← Back to Skills Marketplace
huayang0704

Auto Tech Research

by huayang0704 · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ⚠ suspicious
78
Downloads
0
Stars
1
Active Installs
1
Versions
Install in OpenClaw
/install auto-tech-research
Description
全自动化技术深度调研 Skill。通过 Chrome DevTools Protocol 操控浏览器, 模拟真人搜索行为,在国内外主流平台检索内容,确保搜索结果与人工操作一致。 输出 HTML 格式的结构化调研报告。 核心原则: - 不使用 web_fetch,全程使用 browser(CDP 协议) - 搜索行...
README (SKILL.md)

自动化技术深度调研 Skill v4.0

核心变更(v4.0)

v3.0 → v4.0 最大变化:全面弃用 web_fetch,改用 browser(CDP 协议)

维度 v3.0 v4.0
搜索工具 web_fetch 优先,browser 降级 browser 唯一工具
搜索体验 程序式 HTTP 请求 模拟真人浏览器操作
JS 渲染 不支持 完整支持
登录态 不支持 支持(chrome-relay)
反爬绕过 经常 403 与真人一致,极少被拦
搜索结果 API 返回格式 与人工搜索完全一致

工作流程概览

输入技术主题(+ 可选的用户补充链接)
    ↓
[Phase 1] 关键词扩展(中英双语)
    ↓
[Phase 2] 平台相关性评估 + 动态数量分配
    ↓
[Phase 3] 浏览器搜索(CDP 操控,模拟真人)
    ├─ 启动浏览器(独立 openclaw 实例 或 chrome-relay)
    ├─ 逐平台搜索:打开搜索页 → 输入关键词 → 获取结果列表
    ├─ 逐条打开:点击链接 → 等待加载 → snapshot 提取内容
    ├─ 记录每个平台的获取状态和原因
    └─ 用户补充链接:直接 navigate → snapshot
    ↓
[Phase 4] 内容分级整理(L1-L4)
    ↓
[Phase 5] 生成 HTML 报告
    ├─ 技术概览(200-1000字)
    ├─ 分级资源(带平台标签和可点击链接)
    ├─ 平台获取诊断面板
    └─ 平台统计明细

浏览器策略

浏览器选择

场景 浏览器 Profile 说明
默认 openclaw(省略 profile) 独立受控浏览器,无登录态,适合公开内容
需要登录态 chrome-relay 用户已登录的 Chrome,适合知乎、B站等

启动顺序

1. 优先使用 openclaw 默认浏览器(独立、干净)
2. 如果平台需要登录(知乎文章、B站视频详情等),切换 chrome-relay
3. 每个平台搜索完成后,关闭标签页释放资源

Phase 1: 关键词扩展

与 v3.0 一致,中英双语扩展:

示例:Kubernetes
中文:
  核心: ["Kubernetes", "K8s", "容器编排"]
  教程: ["Kubernetes入门", "K8s教程"]
  深度: ["Kubernetes架构", "K8s原理"]
英文:
  core: ["Kubernetes", "K8s", "container orchestration"]
  tutorial: ["Kubernetes tutorial", "K8s getting started"]
  advanced: ["Kubernetes architecture", "K8s deep dive"]

Phase 2: 平台相关性评估

与 v3.0 一致,根据主题动态评估每个平台的相关性和获取数量。

Phase 3: 浏览器搜索(核心变更)

3.1 搜索引擎策略

目标平台 搜索方式 URL 模板
通用英文 Google 搜索 https://www.google.com/search?q={关键词}
通用中文 百度搜索 https://www.baidu.com/s?wd={关键词}
知乎 知乎站内搜索 https://www.zhihu.com/search?type=content&q={关键词}
CSDN CSDN 搜索 https://so.csdn.net/so/search?q={关键词}
B站 B站站内搜索 https://search.bilibili.com/all?keyword={关键词}
GitHub GitHub 搜索 https://github.com/search?q={关键词}&type=repositories
arXiv arXiv 搜索 https://arxiv.org/search/?query={关键词}
YouTube YouTube 搜索 https://www.youtube.com/results?search_query={关键词}
HackerNews HN 搜索 https://hn.algolia.com/?q={关键词}
微信公众号 搜狗微信搜索 https://weixin.sogou.com/weixin?query={关键词}
小宇宙 小宇宙搜索 https://www.xiaoyuzhoufm.com/search?q={关键词}

3.2 每个平台的搜索步骤

对于每个平台:
  1. browser(action="navigate", url=搜索URL)
  2. browser(action="snapshot") → 获取搜索结果列表
  3. 从 snapshot 中提取 Top N 个结果的标题和链接
  4. 对每个结果:
     a. browser(action="navigate", url=结果链接)
     b. browser(action="snapshot") → 获取文章内容
     c. 提取:标题、作者、发布时间、正文摘要
  5. 记录获取状态(成功条数、失败原因)
  6. 关闭多余标签页

3.3 关键操作示例

Google 搜索

browser(action="navigate", url="https://www.google.com/search?q=Kubernetes+tutorial")
browser(action="snapshot")  → 获取搜索结果
# 从 snapshot 中提取链接
browser(action="navigate", url="第一个结果链接")
browser(action="snapshot")  → 获取文章内容

知乎搜索(需要 chrome-relay):

browser(action="navigate", url="https://www.zhihu.com/search?type=content&q=Kubernetes",
        profile="chrome-relay")
browser(action="snapshot", profile="chrome-relay")  → 获取搜索结果
# 提取知乎文章链接(zhuanlan.zhihu.com/p/xxx)
browser(action="navigate", url="https://zhuanlan.zhihu.com/p/xxx",
        profile="chrome-relay")
browser(action="snapshot", profile="chrome-relay")  → 获取文章内容

B站搜索

browser(action="navigate", url="https://search.bilibili.com/all?keyword=Kubernetes")
browser(action="snapshot")  → 获取视频列表(标题、播放量、UP主)
# 提取视频链接(bilibili.com/video/BVxxx)

3.4 内容提取策略

内容类型 提取方式 提取目标
文章 snapshot → 解析文本 标题、作者、正文前500字
视频 snapshot → 解析元数据 标题、UP主、播放量、简介
论文 snapshot → 解析摘要 标题、作者、Abstract
仓库 snapshot → 解析 README 名称、Stars、描述
讨论 snapshot → 解析帖子 标题、分数、评论数

3.5 获取诊断(每个平台必须输出)

📊 平台获取诊断
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✅ Google搜索    | 10条 | 独立浏览器,搜索结果正常
✅ 知乎          | 8条  | chrome-relay,站内搜索 + zhuanlan 直链
✅ GitHub        | 12条 | 独立浏览器,仓库搜索+API
✅ arXiv         | 5条  | 独立浏览器,搜索结果正常
✅ B站           | 6条  | 独立浏览器,视频列表正常
⚠️ CSDN          | 3条  | 独立浏览器,部分文章需VIP,跳过
❌ YouTube        | 0条  | 独立浏览器,地区限制无法加载
✅ HackerNews    | 4条  | 独立浏览器,Algolia搜索正常
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
总计: 48条 (7/8 平台成功)

3.6 用户补充链接处理

当用户发送额外链接时:

1. 识别 URL 的平台来源
2. browser(action="navigate", url=用户链接)
3. browser(action="snapshot") → 提取内容
4. 分类到对应 Level
5. 更新 HTML 报告(增量)

Phase 4: 内容分级

与 v3.0 一致:L1入门 / L2实践 / L3深度 / L4前沿

Phase 5: HTML 报告生成

与 v3.0 一致,使用 references/report-template.html 模板生成自包含 HTML。

性能优化

减少浏览器操作次数

  1. 批量提取:一次 snapshot 提取搜索页面的所有结果,而不是逐个点击
  2. 标签复用:在同一标签页中 navigate 不同页面,减少标签开关
  3. 跳过低质量:搜索结果中明显低质量的跳过不打开
  4. 并行标签:如果浏览器支持,可同时打开多个标签

超时和错误处理

- 页面加载超时(>15s)→ 标记失败,继续下一个
- 页面内容为空 → 尝试等待 2s 后重新 snapshot
- 弹窗/登录拦截 → 尝试关闭弹窗,或切换 chrome-relay
- 验证码 → 标记为"需人工干预",跳过

与 v3.0 的完整对比

维度 v3.0 v4.0
搜索工具 web_fetch 优先 browser(CDP)唯一
搜索体验 HTTP 请求 模拟真人浏览器
JS 渲染
知乎 403 失败 ✅ chrome-relay 搜索
B站 空内容 ✅ JS 渲染后提取
CSDN 404 频繁 ✅ 浏览器正常加载
YouTube 超时 ⚠️ 可能地区限制
搜索结果一致性 低(API格式) 高(与人工一致)
输出格式 HTML HTML(不变)
失败诊断 ✅(不变)

注意事项

  1. 浏览器资源管理:搜索完一个平台后关闭多余标签页,避免内存泄漏
  2. 搜索频率控制:不要短时间内对同一平台发起过多请求,间隔 2-3 秒
  3. 内容提取精度:snapshot 返回的是 accessibility tree,需要从中筛选有效内容
  4. cookie/登录态:需要登录的平台使用 chrome-relay,公开内容用独立浏览器
  5. 隐私保护:不在独立浏览器中输入任何账号密码

版本:v4.0.0 更新日期:2026-03-27 核心变更:全面弃用 web_fetch,改用 browser CDP 协议模拟真人搜索

Usage Guidance
Before installing or running this skill: (1) Review scripts/auto-research.py fully — any included Python can run network requests or exfiltrate data. (2) Verify how the agent will access 'chrome-relay' and whether you are comfortable allowing snapshots of pages accessed while logged into your accounts (this can expose cookies, messages, private profiles). (3) If you only need public-web scraping, avoid enabling chrome-relay or run the skill in an isolated/browser profile with no logins. (4) Ask the publisher to explain why a runnable script is included despite 'instruction-only' metadata, and request explicit documentation of data flows (what is stored, where reports are written, and whether any external endpoints are contacted). (5) If you will run the Python script locally, do so in a sandbox or VM and inspect its network activity first.
Capability Analysis
Type: OpenClaw Skill Name: auto-tech-research Version: 1.0.0 The skill bundle is a legitimate framework for automated technical research across multiple platforms. The primary Python script (scripts/auto-research.py) is a simulation that generates structured Markdown reports using mock data and relevance-based logic. The SKILL.md file provides functional instructions for an AI agent to perform browser-based searches using Chrome DevTools Protocol (CDP), including the use of 'chrome-relay' for accessing content on platforms requiring login. No evidence of malicious intent, data exfiltration, or harmful prompt injection was found.
Capability Assessment
Purpose & Capability
The skill claims to perform CDP/browser-driven research and the SKILL.md instructions match that purpose (navigate, snapshot, extract, generate HTML). However the package includes a large Python script (scripts/auto-research.py) and README usage examples that run it, while the registry metadata describes this as instruction-only with no install spec — this mismatch (embedded runnable code but no declared install/run guidance) should be clarified.
Instruction Scope
The runtime instructions explicitly instruct using a 'chrome-relay' profile (the user's logged-in Chrome) for platforms that require login. That implies access to cookies, sessions, and any signed-in accounts. The SKILL.md does not document consent, limits, or which data is read from those profiles, and it instructs broad navigation/snapshotting across many third-party sites — a real privacy surface that should be declared and controlled.
Install Mechanism
No install spec (instruction-only) is lower risk, but the repo contains a ~27KB Python script and examples that run python3 scripts/auto-research.py. It's unclear whether the agent or user is expected to execute that script; the presence of executable code increases the attack surface and should be explained even if no installer is provided.
Credentials
The skill declares no required env vars or credentials, yet its behavior depends on access to browser profiles (openclaw and chrome-relay). Accessing chrome-relay effectively gives the skill access to session cookies and logged-in accounts (WeChat, Zhihu, Bilibili, Google, etc.). That level of access is significant and is not declared as a required credential/config entry.
Persistence & Privilege
always is false and there are no claims of permanent presence or modifying other skills/configs. The agent-autonomous invocation default applies but is not by itself a new concern in this package.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install auto-tech-research
  3. After installation, invoke the skill by name or use /auto-tech-research
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
auto-tech-research v4.0.0 - 全面弃用 web_fetch,所有自动化搜索均通过 browser(Chrome DevTools Protocol)模拟真人操作完成。 - 浏览器自动选择独立 profile 或 chrome-relay(带登录态)以应对不同平台需求,如知乎、B站等需登录内容。 - 支持 JS 渲染页面,搜索和内容获取与人工操作一致,极大提升反爬绕过能力和内容一致性。 - 各平台采集过程与结果透明:HTML 报告中输出获取成功/失败/原因等诊断信息。 - 优化浏览器资源管理和操作效率,减少无效页面打开与超时失败。
Metadata
Slug auto-tech-research
Version 1.0.0
License MIT-0
All-time Installs 1
Active Installs 1
Total Versions 1
Frequently Asked Questions

What is Auto Tech Research?

全自动化技术深度调研 Skill。通过 Chrome DevTools Protocol 操控浏览器, 模拟真人搜索行为,在国内外主流平台检索内容,确保搜索结果与人工操作一致。 输出 HTML 格式的结构化调研报告。 核心原则: - 不使用 web_fetch,全程使用 browser(CDP 协议) - 搜索行... It is an AI Agent Skill for Claude Code / OpenClaw, with 78 downloads so far.

How do I install Auto Tech Research?

Run "/install auto-tech-research" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Auto Tech Research free?

Yes, Auto Tech Research is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Auto Tech Research support?

Auto Tech Research is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Auto Tech Research?

It is built and maintained by huayang0704 (@huayang0704); the current version is v1.0.0.

💬 Comments