agent-skills

Agent Skills package domain-specific instructions, metadata, and resources into modular units that Claude loads automatically. They transform Claude into a specialist for your exact workflow, whether code refactoring, API design, or other tasks, without manual prompt engineering each conversation.

Progressive disclosure loads Skill information in three stages: metadata always available, instructions on trigger, resources as needed. This optimizes token usage and reduces context window strain, letting you define deep expertise without bloating every conversation.

Define a Skill with YAML frontmatter specifying name and description, layer instructions across three loading levels, restrict tool access via allowed-tools, and package optional resources like scripts or templates. The modular structure lets you deploy and reuse the Skill across conversations.

Yes. Use the allowed-tools configuration in the Skill's frontmatter to restrict capabilities to only those your domain requires. This enhances security and focus by preventing the Skill from calling unrelated or sensitive tools.

Agent Skills provide structured progressive disclosure, tool-access control, and reusability across conversations. Custom prompts are single-use and load everything at once. Skills scale better for complex, domain-specific workflows and reduce wasted tokens on unnecessary context.

Agent Skills are modular units that extend Claude's capabilities without replacing existing workflows. They integrate into conversations on demand and support your current tool stack by controlling which capabilities activate for specific tasks.