Three-mode paper reading assistant for Claude Code | 三模式论文阅读助手
TypeScript
22
10 commits
updated May 13, 2026
A Claude Code skill for reading, understanding, and presenting academic papers.
一个用于论文阅读、深度理解和汇报准备的 Claude Code Skill。
| Mode / 模式 | Purpose / 用途 | Time / 耗时 | Interaction / 交互 |
|---|---|---|---|
| Speed Read / 速览 | Quick digest / 快速消化，判断是否值得深读 | ~5 min | One-shot / 一次性输出 |
| Deep Learn / 学习 | In-depth study / 大白话深度理解，边学边存 | 20-40 min | Multi-turn / 多轮交互 |
| Present / 展示 | Slide preparation / 准备演讲 slides | 15-30 min | Page-by-page / 逐页讨论 |
| Batch Search / 批量检索 | Search by topic / 按主题搜索论文 | 3-5 min | Table + select / 表格 + 选择 |
| Batch Download / 批量下载 | Download multiple papers / 粘贴链接批量下载 | 1-3 min | Auto-dedup / 自动去重 |
The CLI skill is a single folder you copy. The optional Web UI lives in the same repo — see Web UI below.
CLI skill 只是一个文件夹。可选的 Web UI 在同一仓库里——见下方 Web UI 章节。
Project-level (recommended) / 项目级安装（推荐）：
git clone https://github.com/nekoneko0831/paper-lens.git /tmp/paper-lens
cp -r /tmp/paper-lens/.claude/skills/paper-lens <your-project>/.claude/skills/
User-level (global) / 用户级安装（全局可用）：
git clone https://github.com/nekoneko0831/paper-lens.git /tmp/paper-lens
cp -r /tmp/paper-lens/.claude/skills/paper-lens ~/.claude/skills/
Note: copying just the skill folder gives you the CLI experience. To use the browser UI, keep the full repo and follow the Web UI instructions.
注意：只拷 skill 文件夹只会得到 CLI 模式。如果想用浏览器界面，请保留整个仓库并按 Web UI 章节启动。
pip install pymupdf
Just talk to Claude Code: / 在 Claude Code 中直接对话：
> Help me read this paper: https://arxiv.org/pdf/1706.03762
> 帮我读这篇论文：https://arxiv.org/pdf/1706.03762
> Search for LLM agent evaluation papers
> 搜索 LLM agent evaluation 相关论文
> Download these: 2501.12948 2309.12288 2401.05566
> 帮我下载这些论文：2501.12948 2309.12288 2401.05566
Claude will auto-detect your intent and enter the matching mode.
Claude 会自动识别意图并进入对应模式。
5-minute digest of any paper. Outputs an info card, TL;DR, key innovation, methodology, results, and discussion questions.
5 分钟快速消化一篇论文。输出基本信息卡片、TL;DR、核心创新、主要方法、实验结果和深度问答。
Trigger: "speed read this paper" / 「帮我速览这篇论文」
20-40 minute interactive deep dive. All concepts explained in plain language — like explaining to a smart friend outside the field.
20-40 分钟交互式深度学习。所有概念用「大白话」解释——想象在给一个聪明但非本领域的朋友讲解。
Key features / 核心特性：
Trigger: "help me learn this paper" / 「帮我学习这篇论文」
Prepare slide content for a paper talk in 15-30 minutes.
15-30 分钟准备一场论文汇报的 slides 内容。
Key features / 核心特性：
Trigger: "prepare slides for this paper" / 「帮我准备 slides」
Search papers by topic keywords. Returns a ranked table with links, then offers to download selected papers.
按主题关键词搜索论文，生成带链接的排序表格，支持一键下载。
Trigger: "search for X papers" / 「搜索 X 相关论文」
Paste multiple arXiv URLs or IDs to download in batch. Supports mixed format input with 3-layer dedup.
粘贴多个 arXiv 链接或 ID，自动去重后批量下载。
Trigger: paste multiple arXiv links / 直接粘贴多个链接
The three reading modes chain seamlessly: / 三种阅读模式可自由串联：
Speed Read → Deep Learn → Present
速览（值不值得读？）→ 学习（深度理解）→ 展示（准备汇报）
Content from Speed Read is reused by Deep Learn — no repeated work.
速览的内容会被学习模式复用，不会重复劳动。
paper-notes/<paper-name>/
├── paper.pdf # Original PDF / 原始 PDF
├── extracted-text.md # Full text / 提取的全文
├── images/ # All extracted images / 所有提取图片
├── figures/ # [Present] Selected figures / 筛选后的论文原图
├── speed-read.md # Speed Read output / 速览输出
├── deep-learn.md # Deep Learn output / 学习输出（增量保存）
└── slides-content.md # Present mode output / 展示模式输出
.claude/skills/paper-lens/
├── SKILL.md # Entry point / 主入口
├── references/
│ ├── speed-read.md # Speed Read instructions / 速览指令
│ ├── deep-learn.md # Deep Learn instructions / 学习指令
│ ├── present.md # Present instructions / 展示指令
│ ├── export-pdf.md # PDF export instructions / PDF 导出指令
│ ├── batch-search.md # Batch Search instructions / 批量检索指令
│ └── batch-download.md # Batch Download instructions / 批量下载指令
└── scripts/
├── extract_figures.py # Figure extraction / 图片提取
└── md_to_pdf.py # Markdown → PDF / MD 转 PDF
The CLI skill works on its own. If you want a chat-style browser UI, this repo also ships an optional Web UI: a FastAPI backend that wraps the skill, plus a Next.js frontend.
CLI skill 本身已可独立使用。如果你想要浏览器聊天界面，仓库还提供可选的 Web UI：FastAPI 后端封装 skill，Next.js 前端。
paper-lens-backend/ # FastAPI server (port 8765) — drives Claude Code CLI via stream-json + MCP
paper-lens-web/ # Next.js app (port 3000) — chat UI, paper browser
claude once and complete /login)claude 完成 /login）⚠️ Do not run the backend with
sudo. The backend spawns aclaudesubprocess that inherits the invoking user's credentials. Asudo-spawned subprocess won't find your login token and every chat will fail withNot logged in · Please run /login.⚠️ 不要用
sudo启动后端。后端会 spawnclaude子进程，并继承调用者的凭证；用sudo起的子进程拿不到你的登录态，每次对话都会失败。
The backend keeps one long-lived Claude subprocess per browser session. Structured questions in the UI are routed through a tiny local MCP server so the agent truly waits for the user's answer before continuing.
后端会为每个浏览器会话保留一个长寿命 Claude 子进程；结构化选择题通过本地 MCP 小服务中转，确保模型真的等用户回答后再继续。
# Backend
cd paper-lens-backend
pip install -r requirements.txt
# Frontend
cd ../paper-lens-web
cp .env.local.example .env.local # adjust NEXT_PUBLIC_BACKEND_URL if needed
npm install
Open one terminal:
cd paper-lens-web
npm run dev
Then open http://localhost:3000 in your browser.
npm run dev starts the FastAPI backend behind the frontend and restarts it if it exits while the frontend dev server is still running. In normal use, users only deal with the frontend URL.
When the skill loads, Phase -1 detects whether ports 3000 and 8765 are alive and tells you the URL — no auto-open. If the Web UI is not running, the suggested startup path is the single frontend command above.
Skill 加载时，Phase -1 会探测 3000 和 8765 端口，活着就给你一行 URL 提示；如果没启动，只提示上面的一条前端启动命令。
Present mode generates slides-content.md with figure references. To convert into a self-contained HTML presentation, hand it to the frontend-slides skill; paper-lens should not build HTML decks itself.
展示模式生成 slides-content.md。要转换为 HTML 演示文稿，交给 frontend-slides skill；paper-lens 自己不写 HTML deck。
Three-mode paper reading assistant for Claude Code | 三模式论文阅读助手
TypeScript
22
10 commits
updated May 13, 2026
A Claude Code skill for reading, understanding, and presenting academic papers.
一个用于论文阅读、深度理解和汇报准备的 Claude Code Skill。
| Mode / 模式 | Purpose / 用途 | Time / 耗时 | Interaction / 交互 |
|---|---|---|---|
| Speed Read / 速览 | Quick digest / 快速消化，判断是否值得深读 | ~5 min | One-shot / 一次性输出 |
| Deep Learn / 学习 | In-depth study / 大白话深度理解，边学边存 | 20-40 min | Multi-turn / 多轮交互 |
| Present / 展示 | Slide preparation / 准备演讲 slides | 15-30 min | Page-by-page / 逐页讨论 |
| Batch Search / 批量检索 | Search by topic / 按主题搜索论文 | 3-5 min | Table + select / 表格 + 选择 |
| Batch Download / 批量下载 | Download multiple papers / 粘贴链接批量下载 | 1-3 min | Auto-dedup / 自动去重 |
The CLI skill is a single folder you copy. The optional Web UI lives in the same repo — see Web UI below.
CLI skill 只是一个文件夹。可选的 Web UI 在同一仓库里——见下方 Web UI 章节。
Project-level (recommended) / 项目级安装（推荐）：
git clone https://github.com/nekoneko0831/paper-lens.git /tmp/paper-lens
cp -r /tmp/paper-lens/.claude/skills/paper-lens <your-project>/.claude/skills/
User-level (global) / 用户级安装（全局可用）：
git clone https://github.com/nekoneko0831/paper-lens.git /tmp/paper-lens
cp -r /tmp/paper-lens/.claude/skills/paper-lens ~/.claude/skills/
Note: copying just the skill folder gives you the CLI experience. To use the browser UI, keep the full repo and follow the Web UI instructions.
注意：只拷 skill 文件夹只会得到 CLI 模式。如果想用浏览器界面，请保留整个仓库并按 Web UI 章节启动。
pip install pymupdf
Just talk to Claude Code: / 在 Claude Code 中直接对话：
> Help me read this paper: https://arxiv.org/pdf/1706.03762
> 帮我读这篇论文：https://arxiv.org/pdf/1706.03762
> Search for LLM agent evaluation papers
> 搜索 LLM agent evaluation 相关论文
> Download these: 2501.12948 2309.12288 2401.05566
> 帮我下载这些论文：2501.12948 2309.12288 2401.05566
Claude will auto-detect your intent and enter the matching mode.
Claude 会自动识别意图并进入对应模式。
5-minute digest of any paper. Outputs an info card, TL;DR, key innovation, methodology, results, and discussion questions.
5 分钟快速消化一篇论文。输出基本信息卡片、TL;DR、核心创新、主要方法、实验结果和深度问答。
Trigger: "speed read this paper" / 「帮我速览这篇论文」
20-40 minute interactive deep dive. All concepts explained in plain language — like explaining to a smart friend outside the field.
20-40 分钟交互式深度学习。所有概念用「大白话」解释——想象在给一个聪明但非本领域的朋友讲解。
Key features / 核心特性：
Trigger: "help me learn this paper" / 「帮我学习这篇论文」
Prepare slide content for a paper talk in 15-30 minutes.
15-30 分钟准备一场论文汇报的 slides 内容。
Key features / 核心特性：
Trigger: "prepare slides for this paper" / 「帮我准备 slides」
Search papers by topic keywords. Returns a ranked table with links, then offers to download selected papers.
按主题关键词搜索论文，生成带链接的排序表格，支持一键下载。
Trigger: "search for X papers" / 「搜索 X 相关论文」
Paste multiple arXiv URLs or IDs to download in batch. Supports mixed format input with 3-layer dedup.
粘贴多个 arXiv 链接或 ID，自动去重后批量下载。
Trigger: paste multiple arXiv links / 直接粘贴多个链接
The three reading modes chain seamlessly: / 三种阅读模式可自由串联：
Speed Read → Deep Learn → Present
速览（值不值得读？）→ 学习（深度理解）→ 展示（准备汇报）
Content from Speed Read is reused by Deep Learn — no repeated work.
速览的内容会被学习模式复用，不会重复劳动。
paper-notes/<paper-name>/
├── paper.pdf # Original PDF / 原始 PDF
├── extracted-text.md # Full text / 提取的全文
├── images/ # All extracted images / 所有提取图片
├── figures/ # [Present] Selected figures / 筛选后的论文原图
├── speed-read.md # Speed Read output / 速览输出
├── deep-learn.md # Deep Learn output / 学习输出（增量保存）
└── slides-content.md # Present mode output / 展示模式输出
.claude/skills/paper-lens/
├── SKILL.md # Entry point / 主入口
├── references/
│ ├── speed-read.md # Speed Read instructions / 速览指令
│ ├── deep-learn.md # Deep Learn instructions / 学习指令
│ ├── present.md # Present instructions / 展示指令
│ ├── export-pdf.md # PDF export instructions / PDF 导出指令
│ ├── batch-search.md # Batch Search instructions / 批量检索指令
│ └── batch-download.md # Batch Download instructions / 批量下载指令
└── scripts/
├── extract_figures.py # Figure extraction / 图片提取
└── md_to_pdf.py # Markdown → PDF / MD 转 PDF
The CLI skill works on its own. If you want a chat-style browser UI, this repo also ships an optional Web UI: a FastAPI backend that wraps the skill, plus a Next.js frontend.
CLI skill 本身已可独立使用。如果你想要浏览器聊天界面，仓库还提供可选的 Web UI：FastAPI 后端封装 skill，Next.js 前端。
paper-lens-backend/ # FastAPI server (port 8765) — drives Claude Code CLI via stream-json + MCP
paper-lens-web/ # Next.js app (port 3000) — chat UI, paper browser
claude once and complete /login)claude 完成 /login）⚠️ Do not run the backend with
sudo. The backend spawns aclaudesubprocess that inherits the invoking user's credentials. Asudo-spawned subprocess won't find your login token and every chat will fail withNot logged in · Please run /login.⚠️ 不要用
sudo启动后端。后端会 spawnclaude子进程，并继承调用者的凭证；用sudo起的子进程拿不到你的登录态，每次对话都会失败。
The backend keeps one long-lived Claude subprocess per browser session. Structured questions in the UI are routed through a tiny local MCP server so the agent truly waits for the user's answer before continuing.
后端会为每个浏览器会话保留一个长寿命 Claude 子进程；结构化选择题通过本地 MCP 小服务中转，确保模型真的等用户回答后再继续。
# Backend
cd paper-lens-backend
pip install -r requirements.txt
# Frontend
cd ../paper-lens-web
cp .env.local.example .env.local # adjust NEXT_PUBLIC_BACKEND_URL if needed
npm install
Open one terminal:
cd paper-lens-web
npm run dev
Then open http://localhost:3000 in your browser.
npm run dev starts the FastAPI backend behind the frontend and restarts it if it exits while the frontend dev server is still running. In normal use, users only deal with the frontend URL.
When the skill loads, Phase -1 detects whether ports 3000 and 8765 are alive and tells you the URL — no auto-open. If the Web UI is not running, the suggested startup path is the single frontend command above.
Skill 加载时，Phase -1 会探测 3000 和 8765 端口，活着就给你一行 URL 提示；如果没启动，只提示上面的一条前端启动命令。
Present mode generates slides-content.md with figure references. To convert into a self-contained HTML presentation, hand it to the frontend-slides skill; paper-lens should not build HTML decks itself.
展示模式生成 slides-content.md。要转换为 HTML 演示文稿，交给 frontend-slides skill；paper-lens 自己不写 HTML deck。