Building

agents

with

Skills:

Equipping

agents

for

specialized

work

A lot has changed in the past year. MCP became the standard for agent connectivity with rapid adoption from industry leaders and the developer community. Claude Code launched as a general-purpose coding agent. And we launched the Claude Agent SDK, which now provides a production-ready agent out of the box.过去一年发生了很多变化。MCP 成为代理连接的标准，得到了行业领导者和开发者社区的迅速采用。Claude Code 作为通用编码代理推出。我们还推出了 Claude Agent SDK，现在它提供了一个开箱即用的生产级代理。

But as we've built and deployed these agents, we keep running into the same gap: agents have intelligence and capabilities, but not always the expertise to effectively tackle real work. This led us to create Agent Skills. Skills are organized collections of files that package domain expertise - workflows, best practices, scripts - in a format agents can access and apply. They turn a capable generalist into a knowledgeable specialist.但在构建和部署这些代理的过程中，我们不断遇到同样的差距：代理拥有智能和能力，但并非总是具备有效处理实际工作的专业知识。这促使我们创建了 Agent Skills。Skills 是组织好的文件集合，将领域专业知识——工作流程、最佳实践、脚本——打包成代理可以访问和应用的格式。它们将一个有能力的通才变成了一个知识渊博的专家。

In this post, we'll explain why we stopped building specialized agents and started building skills instead, and how this shift is changing how we think about extending agent capabilities.在这篇文章中，我们将解释为什么我们停止构建专门的代理，转而开始构建技能，以及这一转变如何改变我们对扩展代理能力的思考。

The new paradigm: code is all you need新范式：代码即一切

We used to think agents in different domains would look very different. A coding agent, a research agent, one for finance, one for marketing—each seemed to need its own tools and scaffolding. The industry initially embraced this model of domain-specific agents. But as models improved in intelligence and agent capabilities progressed, we converged on a different approach.我们过去认为不同领域的代理会看起来非常不同。编码代理、研究代理、金融代理、营销代理——每个似乎都需要自己的工具和框架。行业最初接受了这种领域特定代理的模式。但随着模型智能的提升和代理能力的进步，我们趋同于一种不同的方法。

We came to see code less as just a use case and more as an interface for agents to do almost any digital work. Claude Code is a coding agent, but also a general-purpose agent that happens to work through code.我们开始将代码不仅仅视为一个用例，而是作为代理执行几乎任何数字工作的接口。Claude Code 是一个编码代理，但也是一个通过代码工作的通用代理。

Consider working with Claude Code to generate a financial report. It can call APIs for research, store data in the filesystem, analyze it with Python, and synthesize insights. All of that happens through code. The scaffolding becomes as simple as bash and a filesystem.考虑使用 Claude Code 生成财务报告。它可以调用 API 进行研究，将数据存储在文件系统中，用 Python 进行分析，并综合见解。所有这些都通过代码完成。框架变得像 bash 和文件系统一样简单。

But general capability isn't the same as expertise. When we started using Claude Code for real work, a gap emerged.但通用能力并不等同于专业知识。当我们开始将 Claude Code 用于实际工作时，一个差距出现了。

The missing piece: domain expertise缺失的部分：领域专业知识

Who would you want filing your taxes: a math genius figuring it out from first principles, or an experienced tax professional who's filed thousands of returns? Most people would choose the tax professional. Not because they're smarter, but because they have the right expertise.你希望谁来帮你报税：一个从基本原理出发的数学天才，还是一个经验丰富、处理过数千份报税单的税务专家？大多数人会选择税务专家。不是因为他们更聪明，而是因为他们拥有正确的专业知识。

Agents today are like that math genius: brilliant at reasoning through novel situations, but often lacking the accumulated expertise of a seasoned professional. They can do amazing things with proper guidance. However, they're often missing important context, can't easily absorb your organization's expertise, and don't automatically learn from repeated tasks.今天的代理就像那个数学天才：在推理新情况方面非常出色，但往往缺乏经验丰富的专业人士积累的专业知识。在适当指导下，它们可以做出惊人的事情。然而，它们常常缺少重要的上下文，无法轻松吸收你组织的专业知识，也不会自动从重复任务中学习。

Skills bridge this gap by packaging domain expertise in a format that agents can progressively access and apply.技能通过将领域专业知识打包成代理可以逐步访问和应用的格式，弥合了这一差距。

What are Agent Skills?什么是 Agent Skills？

Skills package domain expertise and procedural knowledge for agents. 技能为代理打包领域专业知识和程序性知识。

anthropic_brand/
├── SKILL.md
├── docs.md
├── slide-decks.md
└── apply_template.py

The simplicity of skills is deliberate. Files are a universal primitive that works with what you already have. You can version them with Git, store them in Google Drive, and share them with your team. This simplicity also means skill creation isn't limited to engineers. Product managers, analysts, and domain experts are already building skills to codify their workflows.技能的简单性是故意的。文件是一种通用的基本元素，可以与已有的东西配合使用。你可以用 Git 进行版本控制，存储在 Google Drive 中，并与团队共享。这种简单性也意味着技能创建不仅限于工程师。产品经理、分析师和领域专家已经在构建技能来编码他们的工作流程。

Progressive disclosure渐进式披露

Skills can contain extensive information. To protect the context window and make skills composable, they use progressive disclosure: at runtime, only the metadata (name and description from the YAML frontmatter) is shown to the model.技能可以包含大量信息。为了保护上下文窗口并使技能可组合，它们使用渐进式披露：在运行时，只有元数据（来自 YAML 前置数据的名称和描述）会显示给模型。

---
name: Anthropic Brand Style Guidelines
description: Anthropic's official brand colors and typography…
---

If Claude determines a skill is needed, it reads the full SKILL.md file. For additional detail, skills can include a references/ directory with supporting documentation loaded only on demand.如果 Claude 确定需要某个技能，它会读取完整的 SKILL.md 文件。如需更多细节，技能可以包含一个 references/ 目录，其中包含按需加载的支持文档。

This three-tier approach means you can equip an agent with hundreds of skills without overwhelming its context window—metadata uses ~50 tokens, full SKILL.md files ~500 tokens, and reference files 2,000+ tokens and only when specifically needed.这种三层方法意味着你可以为代理配备数百个技能，而不会压倒其上下文窗口——元数据使用约 50 个 token，完整的 SKILL.md 文件约 500 个 token，参考文件 2000+ 个 token，且仅在需要时加载。

Skills can include scripts as tools技能可以包含脚本作为工具

Traditional tools have problems: some have poorly written instructions, the model can't always modify or extend them, and they often bloat the context window. Code, on the other hand, is self-documenting, modifiable, and doesn't need to be in context at all times.传统工具有问题：有些指令写得不好，模型不能总是修改或扩展它们，而且它们常常使上下文窗口膨胀。另一方面，代码是自文档化的、可修改的，并且不需要始终在上下文中。

Here's a real example: we kept seeing Claude write the same script to apply Anthropic styling to slides. So we asked Claude to save it as a tool for itself:这是一个真实的例子：我们不断看到 Claude 编写相同的脚本，将 Anthropic 样式应用于幻灯片。所以我们让 Claude 将其保存为自身的工具：

# anthropic/brand_styling/apply_template.py
import sys
from pptx import Presentation

if len(sys.argv) != 2:
    print("USAGE: apply_template.py <pptx>")
    sys.exit(1)

prs = Presentation(sys.argv[1])
for slide in prs.slides:
    ...

The corresponding documentation in slide-decks.md simply references this script:slide-decks.md 中的相应文档简单地引用了这个脚本：

## Anthropic Slide Decks
- Intro/outro slides
  - background color: `#141413`
  - foreground color: oat
- Section slides:
  - background color: `#da7857`
  - foreground color: `#141413`

Use the `./apply_template.py` script to update a pptx file in-place.

The skills ecosystem技能生态系统

The skills ecosystem has emerged quickly, and so far we've seen three major types of skills being built:技能生态系统迅速出现，到目前为止，我们看到正在构建三种主要类型的技能：

Foundational skills基础技能

These provide core capabilities everyone needs: working with documents, spreadsheets, presentations, etc. They encode best practices for document generation and manipulation. You can see what this looks like in practice by exploring the foundational skills in our public repository.这些提供了每个人都需要的核心能力：处理文档、电子表格、演示文稿等。它们编码了文档生成和操作的最佳实践。你可以通过探索我们公共仓库中的基础技能来了解实际应用。

Partner skills合作伙伴技能

As skills standardize how agents interact with specialized capabilities, companies are building skills to make their services agent-accessible. K-Dense, Browserbase, Notion, and many others are creating skills that integrate their services directly, extending Claude's capabilities in specific domains while maintaining the simplicity of the skills format.随着技能标准化代理与专业能力的交互方式，公司正在构建技能，使其服务对代理可访问。K-Dense、Browserbase、Notion 等许多公司正在创建直接集成其服务的技能，在保持技能格式简单的同时，扩展 Claude 在特定领域的能力。

Enterprise skills企业技能

Organizations build proprietary skills encoding their internal processes and domain expertise. Skills help capture the specific workflows, compliance requirements, and institutional knowledge that make an agent useful for enterprise work.组织构建专有技能，编码其内部流程和领域专业知识。技能有助于捕获使代理对企业工作有用的特定工作流程、合规要求和机构知识。

Trends we see我们看到的趋势

As skills adoption grows, several patterns are emerging that point to where this paradigm may be heading. These trends shape how we think about skill design and the tooling we're building to support skill developers.随着技能采用的增长，出现了一些模式，指向这一范式可能的发展方向。这些趋势塑造了我们对技能设计的思考，以及我们正在构建的支持技能开发者的工具。

Increasing complexity复杂性增加

Early skills were simple documentation references. Now we're seeing sophisticated multi-step workflows that coordinate data retrieval, complex calculations, and formatted output across multiple tools.早期的技能是简单的文档参考。现在我们看到复杂的多步骤工作流程，协调数据检索、复杂计算和跨多个工具的格式化输出。

• Simple: "Status report writer" (~100 lines) - Templates and formatting简单：“状态报告编写器”（约 100 行）——模板和格式化
• Intermediate: "Financial model builder" (~800 lines) - Data retrieval, Excel modeling with Python中级：“财务模型构建器”（约 800 行）——数据检索、使用 Python 进行 Excel 建模
• Complex: "RNA sequencing pipeline" (2,500+ lines) - Coordinates HISAT2, StringTie, DESeq2 analysis复杂：“RNA 测序流程”（2500+ 行）——协调 HISAT2、StringTie、DESeq2 分析

Skills and MCP技能与 MCP

Skills and MCP servers work together naturally. A competitive analysis skill might coordinate web search, internal databases via MCP, Slack message history, and Notion pages to synthesize a comprehensive report.技能和 MCP 服务器自然地协同工作。一个竞争分析技能可能协调网络搜索、通过 MCP 的内部数据库、Slack 消息历史和 Notion 页面，以综合一份全面的报告。

Non-developer adoption非开发者采用

Skill creation is expanding beyond engineers to product managers, analysts, and domain experts across disciplines. They can create and test their first skill in under 30 minutes using the skill-creator tool, which guides them through the process interactively. We're working  to make skill creation even more accessible, with improved tooling and templates that let anyone capture and share expertise.技能创建正在从工程师扩展到产品经理、分析师和跨学科的领域专家。他们可以使用技能创建工具在 30 分钟内创建和测试第一个技能，该工具通过交互式方式引导他们完成过程。我们正在努力使技能创建更加易于访问，改进工具和模板，让任何人都能捕获和分享专业知识。

The complete architecture完整架构

Putting it all together, the emerging agent architecture looks like a combination of:综合来看，新兴的代理架构看起来像是以下组合：

1. Agent loop: The core reasoning system that decides what to do next代理循环：决定下一步做什么的核心推理系统
2. Agent runtime: Execution environment (code, filesystem)代理运行时：执行环境（代码、文件系统）
3. MCP servers: Connections to external tools and data sourcesMCP 服务器：与外部工具和数据源的连接
4. Skills library: Domain expertise and procedural knowledge技能库：领域专业知识和程序性知识

Each layer has a clear purpose: the loop reasons, the runtime executes, MCP connects, and skills guide. This separation makes the system comprehensible and allows each piece to evolve independently.每一层都有明确的目的：循环推理，运行时执行，MCP 连接，技能指导。这种分离使系统易于理解，并允许每个部分独立发展。

Consider what happens when you add a single skill to this architecture. The frontend design skill transforms Claude's frontend capabilities instantly. It provides specialized guidance on typography, color theory, and animation, activating only when building web interfaces. Progressive disclosure means it loads only when relevant. Adding new capabilities is straightforward.考虑当你向这个架构添加一个技能时会发生什么。前端设计技能立即改变了 Claude 的前端能力。它提供关于排版、色彩理论和动画的专业指导，仅在构建 Web 界面时激活。渐进式披露意味着它只在相关时加载。添加新功能很简单。

Deploying skills to new verticals将技能部署到新领域

This emerging pattern of general agents equipped with MCP servers and skills is already helping us deploy Claude to new verticals.这种配备 MCP 服务器和技能的通用代理的新兴模式已经帮助我们向新领域部署 Claude。

Financial Services金融服务

Just after launching skills, we enhanced Claude for the financial services sector with skills that make Claude more useful for finance professionals:在推出技能后不久，我们通过技能增强了 Claude 在金融服务领域的应用，使 Claude 对金融专业人士更有用：

• DCF model builder: Constructs discounted cash flow models with proper WACC calculations and sensitivity analysisDCF 模型构建器：构建贴现现金流模型，包含正确的 WACC 计算和敏感性分析
• Comparable company analysis: Generates comps tables with relevant multiples and benchmarking可比公司分析：生成包含相关倍数和基准测试的可比公司表
• Earnings analysis: Processes quarterly results and creates investment update reports收益分析：处理季度结果并创建投资更新报告
• Initiation coverage: Builds comprehensive research reports with financial models启动覆盖：构建包含财务模型的综合研究报告
• Due diligence: Structures M&A analysis with standardized frameworks尽职调查：使用标准化框架构建并购分析
• Pitch materials: Creates client presentations following industry standards推介材料：按照行业标准创建客户演示文稿

Healthcare & Life Sciences医疗保健与生命科学

We've also enhanced our healthcare and life sciences offerings with skills that make Claude more useful for researchers, clinicians, and healthcare developers:我们还通过技能增强了医疗保健和生命科学产品，使 Claude 对研究人员、临床医生和医疗保健开发者更有用：

• Bioinformatics bundles: Skills for scVI-tools and Nextflow deployments, essential for managing genomic pipelines and single-cell RNA sequencing生物信息学包：用于 scVI-tools 和 Nextflow 部署的技能，对管理基因组流程和单细胞 RNA 测序至关重要
• Clinical trial protocol generation: Accelerates protocol development for clinical research临床试验方案生成：加速临床研究的方案开发
• Scientific problem selection: Helps researchers identify and frame impactful research questions科学问题选择：帮助研究人员识别和构建有影响力的研究问题
• FHIR development: Helps developers write more accurate code for health data interoperability, connecting healthcare systems faster with fewer errorsFHIR 开发：帮助开发者编写更准确的健康数据互操作性代码，更快地连接医疗系统，减少错误
• Prior authorization review: Cuts administrative burden and accelerates patient access to needed care by cross-referencing coverage requirements, clinical guidelines, and patient records事先授权审查：通过交叉引用覆盖要求、临床指南和患者记录，减少行政负担，加速患者获得所需护理

Standardizing Agent Skills标准化 Agent Skills

To enable this vision, we're publishing Agent Skills as an open standard. Like MCP, we believe skills should be portable across tools and platforms. The same skill should work whether you're using Claude or other AI platforms. We've been collaborating with members of the ecosystem on the standard, and we're excited to see early adoption.为了实现这一愿景，我们将 Agent Skills 作为开放标准发布。像 MCP 一样，我们相信技能应该跨工具和平台可移植。无论你使用 Claude 还是其他 AI 平台，同一个技能都应该有效。我们一直在与生态系统成员合作制定标准，并很高兴看到早期采用。

When someone starts using an AI agent for the first time, it should already know what you and your team care about because skills capture and transfer that expertise. As this ecosystem grows, a skill built by someone else in the community can make your agent more useful, reliable, and capable - regardless of which AI platform they're using.当有人第一次开始使用 AI 代理时，它应该已经知道你和你的团队关心什么，因为技能捕获并传递了这些专业知识。随着这个生态系统的发展，社区中其他人构建的技能可以让你的代理更有用、更可靠、更有能力——无论他们使用哪个 AI 平台。

Getting started开始使用

We're converging on an architecture for general agents, and skills provide a paradigm for shipping and sharing new capabilities. The real value emerges from the collective knowledge base we build together: capturing expertise, transferring it across teams, and making every agent more capable than the last.我们正在趋同于通用代理的架构，技能提供了一种交付和共享新能力的范式。真正的价值来自于我们共同构建的集体知识库：捕获专业知识，跨团队传递，并使每个代理都比上一个更有能力。

Resources:资源：

• Don’t Build Agents, Build Skills Instead (YouTube Video)不要构建代理，而是构建技能（YouTube 视频）
• Skills documentation技能文档
• GitHub repositoryGitHub 仓库
• Skills cookbook技能食谱
• Using skills in Claude在 Claude 中使用技能
• Skills API quickstart技能 API 快速入门
• Skills best practices documentation技能最佳实践文档

Acknowledgments:致谢：

Barry Zhang, Mahesh Murag, Keith Lazuka, Ryan WhiteheadBarry Zhang、Mahesh Murag、Keith Lazuka、Ryan Whitehead