Anthropic刚刚用Claude技能解决了提示混乱问题吗？
Claude刚刚推出了一些有趣的功能：代理技能。这些技能类似于“提示插件”，但以一种非常适合开发者的方式实现。
一个技能就是一个小文件夹：包含指令、示例，可能还有一个脚本。Claude只在相关时加载它，因此你不必每次都提供上下文。而且，由于你可以包含真实代码，输出变得可预测，而不是“基于感觉”的。
我一直在测试这些技能，用于搭建项目、执行团队规范、生成模板和清理数据。效果令人震惊。你可以在这里了解我测试的更多信息：https://www.shuttle.dev/blog/2025/12/02/claude-skills-complete-guide
这感觉像是AI辅助开发工作中缺失的一个基本功能。
还有其他人正在玩这些技能吗？
查看原文
Claude just shipped something interesting: Agent Skills. They’re like “prompt plugins,” but done in a very dev-friendly way.<p>A skill is a tiny folder: instructions, examples, maybe a script. Claude only loads it when relevant, so you don’t have to shovel context every time. And since you can include real code, the output becomes predictable, not “vibes-based.”<p>I’ve been testing them for scaffolding projects, enforcing team conventions, generating boilerplate and cleaning data. It’s shockingly effective. You can learn more about my testing here: https://www.shuttle.dev/blog/2025/12/02/claude-skills-complete-guide<p>Feels like a missing primitive for AI-assisted dev work.<p>Anyone else playing with these yet?