skill-browser

Reusable AI Agent skills for Runtime-aware web interaction — works with GitHub Copilot CLI, OpenClaw/Antigravity, Claude Code, Cursor, and any agent that supports markdown-based skill files.

可复用的 AI Agent skills，面向 Runtime-aware 网页交互 —— 适用于 GitHub Copilot CLI、OpenClaw/Antigravity、Claude Code、Cursor 等所有支持 markdown skill 的 AI Agent。

Core capabilities | 核心能力：

Explore: adapt unfamiliar websites generically | 通用适配陌生网站
Stabilize: turn proven flows into reusable profiles | 把验证过的流程沉淀为可复用 profile
Execute: search, rank, and summarize across platforms using the best available runtime | 用最优执行路径做搜索、排序和摘要

Why this repo | 为什么做这个仓库

This repo packages a practical browser-workflow playbook into shareable Copilot skills.

这个仓库把一套实战验证过的浏览器工作流，整理成了可以分享、安装、复用的 Copilot skills。

Architecture model (2026) | 架构模型

Cognition Layer    → LLM (Planner + Reasoner)
Execution Layer    → opencli (CLI) / fetch (API) / Browser Agent (UI)
Memory Layer       → site profiles, platform registry
Environment        → Web

The key insight: the agent should pick the best execution path per platform, not default to browser scraping.

核心洞察：Agent 应该按平台选最优执行路径，而不是默认走浏览器抓取。

Execution routing | 执行路由

if opencli supports the platform:
    use opencli (deterministic, zero LLM cost, session-reuse)
elif public API exists:
    use fetch (structured JSON, no rendering needed)
elif site is SSR and publicly accessible:
    use web_fetch or Playwright (metadata extraction)
else:
    require login-state browser or declare blocked

Platform access tiers | 平台接入分级

Platform	opencli	Public API	SSR fetch	Tier	Recommended
Bilibili	✅ `bilibili search`	✅ `api.bilibili.com`	partial	T1	opencli
YouTube	✅ `youtube search`	✅ Data API	partial	T1	opencli
HackerNews	✅ `hackernews top`	✅ `hn.algolia.com`	✅ SSR	T1	opencli or fetch
Wikipedia	✅ `wikipedia search`	✅	✅ SSR	T1	any path works
Reddit	✅ `reddit search`	❌	partial	T1	opencli
小红书	✅ `xiaohongshu search`	❌	❌ SPA	T2	opencli only
抖音/TikTok	✅ `tiktok search`	❌	❌ SPA	T2	opencli only
知乎	✅ `zhihu search`	❌	❌ login	T2	opencli only
微博	✅ `weibo search`	❌	❌	T2	opencli only
豆瓣	✅ `douban search`	❌	partial	T2	opencli
Medium	✅ `medium search`	❌	❌	T2	opencli
arXiv	❌	✅ `export.arxiv.org`	✅ SSR	T3	fetch / web_fetch
BBC	✅ `bbc news`	❌	✅ SSR	T3	fetch / web_fetch
今日头条	❌	❌	✅ SSR	T3	web_fetch

Tier definitions | 分级定义：

T1 — multiple working paths; opencli preferred for consistency | 多条路径可用，opencli 优先
T2 — opencli required; if unavailable, declare blocked | opencli 必须，无公开 fallback
T3 — public fetch works reliably; opencli optional | 公开 fetch 可靠，opencli 可选

Layered skill workflow | 分层 skill 工作流

Instead of hardcoding every website one by one, the workflow follows a layered approach:

它的核心思路不是"每个网站都单独硬编码一遍"，而是三层方法：

adapt — explore and adapt generically first | 先做通用适配
profile — stabilize into a reusable profile only when necessary | 只有确实需要时才沉淀专用 profile
summary — execute search and summarization using the best available runtime | 用最优 runtime 执行搜索与摘要

Recommended names | 推荐优先记住的名称

If you only remember three public-safe English names, remember these:

如果你只先记住 3 个公开场景更稳妥的英文名，建议记这 3 个：

web-adapt
site-profile
content-summary

If you want the shortest local English aliases, use:

如果你想用最短的本地英文别名，可以用：

adapt
profile
summary

If you prefer Chinese aliases, use:

如果你更喜欢中文别名，可以用：

网站适配
站点配置
内容摘要

They map to the same three capability families:

它们本质上都映射到同样的三类能力：

adapt / web-adapt / 网站适配: generic website adaptation | 通用网站适配
profile / site-profile / 站点配置: stable external site profile workflow | 稳定站点 profile 沉淀
summary / content-summary / 内容摘要: content search, ranking, and summarization | 内容搜索、筛选与摘要

Included skills | 仓库包含的 skills

Shortest English aliases | 最短英文别名

adapt
profile
summary

Chinese aliases | 中文别名

网站适配
站点配置
内容摘要

Short public aliases | 推荐公开使用的英文短名

web-adapt
site-profile
content-summary

Full names kept for compatibility | 为兼容保留的完整名

universal-web-adaptation
external-site-profile-learning
content-search-summarization

What each skill does | 每个 skill 是做什么的

`web-adapt`

Use this skill when you want to make progress on an unfamiliar public website before writing any site-specific rules.

这个 skill 适合在还没有站点规则时，先用通用方法把网站跑通。

Best for:

适合场景：

detecting search inputs, buttons, openers, and popups
trying Enter, click, opener, or popup strategies
deciding whether a dedicated site profile is actually needed
找搜索框、按钮、弹窗、新标签页
尝试 Enter、click、opener、popup 等策略
判断这个网站是否值得沉淀成专用 profile

`site-profile`

Use this skill when a website already looks automatable and you want to turn it into a stable, repeatable profile.

这个 skill 适合把“已经能试出来”的网站，升级成“稳定可复用”的 site profile。

Best for:

适合场景：

investigating stable selectors
classifying failure modes
adding config-driven site profiles
validating heuristic and LLM-based planning flows
调查稳定 selector
分析失败模式
增加 config-driven site profile
验证 heuristic 与 LLM 规划链路

`content-summary`

Use this skill to search public content platforms, pick the top relevant results, and summarize each item with links and short takeaways.

这个 skill 适合在公开内容平台上搜索主题，筛选 Top 结果，并输出链接和摘要。

Best for:

适合场景：

Bilibili, YouTube, Xiaohongshu, and similar platforms
Top 5 / Top 10 result selection
opening detail pages to enrich metadata
writing Chinese or English summaries
B站、YouTube、小红书等内容平台
Top 5 / Top 10 结果筛选
打开详情页补充元数据
输出中英文摘要

Installation | 安装方式

One-line install (recommended) | 一键安装（推荐）

git clone https://github.com/AllenS0104/skill-browser.git
cd skill-browser
node install.js           # install 4 core skills (adapt, profile, summary, adapter-audit)

Options:

node install.js           # recommended set (4 shortest aliases)
node install.js --cn      # recommended + Chinese aliases
node install.js --all     # install all 13 aliases
node install.js --list    # list available skills
node install.js --clean   # uninstall all skill-browser skills

After install, reload in Copilot CLI:

安装后，在 Copilot CLI 中执行：

/skills reload
/skills list

Manual install | 手动安装

Copy any skill folder into your Copilot CLI skills directory:

也可以手动拷贝单个 skill 文件夹：

# Windows
%USERPROFILE%\.copilot\skills\<skill-name>\SKILL.md

# macOS / Linux
~/.copilot/skills/<skill-name>/SKILL.md

Example prompts | 示例提示词

You do not have to use an exact slash-style command every time.

不一定每次都必须写成精确的 /skill-name 命令。

Recommended invocation strength:

推荐触发强度：

explicit slash command such as /web-adapt — most deterministic
explicit skill-name mention such as use the web-adapt skill
natural-language task description that clearly matches the skill intent
显式 /web-adapt 这类命令 —— 最稳定
显式提到 skill 名，例如 use the web-adapt skill
只写自然语言任务，但任务意图要足够清晰

If you want reliable routing, prefer either the slash form or the explicit skill name.

如果你想让路由更稳定，优先使用 slash 形式或明确提到 skill 名。

`web-adapt`

使用 /web-adapt 这个 skill，先通用适配这个陌生网站。

Use the /web-adapt skill to adapt this unfamiliar website before creating any site-specific rules.

Please adapt this unfamiliar public website generically first, and tell me whether it really needs a dedicated site profile.

请用 /网站适配 先通用适配这个陌生网站，再告诉我是否真的需要专用 profile。

`site-profile`

使用 /site-profile 这个 skill，把这个网站沉淀成稳定的 site profile。

Use the /site-profile skill to turn this website into a stable external site profile.

Please turn this site into a stable reusable external profile, and validate both heuristic and Gemini planning flows.

请用 /站点配置 把这个已经跑通的网站沉淀成稳定可复用的 profile，并验证两条规划链路。

`content-summary`

使用 /content-summary 这个 skill，帮我从 B 站找 5 个关于 AI 的视频并总结。

Use the /content-summary skill to find the top 5 Bilibili videos about AI and summarize them.

Find the top 5 Bilibili videos about AI, rank them by relevance, and summarize each item with links and confidence notes.

请用 /内容摘要 在 B 站搜索 AI 相关内容，筛选 Top 5，并输出带链接和置信度的摘要。

Suggested workflow | 推荐使用顺序

If you are new to this repo, use the skills in this order:

如果你是第一次使用，建议按这个顺序：

web-adapt
site-profile
content-summary

This creates a practical loop:

这样最容易形成一条完整闭环：

first adapt the site
then stabilize the knowledge into a profile
finally reuse the workflow for search and summarization tasks
先让 AI 适应网站
再把经验沉淀成稳定规则
最后把流程复用到搜索与摘要任务

Who this is for | 适合谁使用

This repo is useful for:

这个仓库适合：

developers extending any AI agent with browser/web skills (Copilot CLI, OpenClaw, Claude Code, Cursor, etc.)
teams building browser automation workflows
people productizing AI-assisted web workflows
anyone who wants reusable website adaptation skills instead of one-off scripts
想给任意 AI Agent 扩展浏览器/网页能力的开发者（Copilot CLI、OpenClaw、Claude Code、Cursor 等）
在做浏览器自动化工作流的团队
想把 AI 网页工作流产品化的人
不想每次都临时写一次性脚本的人

Design principles | 设计原则

generic first, site-specific second
prefer reusable workflows over one-off hacks
keep profile logic config-driven when possible
validate with real websites, not only toy examples
通用优先，专用其次
优先沉淀可复用工作流，而不是一次性 hack
尽量让 profile 逻辑保持 config-driven
尽量用真实网站验证，而不只是 demo

Notes | 说明

These skills are agent-agnostic — any AI agent that reads markdown instruction files can use them directly.

这些 skills 是不绑定特定 Agent 的 —— 任何能读取 markdown 指令文件的 AI Agent 都可以直接使用。

Compatibility | 兼容性验证

Our skill files are plain markdown with YAML frontmatter, containing no executable code, no agent-specific API calls, and no framework dependencies. This makes them portable across agents.

我们的 skill 文件是纯 markdown + YAML frontmatter，不包含可执行代码、不调用任何 Agent 特定 API、不依赖任何框架。因此可以跨 Agent 移植。

Agent	How it loads instructions	How to use our skills	Verified
GitHub Copilot CLI	`~/.copilot/skills/<name>/SKILL.md`	`node install.js` (auto-deploy)	✅ tested
OpenClaw / Antigravity	`SKILL.md` via AGENTS.md → TOOLS.md skill reference	Copy skill folder to workspace, reference in AGENTS.md	✅ tested
Claude Code	`CLAUDE.md` + `.claude/rules/*.md` + `@import` syntax	`@import` our SKILL.md, or paste content into `.claude/rules/browser.md`	✅ format compatible
Cursor	`.cursorrules` or `.cursor/rules/*.md` (project rules)	Copy SKILL.md content into `.cursor/rules/browser-skill.md`	✅ format compatible
Other LLM agents	System prompt / context injection	Paste SKILL.md content into agent's system prompt or instruction context	✅ plain markdown

Why it works everywhere: our SKILL.md files are pure instructional text — they teach the agent how to think about web adaptation, site profiling, and content summarization. They don't call any agent-specific runtime. Any agent that can read markdown context can follow these instructions.

为什么到处能用：我们的 SKILL.md 是纯粹的教学文本 —— 教 Agent 如何思考网站适配、站点 profile 沉淀和内容搜索摘要。不调用任何 Agent 专属运行时。任何能读 markdown 上下文的 Agent 都能遵循这些指令。

Contributing | 参与贡献

If you want to improve the repo, add a new skill, or propose a better alias, see CONTRIBUTING.md.

如果你想继续完善仓库、补充新 skill，或者提出更好记的别名，可以先看 CONTRIBUTING.md。

Acknowledgments | 致谢

This project is built on top of and deeply inspired by the following open-source work:

本项目建立在以下开源项目之上，并深受其启发：

opencli by @jackwener — the Runtime-aware execution routing in our skills relies heavily on opencli's deterministic CLI adapters, Browser Bridge session reuse, and dual-engine (YAML + TS) architecture. opencli is the reason our content-summary skill can reliably access 50+ platforms including Bilibili, 小红书, TikTok, 知乎, and many more. We are also proud contributors to the opencli project (PR #404, PR #414).
GitHub Copilot CLI — the skill system that makes this entire repo possible.

Thank you to all open-source contributors who make AI-powered workflows better for everyone. 🙏

感谢所有让 AI 工作流变得更好的开源贡献者。🙏