Deep Reader · 深度阅读与论文分析

一个 Claude Code skill，帮你快速拆解一本书的核心骨架、提炼一篇论文的核心想法、追溯一篇论文的研究脉络。输出结构化笔记，直接推送到飞书知识库。

---

English | 中文

---

这是什么

Deep Reader 是一个给 Claude Code 用的 skill。你用一句自然语言触发它，它帮你：

• 读书：用五步法提取一本书的骨架（动机、未证前提、分析框架、核心结论、便携知识）
• 读论文：用四步法提炼一篇论文（核心想法、创新点、研究脉络、关键技术概念翻译）
• 追溯论文：从一篇论文出发，向前找它的前序工作，向后找它的后续引用，画出完整研究脉络
• 找书 / 找论文：按主题搜索推荐候选，你挑选后再提取

所有输出直接推送到飞书知识库，不在本地保存文件。

触发词

你说	它做什么
读书：人生的智慧	提取《人生的智慧》的骨架，推送飞书
读论文：Attention Is All You Need	提炼 Transformer 论文，推送飞书
追溯：Attention Is All You Need	追溯 Transformer 论文的前后相关论文
找书：人生哲学	搜索人生哲学相关书单，你挑完后提取
找论文：大模型	搜索大模型方向论文，你挑完后提炼

前置条件

1. Claude Code：Anthropic 的 CLI 编码工具
2. lark-cli：飞书/Lark 的 CLI 工具（用于推送文档到飞书知识库）
3. 飞书知识库：你需要在飞书中创建一个知识库，并获取节点 token

安装

将 SKILL.md 和 references/ 目录复制到你的 Claude Code skills 目录下：

# 例如，如果你的项目在 ~/my-project
mkdir -p ~/my-project/.claude/skills/deep-reader
cp SKILL.md ~/my-project/.claude/skills/deep-reader/
cp -r references ~/my-project/.claude/skills/deep-reader/

配置

使用前需要配置你自己的飞书知识库节点 token。

打开 SKILL.md，找到「飞书知识库结构」部分，将占位符替换为你的实际 token：

<YOUR_SPACE_ID>              → 你的飞书知识空间 ID
<YOUR_INPUT_NODE_TOKEN>      → "输入"节点的 token
<YOUR_PAPER_NODE_TOKEN>      → "论文"节点的 token
<YOUR_READING_NODE_TOKEN>    → "阅读"节点的 token

同样，打开 references/ 下的三个 .md 文件，替换末尾的占位符 token。

获取 token 的方法：

lark-cli wiki spaces get_node --params '{"token":"你的wiki页面URL中的token"}'

三种方法

1. 书籍骨架提取（五步法）

步骤	问题
诊断动机	作者为什么写这本书？
未证前提	哪些论点是假设成立、没证明的？
分析框架	作者独有的视角和术语是什么？
核心结论	全书最关键的一个判断
便携知识	公式 / 一句话 / 结构图

2. 论文提炼（四步法）

步骤	内容
核心想法	大白话讲清楚论文做了什么
创新点	跟之前的工作比，新在哪
研究脉络	在研究领域里的位置
技术概念翻译	3-5 个核心概念，翻译成普通人能懂的话

3. 论文追溯（四步法）

步骤	内容
源论文核心	一句话讲清楚源论文的贡献
向前追溯	3-5 篇最重要的前序论文
向后追踪	3-5 篇引用/扩展了源论文的后续工作
研究脉络	时间线 + 演化关系 + 下一步建议

项目结构

deep-reader-skill/
├── SKILL.md                          # 主文件：触发词路由、方法论概览
├── references/
│   ├── book-method.md                # 书籍骨架提取详细方法论
│   ├── paper-method.md               # 论文提炼详细方法论
│   └── paper-river.md                # 论文追溯详细方法论
├── README.md                         # 本文件
└── LICENSE                           # MIT

致谢

方法论改编自李继刚的 ljg-book / ljg-paper / ljg-paper-river skills。

License

MIT

---

Deep Reader · Deep Reading & Paper Analysis

A Claude Code skill for extracting book skeletons, distilling AI papers, and tracing research lineages. Outputs structured notes directly to your Lark/Feishu wiki.

What Is This

Deep Reader is a skill for Claude Code. You trigger it with a natural language command, and it helps you:

• Read a book: Extract the skeleton using a 5-step method (motivation, unproven assumptions, analytical framework, core conclusion, portable knowledge)
• Read a paper: Distill a paper using a 4-step method (core idea, innovation, research context, key concept translation)
• Trace a paper: From one paper, find its predecessors and successors, and draw the complete research lineage
• Discover books / papers: Search by topic, get recommendations, then extract after you pick

All outputs are pushed to your Lark/Feishu wiki — no local files.

Trigger Commands

You say	It does
读书：The Art of War	Extracts the skeleton of "The Art of War", pushes to Lark
读论文：Attention Is All You Need	Distills the Transformer paper, pushes to Lark
追溯：Attention Is All You Need	Traces related papers before and after Transformer
找书：philosophy of life	Searches for philosophy book recommendations
找论文：large language models	Searches for LLM papers

Note: Trigger keywords are in Chinese. You can modify them in SKILL.md for your preferred language.

Prerequisites

1. Claude Code: Anthropic's CLI coding tool
2. lark-cli: Lark/Feishu CLI (for pushing docs to your wiki)
3. A Lark/Feishu wiki: You need a wiki space with proper node tokens configured

Installation

Copy SKILL.md and the references/ directory to your Claude Code skills folder:

mkdir -p your-project/.claude/skills/deep-reader
cp SKILL.md your-project/.claude/skills/deep-reader/
cp -r references your-project/.claude/skills/deep-reader/

Configuration

Before using, configure your own Lark/Feishu wiki node tokens.

Open SKILL.md, find the "飞书知识库结构" section, and replace placeholders with your actual tokens:

<YOUR_SPACE_ID>              → Your Lark wiki space ID
<YOUR_INPUT_NODE_TOKEN>      → Token for the "Input" parent node
<YOUR_PAPER_NODE_TOKEN>      → Token for the "Papers" node
<YOUR_READING_NODE_TOKEN>    → Token for the "Reading" node

Also replace placeholder tokens in the three .md files under references/.

The Three Methods

1. Book Skeleton Extraction (5 steps)

Step	Question
Diagnose motivation	Why did the author write this book?
List unproven assumptions	What claims are taken for granted?
Isolate the framework	What's the author's unique perspective and terminology?
Distill core conclusion	The single most important judgment
Compress to portable knowledge	Formula / one sentence / structure diagram

2. Paper Distillation (4 steps)

Step	Content
Core idea	Explain what the paper does in plain language
Innovation	What's new compared to prior work
Research context	Where does this paper sit in the field
Key concept translation	3-5 technical concepts translated for non-experts

3. Paper Lineage Tracing (4 steps)

Step	Content
Source paper core	One sentence on the source paper's contribution
Trace backward	3-5 most important predecessor papers
Trace forward	3-5 papers that cite/extend the source
Research lineage	Timeline + evolution + what to read next

Project Structure

deep-reader-skill/
├── SKILL.md                          # Main file: trigger routing, method overview
├── references/
│   ├── book-method.md                # Detailed book skeleton methodology
│   ├── paper-method.md               # Detailed paper distillation methodology
│   └── paper-river.md                # Detailed paper tracing methodology
├── README.md                         # This file
└── LICENSE                           # MIT

Acknowledgments

Methodology adapted from Li Jigang's ljg-book / ljg-paper / ljg-paper-river skills.

License

MIT