Textbook Learning (书本知识吸收)

When the user feeds you a textbook, book, or study material, absorb its knowledge thoroughly and make it available for downstream tasks in the current session.

When to Use

• User says "投喂" (feed), "学习这本书", "吸收知识", or sends a PDF/EPUB for systematic study
• User sends course materials and expects you to understand them for future questions
• User asks you to "read and absorb" a textbook before working on related problems

Workflow

Step 1: Deep Read

Use pdf-deep-reading skill for the extraction pass. Get the full content — every chapter, every section. Don't skim.

Special case — large EPUBs with many short chapters (e.g., video-tutorial-style books): When an EPUB has 200+ chapters (each covering one video/section), bulk-extract first:
1. Use ebooklib + BeautifulSoup to dump all chapter text to a JSON file (/tmp/ctf_full/chapters.json)
2. Build a part/chapter index to understand the book structure (chapter titles → major parts)
3. Do NOT try to read every chapter individually — group by part, then read the key technical chapters (the ones with substantive content, not just section separators)
4. Chapters with auto-generated titles like chapter3 (<h1> missing) have their title embedded in the first line of the text body (e.g., "1.1 Burp Suite..."). Extract it from there.

Detection: an EPUB with >50 document items where many have identical <title> content is likely a video-tutorial book. Read by part, not by individual short chapter.

Step 2: Structured Knowledge Extraction

Extract and organize knowledge into these categories. Output them to the conversation so both you and the user can refer to them:

A. 核心概念 (Core Concepts)
- Key definitions, terminology, and their precise meanings
- Include page numbers for reference

B. 公式与定理 (Formulas & Theorems)
- Every important formula with variable definitions
- Theorem statements and conditions
- Number them for easy reference: F1, F2, T1, T2...

C. 算法与流程 (Algorithms & Procedures)
- Step-by-step algorithms, encryption/decryption flows, computational procedures
- Pseudo-code or numbered steps
- Label them: A1, A2...

D. 解题模式 (Problem-Solving Patterns)
- Common question types and their solution templates
- Typical pitfalls and edge cases
- How to recognize which approach to use

E. 知识框架 (Knowledge Framework)
- How concepts connect: dependencies, hierarchies, contrasts
- "先学X再学Y" relationships
- Chapter-to-chapter logical flow

Step 3: Knowledge Manifesto

After extraction, produce a compact Knowledge Index (知识索引) that acts as a quick-reference table of contents for the book's knowledge. Format:

## 知识索引：《书名》

### 第一章: 标题
- 概念: C1.1 xxx (p.N), C1.2 xxx (p.M)
- 公式: F1.1 xxx, F1.2 xxx
- 算法: A1.1 xxx

### 第二章: 标题
...

This index lives in the conversation. When the user later asks a question, scan this index to locate the relevant knowledge instantly.

Step 4: Active Recall

When the user asks questions related to the absorbed material:
1. Scan the knowledge index first
2. Cite specific concepts/formulas/algorithms by their labels (e.g., "根据第三章 F3.2...")
3. Show the reasoning from the textbook's framework, not general knowledge
4. If the textbook has a specific approach that differs from the conventional one, use the textbook's approach

Knowledge Retention Strategy

Within a session, the extracted knowledge stays in context. For cross-session tasks:
- The knowledge index itself is compact enough to reference via session_search
- Key frameworks that will be reused (e.g., "密码学中AES的完整加密流程") should be saved as skills when they stabilize
- For exam prep, save study plans via study-plan-creation skill

Pitfalls

1. Don't assume standard conventions — If the textbook uses non-standard notation (e.g., MixColum instead of MixColumns), use the textbook's version. The exam will too.

2. Chinese math notation quirks — Chinese textbooks often use different symbols than English ones. Respect the textbook's notation.

3. Scanned PDFs — vision-based extraction workflow — Many of the user's textbooks are scanned (no text layer). Detection: page.get_text() returns empty string, or len(page.get_text('blocks')) == 0. Workflow for scanned PDFs:
   - (a) Use PyMuPDF doc.get_toc() to extract embedded TOC metadata (this works even on scanned pages).
   - (b) Extract key pages as high-res PNG images: page.get_pixmap(dpi=300) then pix.save(). Key pages = chapter start pages from TOC, TOC pages themselves.
   - (c) Use vision_analyze on each saved PNG to OCR the text content.
   - (d) Assemble structured notes from the OCR results + known book structure.
   - Do NOT try pdftotext, marker-pdf, or pdf-deep-reading first on scanned PDFs — they will all fail silently. Check for text layer first (one page is enough).

4. Scanned PPTX files — python-pptx + vision workflow — The user's 密码学及其应用 course materials arrive as .pptx files that are essentially scanned images with no editable text. Detection: page.get_text() returns empty on every slide. Workflow:
   - (a) Install python-pptx (pip install python-pptx).
   - (b) Loop through each slide: extract text with shape.text (skip empty ones), collect meaningful content.
   - (c) If all text is empty (fully scanned): export key slides as high-res images with slide.get_pixmap(dpi=300) → PyMuPDF can't render PPTX directly, so use pptx.Presentation() + shapes.z_order approach, or export via LibreOffice if available. Alternatively: convert PPTX to PDF first with LibreOffice (soffice --headless --convert-to pdf file.pptx), then use PyMuPDF on the resulting PDF.
   - (d) Use vision_analyze on each saved image to OCR the content.
   - (e) For large PPTX files (100+ slides), save images to pages/ subdirectory to avoid memory issues.
   - This session: Successfully extracted 8 chapters of 密码学及其应用 PPT (454 total pages across all files) using this approach.

5. Don't skip "obvious" chapters — Even introductory chapters contain definitions the exam will test. Extract everything.

6. Stay in the textbook's mental model — When answering questions, think like the textbook author, not like a generic AI. The grading rubric follows the textbook's framework.

7. Output file format — User prefers .json for structured data files (question banks, etc.). Do not use .hjson unless explicitly asked. When delivering structured data as a file, use .json extension.

8. Video-tutorial chapters may lack h1/h2 — Chinese CTF/practical books rewritten from video series often have 200+ chapter files with auto-generated <title> tags. The actual section number and title (e.g., "1.1 Burp Suite") appears in the body text as the first meaningful line. When BeautifulSoup fails to find h1/h2, extract the first non-empty text line that looks like a section header (starts with a digit followed by a period).

9. Chunk EPUB by part before question generation — For multi-part books (6 parts × 50 chapters each), generate questions part-by-part using delegate_task in parallel, then merge. Each subagent gets one part's text file as context, avoiding context window overflow.

Step 5: Question Bank Generation (题库生成)

When the user asks to generate a question bank from absorbed knowledge:

1. Parse format requirements — The user has a standardized JSON schema for questions. Read back the format spec from the conversation if provided. Key fields: id, content, options, answer, explanation, type, difficulty, source, nodes, formulas, siblings.

2. Type mapping — true_false (options=["对","错"]), single_choice (answer like "A"), multi_choice (answer like "BCD", letters ordered alphabetically).

3. ID convention — Use a subject abbreviation (e.g., llmpe for this book), chapter-based nodes (ch0 for preface/intro, ch1-chN for chapters, tf for standalone true/false). Sequential numbering zero-padded to 3 digits.

4. Difficulty calibration — Basic concepts → 2, understanding/application → 3, comprehensive analysis → 4.

5. Coverage — Generate questions evenly across all chapters. Include both single-choice and true/false types. Each question must have an explanation citing specific book content.

6. Output format — Save as .json (not .hjson unless explicitly asked). User prefers standard JSON. Output the JSON inline without markdown code block wrapping unless the user asks for a file.

Integration with Other Skills

Handling Practice Exam Books

When the user feeds a practice exam book (e.g., CISA Practice Exams, CISSP questions), the "knowledge" is structured as Q&A pairs, not conceptual chapters. In this case:

1. Still do a deep read to understand the domain structure and question distribution
2. Produce a compact domain overview with key concepts (the A/B/D/E categories still apply in condensed form)
3. Skip the full knowledge index — instead chain directly to question-bank-formatter for structured JSON output
4. Note the domain-to-chapter mapping (exam books often reorder chapters vs. domain numbers)
   | Feed PDF → deep analysis | pdf-deep-reading → textbook-learning |

Example

User: "投喂《现代密码学》杨波第4版"

Response flow:
1. Load pdf-deep-reading, extract all content
2. Output structured knowledge (概念/公式/算法/解题模式/框架) per chapter
3. Output compact knowledge index
4. "知识已吸收，可以开始提问或做实验"