creating-agent-skills

Structure a Claude Code skill by using predictable frontmatter and a concise XML body, placing detailed implementation references in separate files to enable progressive disclosure and reliable execution.

A router pattern in agent skills directs execution to specialized sub-skills based on intent. Use a router architecture for complex, multi-domain tasks, and choose a simple single-file skill for focused, narrow operations.

Validate a SKILL.md file by auditing its naming conventions, frontmatter consistency, and reference patterns against established criteria, iterating on the structure to ensure reliable discovery and execution.

Progressive disclosure in prompt design reduces ambiguity by keeping the main SKILL.md concise and moving detailed instructions, scripts, and templates into reference files that the agent loads only when needed.

Avoid a router-based architecture when your task is highly focused and self-contained; a simple single-file skill provides sufficient structure without the overhead of routing logic and multiple reference files.

You can use existing templates to start designing a Claude Code skill by reading the current SKILL.md, sketching your intended structure, and creating a minimal, self-contained skeleton following recommended patterns.