CRISP = Comprehend · Review · Internalize · Synthesize · Practice
深度閱讀五階段：理解結構、批判審視、內化連結、整合全貌、付諸行動。
大部分 AI「讀書摘要」給你一份條列清單就結束了。你覺得好像讀了什麼，但什麼都沒留下。
CRISP Reading 不一樣。它不只是摘要——它像一位認真的讀者那樣閱讀：質疑作者的預設立場、壓力測試論證邏輯、把新想法跟你已知的知識連結起來，最後把洞察轉化成具體可執行的行動。
產出不是一堵文字牆。而是一份互動式 HTML 閱讀報告——可展開的論點、精選引句、批判專區、個人行動計畫——設計得像翻開一本精裝書，而不是滑過一個儀表板。
npx skills add kcchien/crisp-reading就這樣。當你請 AI 助手讀書或分析書籍時，它會自動啟用這個技能。
用自然語言跟你的 AI 助手對話即可：
| 你說 | 發生什麼事 |
|---|---|
| 「幫我讀這本書」 + PDF | 提取全文 → 深度分析 → 互動式 HTML 報告 |
| 「分析《老殘遊記》」（僅書名） | 自動從 70,000+ 冊公共領域書庫下載全文 → 同樣的深度分析 |
| 「分析《原子習慣》」（書庫無收錄） | 依 Claude 知識分析，報告中明確標示 |
| 「這本書值不值得讀？」 | 快速 TIPS 四維度評估，不囉嗦 |
| 「我有讀書筆記了，幫我深化」 | 以你的筆記為基礎延伸，不從頭開始 |
| 「先讀前五章」 | 支援大型書籍分批處理，最後合併為一份報告 |
沒有 PDF？沒問題。技能會自動搜尋 Project Gutenberg（透過 Gutendex API）並下載全文進行分析——完全不需手動操作。
- 70,000+ 冊公共領域書籍可用
- 444+ 冊中文古典文學 — 紅樓夢、西遊記、三國演義、老殘遊記、儒林外史、孽海花等
- 中日韓書名搜尋 — 直接輸入中文書名即可
- 數萬冊英文作品（1928 年前）— Shakespeare, Austen, Dickens, Fitzgerald 等經典
- 書庫沒收錄？回退到 Claude 知識分析，報告中清楚標示
三條路徑，同樣的深度分析：
| 輸入 | 發生什麼事 |
|---|---|
| PDF / EPUB 檔案 | 提取全文 → 深度分析 → HTML 報告 |
| 僅輸入書名 | 自動搜尋 70,000+ 冊公共領域書籍（Gutendex API）→ 下載全文 → 同樣的深度分析 |
| 僅書名（書庫無收錄） | 依 Claude 知識分析，報告中明確標示 |
公共領域書庫收錄 444+ 冊中文古典文學（紅樓夢、西遊記、三國演義、老殘遊記等）及數萬冊 1928 年前的英文作品。完整支援中日韓書名搜尋——直接輸入中文書名即可。
職責分離：
| 工作 | 由誰做 | Token 消耗 |
|---|---|---|
| 文字提取、分塊、書籍資訊 | extract-text.py |
零 |
| 公共領域書庫搜尋與下載 | Gutendex API | 零 |
| 閱讀理解、批判、整合 | Claude | 依書籍大小 |
| HTML 模板渲染 | render-report.py |
零 |
大型書籍（>80K tokens）會自動分塊、逐批分析、最後合併為一份連貫的報告——不是碎片拼接。
CRISP Reading 融合六種經過驗證的閱讀方法論，形成一套完整的閱讀實踐。單一方法無法涵蓋所有面向；組合起來才是完整的閱讀。
| 方法 | 它貢獻什麼 | 在 CRISP 中的角色 |
|---|---|---|
| Adler 分析閱讀 | 系統化理解——每位讀者都該問的四個問題 | 結構分析：這本書在談什麼、論點是什麼、有道理嗎、跟我有什麼關係 |
| TIPS 評分法（樊登） | 四維度評分——工具性、啟發性、實用性、科學性 | 快速分流：決定分析深度（4-5 略讀、6-8 標準、9-12 完整深讀） |
| Zettelkasten（Luhmann） | 三層筆記結構——閃念、文獻、永久 | 知識整合：把書中概念連結到你既有的知識網路 |
| 費曼技巧 | 「如果你沒辦法簡單地解釋它，代表你還不夠懂」 | 產出檢驗：報告中每個概念都必須用白話解釋 |
| 自我解釋法 | 暫停，用自己的話重述——What / Why / Connection | 理解確認：標記那些理解不夠深的段落 |
| 鋼鐵人論證 | 批判前先強化作者的論點到最強版本 | 智識誠實：批判最強版本的論點，而不是稻草人 |
- 誠實優先 — 僅憑書名（沒有 PDF）時，明確標示。絕不編造細節或虛構引句。
- 批判與證據等比 — 科學性越弱，批判越深。有紮實研究的書批判佔 15-20%；靠軼事的自我成長書佔 30-40%。
- 方法論不外露 — 報告中永遠不出現「Adler」、「Zettelkasten」、「鋼鐵人論證」。讀者看到的是自然的文字。
- 行動必須具體 — 每個行動承諾三要素缺一不可：何時（時間）、何處（場景）、做什麼（具體行為）。
- 留白引思考 — 報告刻意留下開放式提問，引導讀者形成自己的判斷，而非給出定論。
截圖即將加入——對任何書執行此技能即可產出你自己的報告。
每份報告包含：
| 區塊 | 深度 | 你會看到什麼 |
|---|---|---|
| 一句話評價 | 精華版 | 一句話捕捉這本書的核心價值 |
| TIPS 評分 | 精華版 | 四維度評分，告訴你這本書為什麼值得讀（或不值得） |
| 核心論點 | 精華版 | 3-7 個關鍵論點，可展開查看論證細節 |
| 關鍵概念 | 精華版 | 書中的核心詞彙，用白話解釋 |
| 精選引句 | 精華版 | 3-5 則最佳段落，非中文書附翻譯 |
| 行動計畫 | 精華版 | 3-5 個具體承諾，含時間、場景、行為 |
| 概念關係圖 | 完整版 | SVG 圖表呈現概念間的關聯 |
| 批判視角 | 完整版 | 邏輯跳躍、證據不足、隱含假設 |
| 知識連結 | 完整版 | Zettelkasten 風格的跨書連結 |
| 延伸閱讀 | 完整版 | 下一步該讀什麼，以及為什麼 |
報告是一個完全獨立的 HTML 檔案——不需要伺服器、不需要框架、不需要網路。用任何瀏覽器打開即可。特色：
- 精華版 / 完整版切換 — 先看重點，想深入再展開
- 深色模式 — 跟隨系統偏好，或手動切換
- 列印友善 — 自動展開所有區塊、強制淺色模式
- 響應式 — 手機、平板、桌面都好讀
- 智慧分塊 — 大型書籍自動切割、分批分析、智慧合併
- PDF + EPUB — PDF 用內建 pymupdf4llm；EPUB 可選用 document-to-markdown 技能
- 多語言輸入 — 讀任何語言的書，一律以繁體中文輸出，原文引句保留
- 零 Token 渲染 — HTML 生成完全由腳本處理；Claude 的算力全花在思考上
| Agent | 安裝方式 |
|---|---|
| Claude Code | npx skills add kcchien/crisp-reading |
| 其他相容 Agent | 任何能讀取 SKILL.md 作為指令的 AI Agent |
CRISP Reading 遵循開放的
SKILL.md慣例。任何能發現並載入SKILL.md的 AI Agent 都能使用。
- 支援 Agent Skills 的 AI 編碼助手（例如 Claude Code）
- Python 3.9+
pymupdf4llm（pip install pymupdf4llm）— PDF 文字提取- （選用） document-to-markdown 技能 — EPUB 支援
Turn any book into understanding.
An AI reading companion that deconstructs, critiques, and internalizes books —
then produces a beautiful, interactive HTML reading report.
CRISP = Comprehend · Review · Internalize · Synthesize · Practice
Five stages of deep reading: understand the structure, critically review the arguments, internalize through your own knowledge, synthesize into a coherent picture, and commit to practice.
Most AI "book summaries" give you a bullet list and call it a day. You walk away feeling like you read something, but nothing sticks.
CRISP Reading is different. It doesn't just summarize — it reads the book the way a thoughtful reader would: questioning the author's assumptions, stress-testing the arguments, connecting ideas to what you already know, and turning insights into concrete actions.
The output isn't a wall of text. It's an interactive HTML report — with collapsible arguments, curated quotes, critique sections, and a personal action plan — designed to feel like opening a beautifully typeset book, not scrolling through a dashboard.
npx skills add kcchien/crisp-readingThat's it. Your AI agent will automatically activate when you ask it to read or analyze a book.
Just talk to your AI assistant naturally:
| You say | What happens |
|---|---|
| "Help me read this book" + PDF | Extracts full text → deep analysis → interactive HTML report |
| "Analyze 老殘遊記" (title only) | Auto-downloads from 70,000+ public domain books → same deep analysis |
| "Analyze Atomic Habits" (not in library) | Analyzes from Claude's knowledge, clearly labeled |
| "Is this book worth reading?" | Quick TIPS 4-dimension evaluation, no fluff |
| "I have reading notes — help me go deeper" | Builds on your existing work, not starting over |
| "Read chapters 1-5 first" | Handles large books in chunks, merges into one report |
No PDF? No problem. The skill automatically searches Project Gutenberg (via Gutendex API) and downloads the full text for analysis — zero manual effort.
- 70,000+ public domain books available
- 444+ Chinese classics — 紅樓夢, 西遊記, 三國演義, 老殘遊記, 儒林外史, 孽海花, and more
- CJK search — type book titles in Chinese, Japanese, or Korean directly
- Tens of thousands of English works pre-1928 — Shakespeare, Austen, Dickens, Fitzgerald, and beyond
- Book not in the library? Falls back to Claude's knowledge, clearly marked in the report
Three paths to the same deep analysis:
| Input | What happens |
|---|---|
| PDF / EPUB file | Extract full text → deep analysis → HTML report |
| Book title only | Auto-search 70,000+ public domain books via Gutendex API → download full text → same deep analysis |
| Book title (not in library) | Analyze based on Claude's knowledge, clearly labeled as such |
The public domain library includes 444+ Chinese classics (紅樓夢, 西遊記, 三國演義, 老殘遊記, etc.) and tens of thousands of English works pre-1928. CJK title search is fully supported — just type the book name in Chinese, Japanese, or Korean.
Clear separation of concerns:
| Responsibility | Who | Token Cost |
|---|---|---|
| Text extraction, chunking, book info | extract-text.py |
Zero |
| Public domain book search & download | Gutendex API | Zero |
| Reading comprehension, critique, synthesis | Claude | Context-dependent |
| HTML template rendering | render-report.py |
Zero |
For large books (>80K tokens), the skill automatically splits the text into chunks, analyzes each batch, then merges all partial notes into one cohesive report — not a patchwork of fragments.
CRISP Reading fuses six proven reading methodologies into a single, coherent workflow. No single method covers everything; together they form a complete reading practice.
| Method | What It Contributes | Role in CRISP |
|---|---|---|
| Adler's Analytical Reading | Systematic comprehension — the four questions every reader should ask | Structural analysis: what is the book about, what are the arguments, is it true, what of it? |
| TIPS Evaluation (Fan Deng) | Four-dimensional book scoring — Toolability, Inspirability, Practicality, Scientificity | Quick triage: determines how deep the analysis should go (4-5 = skim, 6-8 = standard, 9-12 = full depth) |
| Zettelkasten (Luhmann) | Three-layer note structure — fleeting, literature, permanent | Knowledge integration: connects book ideas to your existing knowledge graph |
| Feynman Technique | "If you can't explain it simply, you don't understand it" | Output test: every concept in the report must be explainable in plain language |
| Self-Explanation | Pause and re-state in your own words — What / Why / Connection | Comprehension check: flags passages where understanding is shallow |
| Steel-Manning | Strengthen the author's argument before critiquing it | Intellectual honesty: critique the best version of the argument, not a strawman |
- Honesty over completeness — If the skill only has a book title (no PDF), it says so. It never fabricates details or invents quotes.
- Critique scales with evidence — The weaker the scientific backing, the deeper the critique. A well-researched book gets 15-20% critical content; a self-help book built on anecdotes gets 30-40%.
- Methods stay invisible — The report never mentions "Adler", "Zettelkasten", or "Steel-Manning". Readers see natural prose, not methodology labels.
- Actions must be specific — Every action item has three parts: when (time), where (context), and what (concrete behavior). "Read more" is not an action; "Tomorrow morning, spend 15 minutes journaling about Chapter 3's framework" is.
- Leave room for thinking — The report deliberately poses open questions rather than giving definitive verdicts. The goal is to spark the reader's own judgment.
Screenshot coming soon — run the skill on any book to generate your own.
Each report includes:
| Section | Depth | What You Get |
|---|---|---|
| One-line verdict | Essential | A single sentence that captures the book's core value |
| TIPS score | Essential | Four-dimensional rating that tells you why the book matters (or doesn't) |
| Core arguments | Essential | 3-7 key arguments, each expandable with supporting detail |
| Key concepts | Essential | The book's vocabulary, explained in plain language |
| Curated quotes | Essential | 3-5 best passages with translation (for non-Chinese books) |
| Action plan | Essential | 3-5 concrete commitments with time, context, and specific behavior |
| Concept map | Deep | SVG diagram showing how the book's ideas relate to each other |
| Critical perspectives | Deep | Where the author's logic breaks, evidence is thin, or assumptions are hidden |
| Knowledge links | Deep | Zettelkasten-style connections to other books and ideas |
| Further reading | Deep | What to read next, and why |
The report ships as a self-contained HTML file — no dependencies, no server, no JavaScript frameworks. Open it in any browser. It features:
- Essential / Deep toggle — Start with the highlights, expand when you want more
- Dark mode — Follows your system preference, or toggle manually
- Print-friendly — Expands all sections and forces light mode for clean printing
- Responsive — Reads well on phones, tablets, and desktops
- Smart chunking — Automatically splits large books, analyzes in batches, merges intelligently
- PDF + EPUB — PDF via built-in pymupdf4llm; EPUB via optional document-to-markdown skill
- Multilingual input — Reads books in any language, always outputs in Traditional Chinese with original quotes preserved
- Zero-token rendering — HTML generation is fully scripted; Claude's context is spent on thinking, not formatting
| Agent | Install Method |
|---|---|
| Claude Code | npx skills add kcchien/crisp-reading |
| Other compatible agents | Any agent that reads SKILL.md as instructions |
CRISP Reading follows the open
SKILL.mdconvention. Any AI agent that can discover and loadSKILL.mdfiles will work.
crisp-reading/
├── SKILL.md # Agent instructions (entry point)
├── scripts/
│ ├── extract-text.py # PDF/EPUB → plain text
│ └── render-report.py # Analysis JSON → HTML report
├── references/
│ ├── json-schema.md # Output JSON structure spec
│ ├── analysis.md # Reading methodology details
│ ├── design-spec.md # HTML report design guidelines
│ └── ebook-library.md # Public domain book sources & API guide
└── assets/
└── reading-report-template.html # HTML template (851 lines)
- An AI coding assistant that supports Agent Skills (e.g., Claude Code)
- Python 3.9+
pymupdf4llm(pip install pymupdf4llm) — for PDF text extraction- (Optional) document-to-markdown skill — for EPUB support