What are Claude Skills?

Claude Skills are folders of instructions, scripts, and resources that Claude loads dynamically for specialized tasks. They can encode document standards, analysis methods, organization workflows, or personal procedures without keeping the full instructions in every conversation.

Claude reads skill metadata to decide whether a skill is relevant, then loads the full instructions and supporting files progressively. Skills can be created by Anthropic, installed from a directory, uploaded by a user, or provisioned across an organization where supported.

Claude Skills follow the open Agent Skills format, so the underlying package can work beyond Claude in systems that implement the specification. Claude Skills refers to Claude's discovery, activation, execution, and management experience around that format.

A general model may understand a task but still miss the organization's required sequence, terminology, templates, or quality checks. A Claude Skill packages that procedural knowledge so the same method can be applied repeatedly across conversations and team members.

Progressive loading keeps specialized procedures available without filling the context window with every instruction at once. Claude can inspect the skill catalog, activate the relevant package, and pull in a script, reference, or asset only when the task reaches that step.

Skills also separate procedure from access. MCP connections and other tools give Claude access to external systems. A skill teaches Claude how to use those capabilities for a particular workflow. The skill does not replace the connector's authentication, permissions, or approval controls.

A custom Claude Skill starts as a folder whose root contains SKILL.md or skill.md according to the supported packaging flow. Required metadata names the skill and describes what it does and when Claude should activate it. The body contains the working instructions.

Additional files can hold detailed references, templates, examples, or executable scripts. Claude first sees the compact metadata, loads the instructions when the skill is selected, and reads the additional files only when the instructions refer to them. This is the progressive disclosure model.

In Claude, a person can enable, upload, or discover skills through the available customization experience. Organization owners can provision approved skills for teams where supported. Before enabling a third-party skill, reviewers should inspect its instructions, scripts, dependencies, bundled assets, and external network requirements.

A company creates a board-update skill for Claude. The instructions define the narrative order, approved metric definitions, maximum length, citation rules, and the requirement to separate observed results from management interpretation.

A user asks Claude to prepare the monthly update from approved files and CRM exports. Claude activates the skill, loads a metric glossary only when it reaches the KPI section, runs a bundled validation script, and produces the update in the required structure.

The skill flags a mismatch between the CRM export and the finance file instead of choosing one value. Claude links both sources and asks for confirmation. The company can improve this behavior by updating the shared skill rather than rewriting prompts for every user.

Use Claude Skills for repeatable procedures, not permanent preferences or one-off requests. Put broad conversational defaults in custom instructions. Use Projects for persistent background knowledge. Use MCP for external tools and data. Use a skill for the task-specific method that connects those inputs.

Write the description as an activation contract. It should say when the skill applies and when it does not. Keep the main instructions concise, reference additional files explicitly, and document environment requirements such as code execution, packages, network access, or MCP connections.

Review third-party and internally shared skills before enabling them. Inspect executable code, package installation, file writes, network calls, credential access, and instructions that can change tool behavior. Test the skill after Claude, dependencies, tools, or organization policy changes.