As Agents grow more capable every day, our focus should shift from asking whether they can do something to how they can do it better. Recently, Anthropic’s release of Claude Skills offered a glimpse into that future.

Today, we take another step towards it by introducing Skillhub, an open source platform for discovering and sharing Agent Skills, built to make expert-level capabilities accessible to both humans and Agents.

General-purpose Agents are already powerful. They can write code, analyze documents, reason through complex problems, and assist with daily tasks. But as impressive as these capabilities are, they can often hit a wall when you need them most, because they lack the tools that make real-world work possible: procedural knowledge, organizational context, and domain expertise.

This is where the gap becomes clear: An Agent can create compelling marketing content, but without your company’s brand guidelines, the tone and style won’t align with your identity. It can write functional code, but without knowledge of your architectural patterns and conventions, that code won’t integrate seamlessly with your systems.

Each Skill is defined by a SKILL.md file containing Metadata that specifies what it does.

For advanced use cases, a Skill can include additional files that extend its functionality:

• Markdown documents such as reference.md or guidelines.md for domain-specific instructions.
• JSON or YAML configuration files for structured parameters or schemas.
• Python or JavaScript scripts that handle deterministic logic.
• Supporting assets like CSV datasets or templates that the Agent can process dynamically.

This modular structure makes Skills composable, portable, and easy to maintain.

One of the most persistent challenges in large language model (LLM) architecture is the context window. Every piece of information, including prompts, conversation history, retrieved data, and tool instructions, must fit within this finite resource.

Agent SDKs and MCP Servers often preload vast amounts of knowledge upfront, wasting tokens on information that might never be used.

Skills take a different approach: Progressive Disclosure. Their structure allows information to load in layers:

1. Metadata - Name and description load at startup, requiring only dozens of tokens. Agents know which skills are available.
2. Instructions - When a skill becomes relevant, the full SKILL.md loads with complete operational guidance.
3. Reference Materials - For complex tasks, additional files are fetched only when explicitly needed.

A skill that would consume thousands of tokens if loaded upfront costs only dozens until actually needed. At scale, hundreds of skills can remain available without drowning the context window.

This shifts our architecture from just in case to just in time.

But the benefits go beyond token efficiency. Progressive disclosure enables a clear separation of concerns between reasoning and execution.

LLMs excel at interpreting ambiguity and reasoning over complex patterns. But their non-deterministic nature can be a hurdle when you need precise, repeatable behavior. Skills close this gap. They can bundle strict workflows and executable code to achieve deterministic results during the execution phase. Once complete, the Agent receives verified outputs and continues reasoning from there.

We believe Skills represent a fundamental shift in how we empower Agents. The traditional approach has clear limitations: stuffing 200-page style guides into every prompt burns tokens and slows response times. Building separate Agents for each use case creates technical debt and integration headaches when workflows overlap.

A single well-designed Skill can transform what an Agent can do. Scale this to hundreds of Skills, each loaded only when needed, and you have a lean, powerful foundation.

Imagine your senior engineer creates a Skill for the team. That knowledge multiplies instantly. Everyone gains the expertise without the learning curve.

Skills can handle complex workflows like building full-stack applications, planning data pipelines, and generating documents. Their structured, workflow-based nature brings determinism where it wasn’t possible before. Agents follow defined procedures instead of hallucinating their way through tasks.

During our internal testing, we found something interesting: Skills appear static, but they constantly evolve. A coworker can use your Skill as-is, build upon it, or remix it and share a new version. Others add their improvements. Each iteration compounds.

A Skill is never the end - it’s the start of an ongoing chain of shared knowledge.

In the example above, Claude builds a full-stack chatbot in minutes using a single Skill. Once a working blueprint exists, any Agent can pick it up and extend it. The next Agent could add a sidebar, switch to a different model provider, or even port the app to another framework.

As Agents become more durable, running for hours or even weeks, efficient context management becomes essential. Skills give them a structured way to store what works, load it when relevant, and share it when useful. Each modification can become a version or remix, allowing improvements to build iteratively across Agents.

This is agent-to-agent collaboration in practice. A workflow that begins with one Skill and improves with every handoff. It’s how knowledge starts to scale across Agents.

Skillhub provides a central place to explore available Skills, complete with documentation, reference files, and example workflows. Each Skill page includes organized tabs for source files, dependencies, and configuration data.

Skills can also be added directly to Claude Code through the integrated Skillhub Marketplace:

Each Skill detail page includes a one-step copy command to install it directly in Claude Code. This makes it easy to try, adapt, and extend Skills without setup overhead.

To contribute, visit the open source project and contribution guide on GitHub.