Putting the real world on the line.

Get Developer Communications

Product upgrading, operating methods, community focus, etc. Send it to your inbox every month.

Update: We have issued proxy skills as an open standard for cross-platform portability. (18 December 2025)

As models improve, we can now build universal agents that interact with mature computing environments. For example:Claude. Codes can perform complex cross-domain tasks using local codes and file systems. But as these agents become stronger, we need more groupable, scalable and transplantable ways to equip them with specific areas of expertise.

This has led us to create proxy skills: organizational folders of instructions, scripts and resources that agents can dynamically identify and load to better perform specific tasks. The skill expands Claude 's ability by transforming the generic agent into a professional agent appropriate to your needs by packaging your expertise into a combustible resource for Claude.

The skills of an agent are developed as if an entry guide had been prepared for new staff. Now, anyone can specialize in agents with portfolio functions by capturing and sharing their programme knowledge, rather than building decentralized, customized agents for each example. In this paper, we will explain what skills are, demonstrate how they work and share the best practices in building your own skills.

To understand the practical application of skills, let's see a real example: Claude's recently introduced skills in document editing. Claude has learned a lot about understanding PDF, but has limited capacity to directly operate PDF (e.g. to fill out a form). This PDF skill allows us to give Claude these new capabilities.

In a simple way, skills is a directory containing SKILL.md files

I don't know. The document must be YAML Frontmatter starts with some necessary metadata: name

and description

I don't know. Agent preloads name on startup

and description

Adds each installed skill to its system tip.

This metadata is the first level of progressive disclosure: It provided Claude with sufficient information on when each skill should be used without having to load all skills into the context. The actual subject of this document is the second level of detail. If Claude thinks the skill is relevant to the current mission, it will load it by reading the full SKILL.md

Enter the context.

As skills become more complex, they may have too many contexts to fit into a single SKILL.md

, or context related only to a particular scene. In these cases, skills can bind other documents in the skills catalogue and quote them by name in SKILL.md

I don't know. These additional link documents are detailed information for level 3 (and higher), and Claude can choose to navigate and discover only as needed.

Among the PDF skills shown below, SKILL.md

Reference to two additional files (reference.md)

and table.md

Skill author chooses to be bound to the core SKILL.md

I don't know. By filling out notes in forms to a separate document (forms.md)

Skilled authors can maintain the core of their skills and believe Claude can read forms.md

Only when completing the form.

Progressive disclosure is a core design principle that makes agency skills flexible and scalable. Like a good manual for the Organization, first a catalogue, then a specific chapter, then a detailed appendix, and lastly a skill that allows Claude to load information only as needed:

Agents with file systems and code implementation tools do not need to read all their skills into context windows when dealing with specific tasks. This means that the number of contexts that can be tied to skills is actually unlimited.

The figure below shows how the context window changes when the user message triggers skills.

the order of operation indicated:

pdf/Skills.md

table.md

Documentation tied to skills; skills may also include the code Claude decides to implement as a tool.

Large language models are good at many tasks, but some operations are better suited to traditional code implementation. For example, it is much more expensive to sort a list by token than simply running a sorting algorithm. In addition to efficiency issues, many applications require certainty only by code.

In our example, PDF skills include a pre-written one. Python. Script to read PDF and extract all table fields. Claude can run this script without having to load the script or PDF into the context. Because of the certainty of the code, the workflow is consistent and replicable.

The following are some useful guidelines for the introduction of the development and testing skills:

Skills.md

The documents became difficult to process, breaking down their content into separate documents and quoting them. Maintaining path separation will reduce the use of tokens if some contexts are mutually exclusive or rarely used together. Finally, codes can serve as both an implementable tool and a document. It should be clear whether Claude should run the scripts directly or read them as references to the context.

and description

Your skills. Claude will use these when deciding whether to trigger skills to respond to his current mandate. Skills provide Claude with new capabilities through commands and codes. While it makes them strong, it also means that malicious skills may introduce loopholes in the environment in which they are used, or lead Claude to steal data and operate unexpectedly.

We recommend that skills be installed only from credible sources. When skills are installed from less trusted sources, please thoroughly review them before they are used. The first reading skills are tied to document content to understand its use, with particular attention to code dependence and bundled resources (e.g. images or scripts). Similarly, note that the skills indicate that Claude connects to a command or code that is potentially untrustworthy of an external network source.

Now Claude.ai, Claude Code, Claude Agent SDK And Claude Developer Platform supports proxy skills.

In the coming weeks, we will continue to add functions that support the full life cycle of the creation, editing, discovery, sharing and use of skills. We are particularly excited about the opportunity for skills to help organizations and individuals share their background and workflow with Claude. We will also explore how skills can complement model context protocols with more sophisticated workflows involving external tools and software to Acting Professors (see below).MCPServers.

Looking to the future, we want to enable agents to create, edit and evaluate their own skills and to code their own modes of behaviour as reusable functions.

Skills are a simple concept with a corresponding simple format. This simplicity makes it easier for organizations, developers and end-users to construct customized agents and to provide them with new functions.

We're happy to see what people are building with their skills. We started looking at our skills files and recipes today.

Written by Barry Zhang, Keith Lazuka and Mahesh Murag, all very fond of folders. Thank you very much. Anthropic Many others advocate, support and develop skills.

Product upgrading, operating methods, community focus, etc. Send it to your inbox every month.