) and the depth. Tell them you'll build a single self-contained offline page with quizzes.

1. Get the book's structure & content, sourced. Parse the real table of contents and, for

the chosen chapters, the actual core ideas, key equations/definitions, figures, and the chapter's own exercises (most textbooks end sections with them — these seed the quiz faithfully). Every load-bearing claim, formula, and figure must come from the real book, cited to a page/section. See references/source-parsing.md — this is the make-or-break.

1. Design the chapter course. Group the full TOC into 3–5 stages (打基础 → 核心 → 进阶 → 应用)

as a chapter map with a "从这里开始读" path, then deep-dive the chosen chapters.

1. Write each deep-dive chapter as: core takeaways → key equations/definitions (rendered

offline, no KaTeX CDN) → the book's figure → one worked example the learner reveals step by step → an active-recall quiz (MCQs that self-grade + free-recall flip cards), built from the chapter's real exercises and content. See references/quizzes-and-examples.md.

1. Download media with scripts/fetch-media.sh into assets/ (handles Git LFS, blocked

hosts, and SVG; drops anything that isn't a real image). Use relative assets/... paths.

1. Generate the page by copying assets/template.html (a complete worked example — the

Dive into Deep Learning page) and replacing its content, keeping the CSS, the JavaScript (quiz grading, worked-example reveal, flip cards, progress tracker with localStorage, lightbox, scroll-spy, glossary search), and the section scaffolding.

1. Verify it renders. Open the page (a headless browser if available), confirm every

figure loads, a quiz grades right/wrong, the worked example reveals, and the tracker counts. Fix before calling it done.

Read references/workflow.md for step-by-step detail, references/page-design.md for the section list + components, references/source-parsing.md before parsing the book, and references/quizzes-and-examples.md before writing the pedagogy.

Bundled resources

• assets/template.html — the proven page (a worked Dive into Deep Learning example).

Copy it, keep its CSS + JS (quiz grading, worked-example reveal, flip cards, localStorage progress tracker, lightbox, scroll-spy, searchable glossary, copy buttons), replace all content.

• scripts/fetch-media.sh — ./fetch-media.sh /assets name.ext=url …. Auto-fixes Git

LFS pointers, routes blocked hosts through a proxy, accepts SVG, and drops invalid images. Run via the Bash tool (git-bash curl), not Windows cmd curl.

Principles that make the page good

• **Faithful, never fabricated — and sourced to the book.** Every formula, definition, and

quiz answer must be checkable against an actual page/section; cite it (a "出处" link). If a worked example uses numbers the book doesn't print, say so ("演算示例，书中未印此数"). A learning page that teaches a wrong formula is worse than none.

• A course, not a summary. Organize for working through the book: a reading path, a few

chapters done deeply, recall built in — not an encyclopedia dump of every chapter.

• Quizzes are the point. Active recall + worked examples are what make this beat reading

the PDF. Build them from the book's own exercises; make them self-grade and stick (progress persists in localStorage).

• Don't boil the whole book. Deep-dive a representative few chapters; map the rest. Be

honest that the other chapters follow the same pattern.

• Offline-first. Relative assets/ paths; render math without a CDN; verify every

download is a real image before referencing it.

• Confirm scope (which chapters, what depth) before a long run.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: shangjunyang1986
• Source: shangjunyang1986/ai-learning-skills
• License: MIT

Install and usage instructions live in the source repository linked above.