Imagine onboarding a brilliant new colleague who already has vast general knowledge but lacks the specific procedures, templates, and organizational quirks unique to your work. You would not rewrite their brain; you would hand them a well-organized folder of documentation, scripts, and examples.
This is precisely the philosophy behind Anthropic's Skills repository - a collection of modular instruction packages that transform Claude from a general-purpose AI assistant into a specialized expert for particular domains.
Released publicly in October 2025, this repository represents Anthropic's vision for how humans and AI agents can share procedural knowledge in composable, scalable ways.
The Challenge of Contextual Expertise
Large language models like Claude possess remarkable breadth of knowledge, trained on vast swaths of human writing. Yet real-world work demands something different: procedural expertise that cannot be distilled from general training data.
How does your organization structure its financial reports? What specific workflow does your team follow for code reviews? Which exact parameters must be used when generating documents in your company's brand style?
These details exist nowhere in public datasets. Previously, developers faced a frustrating choice: either cram extensive context into every prompt (expensive and unwieldy) or build entirely custom agents for each use case (time-consuming and fragile).
Skills offer a third path - modular capability packages that agents discover and load dynamically, much like a professional consulting their reference materials only when the situation demands.
Key Features
- Progressive Disclosure: Skills load context in three tiers - metadata at startup, full instructions when triggered, and additional references on demand aiming to keep the context window efficient.
- Simple Format: Just folders and markdown files with YAML frontmatter. No proprietary formats, compilation steps, or special runtimes required.
- Code Execution: Skills can bundle Python scripts that Claude executes without loading into context, enabling deterministic operations like PDF manipulation or data processing.
- Cross-Platform Support: Works across Claude.ai, Claude Code, the Claude Agent SDK, and the Claude API with consistent behavior.
- Extensible Architecture: Bundle scripts, references, and assets alongside instructions to create sophisticated capabilities that scale with your needs.
A Skill in Practice
At its simplest, a skill is a folder containing a SKILL.md file. This file begins with YAML frontmatter specifying a name and description - metadata that Claude reads at startup to understand when each skill should be triggered. The markdown body below contains the actual instructions Claude follows when the skill activates.
The elegance lies in progressive disclosure: Claude initially loads only the lightweight metadata for all installed skills, consuming minimal context. When a user request matches a skill's description, Claude then reads the full SKILL.md.
For complex skills, additional reference files can be bundled alongside, loaded only when Claude determines they are relevant. This architecture means skills can contain effectively very large context window without overwhelming the working memory.
---
name: my-skill-name
description: A clear description of what this skill does and when to use it
license: Complete terms in LICENSE.txt
---
# My Skill Name
[Add your instructions here that Claude will follow when this skill is active]
## Overview
Overview of the tasks this skill is for.
**Keywords**: one, two, three etc...
## Examples
- Example usage 1
- Example usage 2
## Guidelines
- Guideline 1
- Guideline 2 Why This Approach Resonates
What strikes me most about the Skills framework is its deliberate simplicity. In an ecosystem increasingly crowded with complex tooling and orchestration frameworks, Anthropic chose plain folders and markdown files.
There is no proprietary format, no compilation step, no runtime to install. A skill can be created in minutes with a text editor, version-controlled with Git, and shared as easily as any other code artifact.
This design philosophy acknowledges a fundamental truth about AI development: the bottleneck is rarely the technology itself but rather the transfer of human expertise into machine-usable form.
By making that transfer as frictionless as possible, Skills lower the barrier for domain experts - not just developers - to customize AI behavior. The repository itself serves as both documentation and inspiration, offering over a dozen example skills spanning creative work, enterprise communication, technical development, and document processing.
Inside the Skills Collection
The repository organizes its offerings into distinct categories that reveal the breadth of possibilities:
- Document Skills: The docx, pdf, pptx, and xlsx skills power Claude's native document creation capabilities on claude.ai. These are source-available (not open source) but demonstrate production-grade skill architecture.
- Creative Skills: The algorithmic-art skill transforms Claude into a generative artist, guiding it to create p5.js compositions with seeded randomness and parameter controls. The theme-factory produces consistent visual themes.
- Development Skills: The mcp-builder skill teaches Claude to create Model Context Protocol servers - essentially helping the AI build tools that extend its own capabilities. The webapp-testing skill enables automated Playwright-based testing of web applications.
- Enterprise Skills: The brand-guidelines and internal-comms skills demonstrate how organizations can encode their specific voice, templates, and workflows.
- Meta Skills: The skill-creator skill guides Claude through the process of building new skills - a capability that hints at self-improving AI systems.
Architecture and Design Patterns
The repository structure reveals thoughtful engineering decisions. The spec directory contains the official Agent Skills specification (version 1.0 as of October 2025), establishing a standard that other implementations can follow.
Each skill follows a consistent pattern: a root SKILL.md with optional subdirectories for scripts (executable code), references (documentation to be loaded as needed), and assets (templates, fonts, images used in outputs).
This separation matters because it determines how content interacts with Claude's context window. Scripts can be executed without loading their contents into context - perfect for deterministic operations like PDF manipulation or data processing. References are loaded selectively based on the task at hand. Assets remain untouched by the language model entirely, used only in the final outputs Claude produces.
The progressive disclosure principle deserves particular attention. When Claude starts up with skills installed, its system prompt contains only skill names and descriptions - perhaps 100 words per skill. When triggered, the full SKILL.md loads - typically under 5,000 words by convention. Additional bundled files load on demand.
This three-tier system means a single agent can have access to dozens of specialized capabilities without any one of them consuming precious context until actually needed. The specification is deliberately minimal, requiring only the name and description fields in YAML frontmatter, with optional license, allowed-tools, and metadata fields for extended functionality.
Real-World Applications
The practical applications span virtually every domain where specialized knowledge improves outcomes.
- Financial analysts can create skills encoding their institution's modeling conventions, regulatory requirements, and report formats.
- Legal teams can package contract review procedures, clause libraries, and jurisdiction-specific guidance.
- Software teams can encode their coding standards, review checklists, and deployment procedures.
- Marketing departments can capture brand voice, campaign templates, and channel-specific formatting rules.
The key insight is that skills work best when they capture procedural knowledge that is both stable enough to be worth encoding and specific enough that Claude would not know it from general training. Company-specific workflows fit this profile perfectly.
Anthropic has begun highlighting partner skills, starting with Notion's official skill package for helping Claude work effectively with Notion workspaces. This signals an emerging ecosystem where software vendors create skills optimized for their platforms, much like they currently create API integrations and documentation.
Community and Contribution
The repository welcomes community participation through GitHub's standard pull request workflow. Most skills carry an Apache 2.0 license, permitting modification and redistribution with attribution. The document skills use a source-available license - viewable for reference but with restrictions on use. For developers looking to create their own skills, the repository includes both a minimal template and the comprehensive skill-creator skill that guides the process interactively. The project maintains active development, with commits from Anthropic engineers adding new skills and refining existing ones through late 2025.
Usage and Licensing
The repository employs a dual licensing approach. Example skills (algorithmic-art, brand-guidelines, mcp-builder, webapp-testing, and others) use the Apache 2.0 license, providing broad permissions for commercial and non-commercial use with attribution requirements.
The document-processing skills (docx, pdf, pptx, xlsx) are source-available rather than open source - Anthropic shares them as reference implementations demonstrating production skill architecture, but with more restrictive terms.
The THIRD_PARTY_NOTICES.md file documents all bundled dependencies, including imageio, FFmpeg, Pillow, and numerous fonts under BSD, GPL, MIT-CMU, and SIL Open Font licenses. Users should review these notices when incorporating skills into their own projects.
Future Directions
Anthropic's engineering blog articulates an ambitious vision: skills that agents can create, edit, and evaluate on their own. This represents a significant step toward AI systems that improve themselves through experience - capturing successful approaches into reusable capabilities rather than starting fresh each session.
The integration with Model Context Protocol (MCP) opens additional possibilities, with skills potentially teaching Claude complex workflows involving multiple external tools.
The Claude Code Marketplace integration already allows skills to be discovered and installed from within the development environment, reducing friction for adoption. As the ecosystem matures, we can expect to see skill-sharing platforms, organizational skill libraries, and increasingly sophisticated capabilities encoded in this simple format.
About Anthropic
Anthropic is an AI safety company founded in 2021, known for developing the Claude family of AI assistants. The company focuses on building reliable, interpretable, and steerable AI systems. Their product lineup includes Claude for consumers and teams, Claude Code for developers, enterprise solutions, and the Claude API for building AI-powered applications. Headquartered in San Francisco, Anthropic has established itself as a leading voice in responsible AI development, publishing research on AI alignment and safety alongside its commercial offerings. The Skills initiative reflects their broader philosophy of making AI more useful through better human-AI collaboration rather than simply increasing raw capability.
Getting Started
The Anthropic Skills repository offers a compelling vision for AI customization that respects both the power of large language models and the irreplaceable value of human expertise. Whether you are a developer seeking to extend Claude's capabilities for a specific technical domain, an organization wanting to encode institutional knowledge, or simply curious about the future of human-AI collaboration, this repository deserves exploration.
Clone it, browse the example skills, try creating one of your own - the barrier to entry is intentionally low, and the potential applications are as varied as the work humans do. In an AI landscape often characterized by complexity and opacity, Skills stand out for their elegant simplicity: just folders, just markdown, just the instructions you would give a capable colleague.
Anthropic Skills: Teaching AI Agents How to Work Like Experts