Agent Skills 课程介绍

摘要

本课程介绍 Agent Skills（智能体技能）的概念与应用。技能是一组指令文件夹，通过提供专业知识来扩展智能体的能力，现已作为开放标准可在多个智能体产品中复用。课程将教授技能的工作原理、最佳实践以及在不同场景（编码、研究、数据分析）下的构建方法。

要点

• 技能定义：包含 SKILL.md 的文件夹，提供专业知识以扩展智能体能力
• 开放标准：标准化格式使其兼容多个智能体产品，一次构建即可跨平台部署
• 渐进式披露：技能名称和描述常驻上下文，完整指令仅在匹配用户请求时加载
• 工具要求：智能体需具备文件系统访问权限和 bash 工具来执行技能命令
• 集成能力：可与 MCP 和子智能体结合，创建强大的智能体工作流

视频信息：Get Introduced to Agent Skills and This Course (Introduction)

中文翻译

欢迎来到这门关于 Agent Skills（智能体技能）的课程，本课程由 DeepLearning.AI 与 Anthropic 合作构建，并由返场讲师 Elie Schoppik 讲授。技能赋予了 Claude 和其他智能体执行任务的新能力。我非常高兴 Elie 能回来教授这门课。谢谢你，Andrew。很高兴能回来和大家一起探讨这个话题。

技能是一组指令文件夹，通过提供专业知识来扩展智能体的能力。在本课程中，你将学习技能是如何工作的，了解创建技能的最佳实践，并为不同的用例构建技能，包括编码、研究、数据分析等。

技能最令人兴奋的一点是它们现在成为了一个开放标准，这意味着它们拥有标准化的格式，适用于任何兼容技能的智能体。因此，你只需构建一次技能，就可以将其部署到多个智能体产品中。任何技能都应包含一个 SKILL.md Markdown 文件，其中包含技能的名称、描述和主要指令。主要指令还可以引用其他文件，如脚本、额外的 Markdown 文件以及模板和图像等资源。

技能是逐步向智能体披露的，这意味着技能的名称和描述始终存在于智能体的上下文窗口中，但直到用户的请求与技能的描述相匹配时，智能体才会将剩余的指令加载到其上下文中。届时，如果有需要，智能体可能还会额外加载参考文件和资源文件。

要使用这种技能，你的智能体需要一套基本的工具：用于读写文件的文件系统访问权限，以及用于执行代码的 bash 工具。这些工具使你的智能体能够执行技能所需的任何命令。你的智能体可以将技能与 MCP（模型上下文协议）和子智能体结合起来，创建强大的智能体工作流。例如，它可以利用 MCP 从外部来源获取数据，然后依靠技能知道如何处理该数据或如何高效地检索数据。它还可以将任务委派给具有隔离上下文的子智能体，子智能体本身也可以利用技能来获取专业知识。

在本课程中，我们将从 Claude AI 开始，创建一个用于营销活动的技能，并将其与 Excel 和 PowerPoint 的预构建技能相结合。然后，我们将为内容创作和数据分析工作流创建两个技能，并在 Claude API 中进行尝试。之后，我们将结合 Claude Code 使用技能来审查和测试代码。最后，我们将使用 Claude Agent SDK 构建一个研究智能体，利用技能来整合研究结果。

我要感谢来自 DeepLearning.AI 的 Hawraa Salami 对本课程的贡献。那么，你如何知道何时该使用技能呢？假设你有一个通过反复要求智能体去执行的工作流。与其每次都解释相同的工作流，不如将其打包成一项技能，这样你的智能体就会自动知道该怎么做。这正是你在第一节课中将随 Elie 一起学到的内容。所以，请继续观看下一个视频以了解更多信息。

English Script

Welcome to this course on agent skills, built in partnership with Anthropic and taught by returning instructor Elie Schoppik. Skills give Claude and other agents new abilities to carry out tasks. I’m thrilled to have Elie return to teach this. Thank you, Andrew. I’m happy to be back and work with you all on this one.

Skills are folders of instructions that extend your agent’s capabilities with specialized knowledge. In this course, you’ll learn how skills work, learn best practices for creating them, and build skills for different use cases including coding and research and data analysis and more.

What’s exciting about skills is they’re now an open standard, which means they have a standardized format that work with any skills compatible agent. So you can build your skills once and deploy them across multiple agent products. Any skill should include a SKILL.md markdown file, which contains the skill’s name, description, and main instructions. The main instructions can also refer to other files such as scripts, additional markdown files, and assets such as templates and images.

Skills are progressively disclosed to the agent, which means that the skill’s name and description always live in your agent’s context window, but the agent does not load the rest of the instructions into its context until a user request matches the skill’s description. At that point, the agent might then additionally load the reference and asset files if needed as well.

To use this skill, your agent needs a basic set of tools, filesystem access to read and write files and a bash tool to execute code. And these tools enable your agent to execute whatever command a skill requires. Your agent can combine skills with MCP and sub-agents to create powerful, agentic workflows. For example, it can use MCP to get data from external sources, then rely on a skill to know what to do with that data or how to retrieve it efficiently. It can also delegate tasks to a sub-agent with isolated context, which can itself use skills for specialized knowledge.

In this course, we’ll start with Claude AI, where we’ll create a skill for a marketing campaign and combine it with the pre-built skills for Excel and PowerPoint. Then, we’ll create two skills for content creation and data analysis workflows and try them with the Claude API. After that, we’ll use skills with Claude code for reviewing and testing code. And finally, we’ll build a research agent with the Claude agent SDK that uses a skill to combine research results.

I’d like to thank Hawraa Salami from DeepLearning.AI who contributed to this course. So, how do you know when to use a skill? Let’s say you have a workflow that you repeatedly ask your agent to implement. Instead of explaining the same workflow every time, you can package it as a skill so your agent automatically knows what to do. That’s exactly what you’ll learn with Elie in the first lesson. So, please go on to the next video to learn more.