Compile a book into a reusable Claude Code Skill — a pipeline that distills a book's judgment into a capability Claude can call. 把「一本书」编译成「一个 Claude Code Skill」的流水线——把书的判断力提炼成 Claude 可随时调用的能力。
EN — book-to-skill takes one book (PDF, scanned/OCR PDF, EPUB, Markdown, or TXT) and compiles it into a single, reusable Claude Code Skill: a SKILL.md plus a six-file references/ knowledge set plus helper scripts/. The output is not a stored copy of the book — it is a capability Claude Code can invoke on demand.
中文 — 「书道」（book-to-skill）是一条把一本书（PDF / 扫描版 PDF / EPUB / Markdown / TXT）编译成一个可长期复用的 Claude Code Skill 的流水线。产物是 SKILL.md + references/ 六件套 + scripts/。它不是把书"存起来"，而是把书做成 Claude Code 可随时调用的能力。
The value of a book = its ability to change how the reader judges and acts.
So a book-skill is not graded on "how much it memorized." It is graded on "can it make a decision in a new situation that is faithful to the spirit of the original book?" Judgment, not recall, is the acceptance criterion.
一本书的价值 = 它改变读者「如何判断与行动」的能力。 因此不以"记住了多少"为验收，以"能否在新场景里做出符合原书精神的判断"为验收。
Stage 0 Preprocess decision 判格式 + 冲突检测（先判断，再动手）
Stage 1 Structuring → book_structured.md（章节边界 + 全角→半角）
Stage 2 Deep parse 多 Agent 并行四层提取：概念→框架→方法→原则
Stage 3 Cognitive architecture 五要素：身份/知识分配/能力边界/行为模式/交互协议
Stage 4 Skill authoring SKILL.md（<500行）+ references 六件套 + scripts
Stage 5 5-dimension validation 覆盖度/关系/应用/判断/边界——以"用"为验，不以"记"为验
Stage 6 Archive & close loop 知识图谱 + 工作日志 + 索引
| Stage | EN | 中文 |
|---|---|---|
| 0 | Preprocess & conflict-detect | 预处理决策 + 冲突检测 |
| 1 | Structuring (chapter boundaries, full→half-width) | 结构化 |
| 2 | Multi-Agent parallel 4-layer parse | 多 Agent 并行四层解析 |
| 3 | Cognitive architecture (5 elements) | 认知架构五要素 |
| 4 | SKILL.md + references six-pack | SKILL 编写 + references 六件套 |
| 5 | 5-dimension validation (use, not recall) | 五维验证 |
| 6 | Archive: knowledge graph + log + index | 归档闭环 |
Install / 安装
# Option A — clone directly into your skills dir
git clone https://github.com/weitzu-com/book-to-skill.git \
~/.claude/skills/book-to-skill
# Option B — clone elsewhere, then symlink (or copy)
git clone https://github.com/weitzu-com/book-to-skill.git
ln -s "$(pwd)/book-to-skill" ~/.claude/skills/book-to-skillOptional script dependencies:
pip install pymupdf # extract_pdf.py (embedded OCR-layer extraction)
brew install poppler # pdftotext, used by pdf_to_text.sh
brew install tesseract # optional: OCR fallback for scans with no text layerTrigger / 触发 — just tell Claude Code to make a skill out of a book and give it the path:
把这本书做成 skill：/path/to/book.pdf
Make this book into a skill: /path/to/book.pdf
Claude Code loads book-to-skill and runs the 7-stage pipeline end to end.
book-to-skill/
├── SKILL.md # Pipeline entry: 7 stages, decision trees, routing table
├── README.md # This file
├── LICENSE # MIT
├── references/ # On-demand knowledge (loaded by SKILL.md routing)
│ ├── 00-pipeline.md # Per-stage done-criteria + failure modes
│ ├── 01-preprocess.md # Format detection, scanned-PDF OCR, chapter boundaries
│ ├── 02-parse.md # Agent prompts, division of labor, 4-layer extraction
│ ├── 03-authoring.md # Cognitive architecture + SKILL.md + six-pack how-to
│ ├── 04-validation.md # 5-dimension criteria + anti-error checklist
│ ├── 05-workflow.md # Workflow parallelism scripts + efficiency benchmarks
│ └── 06-governance.md # Three-Departments-Six-Ministries governance
├── scripts/
│ ├── pdf_to_text.sh # Text PDF → TXT (keeps layout, full→half-width)
│ ├── extract_pdf.py # Scanned PDF → text via PyMuPDF embedded OCR layer
│ ├── detect_structure.py # TXT → detect chapter boundaries → structured skeleton
│ └── check_skill.py # 5-dimension structural validation of a produced skill
└── templates/ # Skeleton templates for the produced skill
├── SKILL.template.md # SKILL.md skeleton (placeholders + ironclad rules)
└── references/ # Six-pack reference stubs
| Script | Purpose / 用途 | Example |
|---|---|---|
pdf_to_text.sh |
Text-type PDF → TXT, preserving layout and converting full-width to half-width characters. 文本型 PDF 转 TXT。 | bash scripts/pdf_to_text.sh book.pdf |
extract_pdf.py |
Pull the embedded OCR text layer out of a scanned PDF (PyMuPDF; ~258 pages in <30s). 扫描版 PDF 提取嵌入 OCR 文字层。 | python3 scripts/extract_pdf.py book.pdf > full_text.txt |
detect_structure.py |
Detect chapter boundaries in a TXT and emit a structured skeleton. 检测章节边界 → 结构化骨架。 | python3 scripts/detect_structure.py full_text.txt |
check_skill.py |
Run 5-dimension structural validation on a produced skill (line count, concept count, isolated nodes, six-pack completeness, routing-load directives). 对产出 skill 做结构验证。 | python3 scripts/check_skill.py <skill_dir> |
Note / 注意:
check_skill.pyvalidates a skill produced by this pipeline (a book compiled into a skill), not this meta-skill repo itself. Running it againstbook-to-skill/will report the six-pack (concepts/frameworks/methods/principles/cases/knowledge_graph) as missing — this is expected, becausebook-to-skillis the compiler, not a book-derived skill.check_skill.py校验的是「书道产出的成品 skill」，对本仓库自身运行会报六件套缺失，属预期。
Dependencies / 依赖: PyMuPDF (pip install pymupdf) for extract_pdf.py; pdftotext from poppler for pdf_to_text.sh; tesseract (optional) for OCR fallback on scans without an embedded text layer.
Five real books, MD / text-PDF / scanned-PDF, 135–314 pages, full conversion in 30–85 minutes.
| Book / 书 | Format / 格式 | Pages / 页数 | Time / 用时 |
|---|---|---|---|
| 会计七原则 (7 Principles of Accounting) | Markdown | 135 | ~30 min |
| Claude Code 手册 (Claude Code Manual) | PDF (text) | 75 | ~40 min |
| Joe Girard (How to Sell Anything) | PDF (OCR) | 190 | ~60 min |
| 经营十二条 (12 Management Principles) | PDF (scanned) | 258 | ~50 min |
| Harness 工程之道 (The Harness Way) | PDF (scanned) | 314 | ~70 min (Workflow parallel) |
The pipeline governs itself with a Three Departments / Six Ministries model (中书起草 draft → 门下审驳 review & veto → 尚书六部执行 execute via six ministries): drafting proposes, the reviewing department can reject, and the six executing ministries carry out the work. Full details in references/06-governance.md.
- v2.4 — Provenance gate / 溯源门禁: every L1–L4 element must trace to a real location in the book; un-sourceable items are marked
[待核]and never fabricated, plus a random 5-element source spot-check before validation. 每个要素须溯源到书中真实位置，溯源不到标[待核]绝不编造；五维验证前先随机抽 5 要素回原书核对。 - v2.3 — 5-book validated 7-stage pipeline (MD / text-PDF / scanned-PDF, 135–314 pp). 经五书验证的七阶段流水线。
MIT © 2026 weitzu-com