Anthropic Gives Claude New ‘Agent Skills’ to Master Real-World Tasks

Anthropic has unveiled Agent Skills, a new framework designed to give its Claude models procedural knowledge and domain-specific expertise. The launch marks a step toward building agents that can dynamically adapt to real-world workflows using files, scripts, and structured context.

Agent Skills allow developers to package instructions, scripts, and resources into organised folders that Claude can load and interpret when needed. The system builds on the principle of progressive disclosure, letting the model read information incrementally, from metadata to detailed guides, without overloading its context window.

At the heart of each skill is a SKILL.md file containing YAML ‘frontmatter’ with metadata like name and description as the company describes. This helps Claude determine when to activate a specific skill. For example, a PDF skill can extend Claude’s native understanding of documents by adding the ability to fill forms or extract fields via bundled Python scripts.

Anthropic said the approach allows for “composable, scalable, and portable” capabilities, turning general-purpose models into specialised agents. By referencing additional files such as forms.md or reference.md, developers can create layered skill structures that Claude reads only when relevant to the task.

Agent Skills also support code execution, enabling Claude to run deterministic scripts for computational tasks instead of relying solely on token generation. This combination of structured context and executable tools helps ensure both accuracy and efficiency.

“Skills are a simple concept with a correspondingly simple format,” the company stated.

Anthropic emphasised security best practices, advising users to install only trusted skills and audit files for malicious dependencies.

The new feature is now available across Claude.ai, Claude Code, the Claude Agent SDK, and the Claude Developer Platform. Anthropic plans to expand capabilities in the coming weeks, allowing agents to eventually create, edit, and evaluate skills autonomously.