Reusable AI Agent skills for Runtime-aware web interaction — adapt websites, build site profiles, search & summarize content. Works with Copilot CLI, OpenClaw/Antigravity, Claude Code, Cursor.

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 Cursor 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 Cursor 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