Agent Skill

An agent skill is a reusable package of instructions, procedures, resources, and optional executable code that gives an AI agent specialized capability for a class of tasks. Examples include preparing a financial model, reviewing a codebase, processing a document format, or following an organization's incident-response procedure.

A skill differs from a single prompt. Prompts usually provide instructions for one interaction, while skills are persistent, discoverable assets that can be selected and loaded when relevant. They may contain:

• metadata describing what the skill does and when to use it;
• procedural instructions and domain-specific constraints;
• templates, examples, schemas, and reference documentation; and
• scripts that perform deterministic operations.

Skills support progressive disclosure in context engineering. Lightweight metadata can remain visible to the agent, while detailed instructions and resources are loaded only after the skill is selected. This reduces context usage when many capabilities are installed.

Skills and tools address different concerns. A tool exposes an operation through an executable interface. A skill explains how and when to combine operations, domain knowledge, and quality checks to complete a broader task.

Skill quality depends on clear triggering criteria, explicit prerequisites, bounded workflows, and verifiable completion checks. Versioning is important because changes can alter agent behavior across many applications.

Skills should be treated like software dependencies. A malicious or poorly reviewed package can instruct an agent to expose data, invoke tools unexpectedly, or run unsafe code. Trusted provenance, code review, sandboxing, and scoped permissions are required.

Anthropic's Agent Skills documentation describes a filesystem-based skill format with metadata, instructions, scripts, and on-demand resources.