AI UNDERDOGSDAILY PICK
AI UNDERDOGS
让 Claude 驱动浏览器
Let Claude drive a browser
lackeyjb/playwright-skill
让 Claude 自己控制浏览器
但不想把整个上下文窗口都搭进去？
这条路终于跑通了
Want Claude to control a browser without torching
your entire context window?
This one actually works
授人以渔
Teach the model to fish
playwright-MCP 把 Playwright 几百个 API 全注册成工具
Claude 每回合都要啃一遍完整描述
token 烧得飞快
playwright-skill 换了个思路：给 Claude 一套轻量指令
让它自己写 Playwright 脚本
自己跑、自己读结果
playwright-MCP registers every Playwright method as a tool
— Claude burns tokens parsing the full API
surface every single turn. playwright-skill flips the model
give Claude a lightweight instruction set, let it
write its own Playwright scripts, run them, and
read the results
★ SIGNAL 1
架构上就赢了
Won at the architecture level
这不是偷懒，是设计上的品味
MCP 的方式是把所有鱼都摆出来
Claude 每次都要挑
这个 Skill 的方式是教它怎么钓鱼
它自己写脚本去执行
上下文占用直接降了一个量级
自主性反而更强了
This isn't laziness — it's taste. MCP lays
out every fish, Claude picks one each time
This Skill teaches the pattern: Claude writes scripts
runs them, iterates. Context drops by an order
of magnitude, autonomy goes up
★ SIGNAL 2
2700 星不是虚的
2.7k stars — the pain is real
一个 Claude Code 的 Skill 拿到 2700 多星
189 个 fork
不是因为它演示效果炫
是因为开发者在 MCP 路线上真的被上下文开销搞怕了
它遵循 Agent Skills 规范
Claude 写脚本、执行
迭代——不是玩具 demo
是可以直接塞进工作流的东西
A Claude Code Skill pulling 2.7k stars and
189 forks — not because it's a slick
demo, but because devs are genuinely hurting from
context overhead on the MCP path. It follows
the Agent Skills spec properly: Claude writes, executes
iterates. Not a toy demo, something you'd actually
slot into your workflow
作者的原话就六个字：比 playwright-MCP 少用上下文
没吹什么革命性、什么极致体验
就是少用上下文
能做到这一件事，就够了
The author's entire pitch: less context than playwright-MCP
No 'revolutionary', no 'next-gen' — just less context
If it does that one thing, it's enough
AI UNDERDOGS
不用烧掉整个上下文窗口
Without burning your context window
lackeyjb/playwright-skill
关注 · 每天发现更多 AI 神作
github.com/lackeyjb/playwright-skill