Arxiv Daily Skill

每日自动搜集 arXiv 论文（LLM/RAG/Agent 方向），带机构信息，飞书推送

Security Scan

The skill's code and instructions match its stated purpose (daily arXiv collection, local formatting, and optional Feishu webhook push); no disproportionate permissions or hidden endpoints were found, but users should verify any webhook they configure and review local file writes before enabling automated runs.

Purpose & Capability

Name/description (daily arXiv paper aggregation + Feishu push) align with included files: fetch_arxiv.py, affiliation extractors, push shell scripts, cron/run helpers. Required capabilities (HTTP access to arxiv.org, optional Feishu webhook) are consistent with the stated purpose.

Instruction Scope

Runtime instructions and scripts perform exactly the described tasks: query arXiv API, parse XML, optionally scrape arXiv HTML to extract affiliations, format Markdown, save a report to /tmp and POST a message to a configured Feishu webhook. No instructions read unrelated system secrets or other system config. Note: deep_analyze.py and other scripts scrape arXiv HTML and look for GitHub links and known institution keywords — this is expected but could produce false positives/negatives in affiliation extraction.

Install Mechanism

No install spec or remote downloads; the skill is instruction-plus-source files. There is no package download or archive extraction. Risk from installation is low, but running the included scripts will execute the bundled Python/bash files.

Credentials

Registry metadata declares no required env vars, which is reasonable because Feishu webhook is optional. However runtime scripts expect either FEISHU_WEBHOOK_URL (env) or a local feishu_config.json to send push messages. The only external endpoints targeted are arxiv.org, github.com (pattern matching), and the user-provided Feishu webhook. Users should treat any webhook URL as a secret/destination and validate it before configuring.

Persistence & Privilege

always:false and no modifications to other skills or system-wide agent config. The skill writes a small affiliations_db.json in its own directory (and saves reports to /tmp); crontab usage is suggested but requires explicit user action. No elevated privileges requested.

Assessment

This skill appears to do what it claims: fetch recent arXiv papers, try to extract affiliation info, format a Markdown report, and optionally POST a preview to a Feishu webhook. Before installing or enabling automated runs: 1) Inspect feishu_config.json and any webhook URL you will set — a webhook is effectively an outbound destination and will receive the report content (authors, abstracts, etc.). Do not point it to untrusted/third-party endpoints. 2) Review the scripts (fetch_arxiv.py, deep_analyze.py, get_affiliations.py) if you care about privacy or accuracy — they scrape arXiv pages and auto-update affiliations_db.json in the skill directory. 3) If enabling cron, add the crontab entry yourself rather than using any automatic installer, and run the scripts interactively first to confirm behavior. 4) Expect occasional affiliation extraction errors; the code heuristics match keywords and GitHub repo names, which can be noisy. If you want more assurance, run the scripts in a restricted environment or container and provide a trusted webhook only.

arXiv 每日论文精选 Skill

功能

每日自动搜集 arXiv 上 LLM/RAG/Agent/Context/Harness 等方向的最新高价值论文（2-5 篇），整理成易于阅读的简洁形式，通过飞书推送给用户。

触发方式

• 定时任务：每个工作日上午 9:00 自动执行
• 手动触发：用户发送"今日论文"、"arxiv 推荐"、"论文推送"、"看看今天的论文"等关键词
• 深度分析：用户发送"详解第 X 篇"、"分析 [arXiv ID]"、"用 ljg-paper 分析这篇"

执行流程

1. 搜集论文

使用 arXiv API 搜索以下方向的最新论文（过去 3 天内）：

• Large Language Model (LLM)
• Retrieval Augmented Generation (RAG)
• AI Agent / Autonomous Agent
• Context Window / Context Management
• AI Harness / Agent Harness
• Transformer Architecture
• Language Model Reasoning
• Multimodal LLM

2. 筛选排序

• 过滤掉非 CS 领域（排除物理、生物、医学等）
• 优先选择 cs.CL, cs.AI, cs.LG, stat.ML 分类
• 根据标题关键词打分排序
• 去重后选择 Top 10（目录）+ Top 5（详解）

3. 机构信息提取

• 从机构数据库 affiliations_db.json 中查找（自动累积）
• 首次遇到的论文自动从 arXiv HTML 页面提取
• 支持知名机构识别：清华、北大、斯坦福、MIT、Google、Meta、华为等

4. 格式化输出

生成 Markdown 格式报告：

# 📚 今日 ArXiv 论文推送 (2026-04-16)

## 📋 目录
1. 论文标题 1
2. 论文标题 2
...

---

## 📖 论文详解

### 1. 论文完整标题
**arXiv**: 2604.13016v1 | 2026-04-14 | Tsinghua University + THUNLP
**作者**: Yaxuan Li, Yuxin Zuo, Bingxiang He et al.
**链接**: https://arxiv.org/abs/2604.13016v1

**中文摘要**:
> ...

**解决的问题**:
- ...

**核心创新**:
- ...

**为什么重要**:
- ...

**局限性**:
- ...

5. 推送

• 飞书推送：目录 + Top 5 详解（通过 Webhook）
• 文件保存：完整报告保存到 /tmp/arxiv_daily_YYYYMMDD.md

输出要求

• 紧凑目录：仅标题，快速浏览
• 详解格式：包含机构、作者、链接、摘要、问题、创新、重要性、局限性
• 机构信息：每篇论文必须显示机构（如"Tsinghua University + THUNLP"）
• 中文友好：摘要保留英文原文，分析使用中文

配置

• 推送时间：工作日 09:00（Asia/Shanghai）
• 推送渠道：飞书（Webhook）
• 论文数量：目录 10 篇 + 详解 5 篇/天
• 时间范围：过去 3 天内的新论文

文件结构

arxiv-daily-skill/
├── SKILL.md                    # 本文件
├── fetch_arxiv.py              # 核心脚本
├── deep_analyze.py             # 深度分析工具
├── cron_run.sh                 # Cron 执行入口
├── run.sh                      # 手动测试
├── affiliations_db.json        # 机构数据库（自动累积）
├── feishu_config.json          # 飞书配置
└── README.md                   # 使用文档

依赖

• arXiv API（无需 API Key）
• Python 3.6+
• 飞书 Webhook（可选，用于推送）
• cron（定时任务）

高级功能

深度分析模式

对单篇论文进行详细解读：

python3 deep_analyze.py 2604.13016

或调用 ljg-paper skill：

• "用 ljg-paper 分析 2604.13016"
• "详解第 3 篇论文"

机构数据库

自动累积的机构信息保存在 affiliations_db.json：

{
  "affiliations": {
    "2604.13016": ["Tsinghua University", "THUNLP"],
    "2604.13029": ["Huawei Technologies Co., Ltd."]
  }
}

自定义搜索方向

修改 fetch_arxiv.py 中的 SEARCH_CONFIGS：

SEARCH_CONFIGS = [
    {"query": "cat:cs.AI AND all:large language model", "label": "LLM", "max": 3},
    # 添加新的搜索方向...
]

飞书配置

1. 在飞书群聊中添加自定义机器人
2. 复制 Webhook 地址
3. 编辑 feishu_config.json，填入 webhook_url
4. 测试推送：./cron_run.sh

示例对话

用户: "今天有什么新论文？" 助理: 推送今日 ArXiv 论文目录 + Top 5 详解

用户: "详解第 3 篇" 助理: 调用 deep_analyze.py 或 ljg-paper 进行深度分析

用户: "用 ljg-paper 分析 2604.13016" 助理: 生成完整的中文解读报告（费曼翻译 + 核心概念 + 博导审稿）

---

版本: v5
最后更新: 2026-04-16
维护者: OpenClaw Agent