Knowledge Curator

自动抓取用户指定链接内容，整理分类为结构化知识库，支持增量更新和多条件内容检索。

Security Scan

The skill mostly matches its stated purpose (scraping links and storing structured markdown entries), but there are several inconsistencies and privacy risks (contradictory trigger descriptions, undocumented runtime requirements, and saved URLs that can include tokens) that warrant caution before installing.

Purpose & Capability

The skill's name/description (web scraping + knowledge storage) aligns with the provided code (fetch.js, summarize.js, store.js, query.js). However SKILL.md lists dependencies/tools like 'web_fetch' and 'exec' even though the shipped code uses Node's https/http modules and filesystem; the registry metadata declares no required binaries or env vars. That mismatch (declared 'none' vs SKILL.md/tool mentions vs Node scripts) is a configuration/documentation inconsistency — not necessarily malicious, but worth flagging.

Instruction Scope

SKILL.md claims the skill only saves content when the user gives an explicit '收藏/保存' command, but README/QUICKSTART contain contradictory wording (some places say '直接发送链接即可' without explicit save). The code fetches arbitrary URLs and saves full scraped content and metadata into the local knowledge-base directory. Example saved files show full original URLs (including query parameters like xsec_token) and scraped metadata (IP 属地, interaction counts). This means sensitive tokens or metadata in URLs or page content can be persistently stored. The instructions do not clearly require sanitization of saved URLs or explicit user confirmation for potentially sensitive content, which is scope creep from a privacy perspective.

Install Mechanism

There is no install spec (instruction-only) which reduces supply-chain risk; code files are included in the package and expect Node.js runtime (package.json engines node >=14). No downloads from third-party URLs or archive extraction are present. Still, because code will be executed (Node scripts), the absence of an install spec but presence of runnable scripts is an inconsistency to note — reviewers should ensure the runtime runs the shipped code only and not an external installer.

Credentials

The skill declares no required env vars or credentials. config.js includes an optional aiService block that references process.env.AI_API_KEY but it's commented and optional. That is proportionate. However the skill will store arbitrary scraped data (including full URLs with query tokens and scraped page metadata). Even without requesting credentials, it can persist tokens and other sensitive data from web pages into the knowledge-base directory—this is a privacy risk and should be considered before use.

Persistence & Privilege

always:false (no forced always-on) and no system-wide configuration changes are requested. The skill writes files to its own knowledge-base/ directory and updates index.json — expected behavior for a local curator. Note: the agent can invoke the skill autonomously (default platform behavior); combined with storage of scraped content this increases blast radius if misused, but autonomous invocation alone is not flagged by policy.

What to consider before installing

What to check before installing or running this skill: - Trigger behavior: SKILL.md says it will only save when the user explicitly asks to '收藏/保存', but README/QUICKSTART contain contradictory examples that imply links may be auto-saved. Confirm (by reading the code or testing in a sandbox) that the skill only saves when you intend it to. - Sensitive data in saved entries: The sample knowledge files include full URLs (with query parameters like xsec_token) and scraped metadata (IP 属地, interaction counts). The skill will persist whatever it scrapes — including tokens or personal data embedded in URLs or page content. If you plan to save links that may contain session tokens or PII, either sanitize the URLs before saving or disable auto-save. - Runtime execution: The package contains Node.js scripts (no install spec). Running the skill will execute these scripts locally; review scripts (fetch.js, store.js, etc.) to confirm there are no unexpected outbound endpoints or obfuscated code. Although no external exfiltration endpoints are present in the provided files, always run first in a restricted environment. - AI integration keys: config.js documents an optional aiService that can use an API key from environment (process.env.AI_API_KEY). If you enable AI summarization, provide a dedicated API key with minimal scope and quota limits. - File system location and backups: The skill writes to knowledge-base/ and exports/ by default. Make sure that directory is in a location you control and that backups/permissions are appropriate. Consider restricting access to that folder if it will contain sensitive data. - Testing and sandboxing: Before giving this skill access in production, run it locally or in a sandbox, try saving known URLs (including those with query params), and inspect resulting markdown files to confirm no secrets are inadvertently stored. Also test the duplicate detection and delete/export commands. - If you need higher assurance: ask the author to remove the contradictory documentation lines (clarify triggers), add URL sanitization options (strip query tokens by default), and explicitly document any network endpoints or telemetry (none are present in the provided code).

knowledge-curator - 知识库管理技能

描述

自动抓取用户发送的链接内容（小红书、B 站、知乎、YouTube 等），整理成结构化知识库，支持按主题分类、增量更新和内容检索。

触发条件

收藏内容（必须满足以下条件之一）

• 用户明确请求保存内容到知识库（"收藏"、"保存到知识库"、"记下来"、"kb save"等）
• 消息中包含"收藏"、"收藏到知识库"、"kb"等关键词 + URL

注意：单纯发送链接不会触发收藏，必须有明确的收藏意图！

查询内容

• 用户查询知识库内容（"/kb search"、"/kb list"、"搜索..."、"查找..."等）

功能

1. 内容抓取

• 小红书: 提取标题、正文、图片描述、标签
• B 站: 提取视频标题、简介、字幕（如有）、标签、UP 主信息
• 知乎: 提取问题、回答、作者、赞同数
• YouTube: 提取视频标题、描述、字幕/转录、标签
• 通用网页: 提取标题、正文、元描述

2. 自动分类

按内容主题自动分类到以下类别：

• 科技: AI、编程、数码、互联网、科学
• 生活: 美食、旅行、家居、日常、情感
• 学习: 教育、课程、教程、知识、技能
• 娱乐: 影视、音乐、游戏、综艺、明星
• 工作: 职场、管理、效率、工具、商业
• 健康: 运动、饮食、医疗、心理、养生

3. 输出格式

每个知识库条目为 Markdown 文件，包含：

# [标题]

**原始链接**: [URL]
**来源平台**: [平台名称]
**收藏日期**: YYYY-MM-DD HH:mm
**分类**: [主题分类]
**标签**: #标签 1 #标签 2 #标签 3

## 摘要
[200-500 字内容摘要]

## 关键知识点
- 知识点 1
- 知识点 2
- 知识点 3

## 原文内容
[完整或精简的原文内容]

## 备注
[用户添加的备注或 AI 的补充说明]

4. 增量更新

• 检测重复链接，避免重复保存
• 支持同一主题下追加新内容
• 维护知识库索引文件 index.json

5. 查询检索

• 关键词搜索（标题、标签、内容）
• 按分类筛选
• 按时间范围筛选
• 语义相似度检索（可选）

使用方法

保存内容（需要明确指令）

正确示例：

用户：收藏 https://www.bilibili.com/video/BV1xx411c7mD
AI: 📚 已保存到知识库【学习/编程】类别
   标题：Python 入门教程
   标签：#Python #编程 #教程
   查看：`/kb search Python`

用户：保存到知识库：https://zhuanlan.zhihu.com/p/123456
AI: ✅ 已收藏到知识库

用户：kb save https://youtube.com/watch?v=xxx
AI: ✅ 已保存到知识库

错误示例（不会触发收藏）：

用户：https://www.bilibili.com/video/BV1xx411c7mD
AI: （仅正常回复链接内容，不收藏）

查询内容

用户：/kb search Python
AI: 🔍 找到 3 条相关内容：
   1. [学习/编程] Python 入门教程 (2026-03-15)
   2. [学习/编程] Python 数据分析实战 (2026-03-10)
   3. [工作/效率] 用 Python 自动化办公 (2026-03-08)

管理命令

• /kb list [category] - 列出知识库内容
• /kb search <keywords> - 搜索内容
• /kb delete <id> - 删除条目
• /kb export - 导出知识库
• /kb stats - 查看统计信息

文件结构

knowledge-curator/
├── SKILL.md              # 本文件
├── README.md             # 详细文档
├── scripts/
│   ├── fetch.js          # 内容抓取
│   ├── summarize.js      # 内容总结
│   ├── categorize.js     # 自动分类
│   ├── store.js          # 存储管理
│   └── query.js          # 查询检索
├── references/
│   └── examples.md       # 使用示例
└── knowledge-base/       # 知识库存储
    ├── 科技/
    ├── 生活/
    ├── 学习/
    ├── 娱乐/
    ├── 工作/
    ├── 健康/
    └── index.json        # 索引文件

依赖

• web_fetch 工具：抓取网页内容
• exec 工具：运行 Node.js 脚本
• 文件系统：存储知识库

注意事项

• 部分平台可能有反爬限制，需适当延迟请求
• 视频字幕提取依赖平台 API 或第三方服务
• 建议定期备份知识库目录
• 敏感内容需用户确认后再保存

版本

v1.0.0 - 初始版本