Equipping agents for the real world with Agent Skills
02:00 · October 16, 2025 · Antropic Engineering Blog
Claude is powerful, but real work requires procedural knowledge and organizational context. Introducing Agent Skills, a new way to build specialized agents using files and folders.
Summary
Anthropic has developed Agent Skills as a structured mechanism for extending general-purpose agents like Claude with domain-specific capabilities. Each skill takes the form of a directory containing a SKILL.md file that begins with YAML frontmatter specifying a name and description. At startup the agent loads only this metadata into its system prompt, allowing it to decide relevance without consuming unnecessary context. When a skill is judged pertinent, the full SKILL.md content is read in; further files referenced from within that document can be retrieved on demand, implementing progressive disclosure that keeps the active context window bounded while permitting arbitrarily rich procedural knowledge.
The approach is illustrated by a PDF skill that augments Claude’s document-editing abilities. Although the model already interprets PDF content, it cannot directly manipulate form fields. The skill therefore bundles additional files such as forms.md alongside a pre-written Python script that extracts form fields deterministically. Claude invokes the script through its code-execution environment rather than generating the extraction logic token by token, yielding both efficiency and repeatable results. The same pattern supports any operation better performed by conventional code than by language-model inference.
Because skills package instructions and executable resources together, they function like onboarding materials that capture an organization’s workflows in reusable form. Authors are advised to keep core SKILL.md files concise, to isolate scenario-specific guidance in separate files, and to include only code whose dependencies and network interactions have been reviewed. Security guidance emphasizes installing skills exclusively from trusted sources and auditing bundled scripts and external references before use. The format is already supported across Claude.ai, Claude Code, the Agent SDK, and the Developer Platform, with further tooling planned for creation, discovery, and sharing.
Why it matters
Directly actionable for Product Teams and Builders: provides concrete implementation patterns, evaluation guidelines, and code patterns for building specialized agents. Addresses lifecycle, observability via progressive loading, and risks like malicious skills.
