Anthropic Skills (anthropics/skills) | Ry Walker Research

Key takeaways

Anthropic's skills repository is a catalog; the separate Agent Skills standard defines the portable file format.
Progressive disclosure delays loading instructions and resources, but discovery metadata still consumes context.
Apache-licensed examples and source-available document skills have different reuse rights.
A shared format does not make scripts, permissions, or Claude-specific features portable across runtimes.

FAQ

What are Anthropic Agent Skills?

Folders of instructions, optional scripts, and resources that Claude can load for a task. Anthropic publishes examples and document skills in anthropics/skills; the underlying format is specified separately at agentskills.io.

Are all the skills open source?

No. Many examples use Apache 2.0, while the document skills are source-available under separate terms. Read the license inside the skill you intend to use.

How do you install the catalog in Claude Code?

Register anthropics/skills with /plugin marketplace add, then install document-skills@anthropic-agent-skills or example-skills@anthropic-agent-skills.

Will one skill work identically in every agent?

No. The file format is shared, but tool availability, dependencies, network access, permissions, and client-specific frontmatter can change behavior.

Executive Summary

Anthropic Skills is Anthropic's public catalog of reusable Claude workflows, including examples and document-processing skills. The repository is an implementation of the format, not the standard itself. Its README also warns that these educational examples may differ from the implementations delivered inside Claude.^[1]

Anthropic introduced Skills in October 2025 and published the format as an open standard on December 18, 2025. The useful design is incremental context loading: make task instructions discoverable, then read their details when needed. That is a packaging and execution pattern, not a guarantee that an agent follows every instruction or produces correct output.^[2]

Attribute	Assessment
Maintainer	Anthropic
Catalog	Examples plus document skills
Licensing	Mixed: Apache 2.0 examples and source-available document skills
Category	Reusable skills catalog; see agentic skills frameworks

The catalog and licensing distinctions above are documented by Anthropic.^[1]

Product Overview

The standard requires a directory containing SKILL.md, with YAML name and description fields and a Markdown body. Optional scripts, references, and assets directories hold executable helpers, supporting material, and templates. A compatibility field can describe environmental requirements; the experimental allowed-tools field may behave differently between clients.^[3]

Progressive Disclosure

Stage	What loads
Discovery	Names and descriptions of available skills
Activation	The selected skill's instructions
Execution	Referenced material or scripts as needed

This saves loading every complete skill upfront; it does not make discovery free or impose a hard upper bound on execution context. The specification recommends a main file under 500 lines and moving detailed material to references.^[3]

Technical Architecture

In Claude Code, skills can be personal, committed to a project, or distributed through plugins. The catalog's installation commands are:^[1]

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

Claude Code adds features beyond the common format, including invocation controls, subagent execution, and dynamic context injection. Its current guide also describes syncing enabled claude.ai skills into Cowork/cloud sessions and opt-in downloads for local sessions. These extensions and sync rules should not be assumed to exist in another client.^[4]

Documentation conflict, checked September 15, 2026: the general Platform overview still says custom skills do not sync between surfaces, while the specific Claude Code guide documents the syncing behavior above. Follow the guide for the exact surface and deployed version; this review did not exercise account syncing.^[5]^[4]

Strengths

Small conceptual surface. Instructions and supporting files remain inspectable and versionable; the standard can represent a simple procedure without a separate service.^[3]
Reusable procedural knowledge. Anthropic's engineering design separates task discovery from the larger instructions needed to execute it.^[2]
Useful reference implementations. The catalog includes creative, technical, enterprise, and document workflows that can inform custom skill design.^[1]

Cautions

Execution inherits runtime risk. Anthropic warns that malicious skills can misuse tools or expose data. Review scripts and instructions as software dependencies.^[5]
The API is a different environment. Its documented skills runtime has no network access or runtime package installation. Agent Skills are not covered by the Platform's zero-data-retention arrangements.^[5]
Automatic selection needs evaluation. Claude Code documents troubleshooting for skills that fail to trigger or trigger too often; a descriptive file does not mechanically enforce a workflow.^[4]

What Developers Say

In an October 16, 2025 hands-on post, independent developer Simon Willison tested the Slack GIF skill with Sonnet 4.5. He valued the reusable validation code and simple packaging while judging the generated GIF poor. His example supports the distinction between making a capability available and obtaining a good result; it is not a benchmark of the current catalog.^[6]

In the launch discussion, Hacker News user pants2 described already using a similar approach through an MCP service and questioned whether it needed a formal specification. That is a useful design critique from one participant, not evidence of general developer sentiment.^[7]

Pricing & Licensing

There is no single license covering every skill. The repository labels many examples Apache 2.0 and the document skills source-available.^[1] The current docx license ties use to an Anthropic agreement and includes restrictions on copying, derivatives, and distribution. Public source availability should not be interpreted as permission to redistribute that skill or port it to an unrelated service.^[8]

Running a skill also uses its host product and compute; downloading a folder is not a free Claude service entitlement. Verify the applicable product plan and each skill's terms when deploying it.

Competitive Positioning and Fit

For teams comparing Superpowers, BMAD, and other agentic skills frameworks, this catalog is best evaluated as source material for particular tasks. The editorial recommendation is to start with a small, reviewed skill that addresses a recurring need, then measure completion quality and selection accuracy in the intended runtime.

Best fit: Teams documenting repeatable workflows and developers studying skill structure.

Poor fit: Teams expecting the catalog to provide a complete development methodology, cross-client permission consistency, or automatic quality assurance.

Viability Assessment

Anthropic continues to publish the catalog alongside documentation for the format and its product integrations. That provides a concrete maintenance context, but it does not establish a support commitment or compatibility guarantee for every example. The repository explicitly tells users to test skills in their own environment.^[1]

Bottom Line

Anthropic Skills is useful as a practical reference for packaging agent procedures. Evaluate the specific skill, its license, and its runtime together. The open format is valuable; identical behavior and unrestricted reuse do not follow from it.

Research by Ry Walker • methodology

Sources