Linked File Patterns

The third layer is linked files — external resources that Claude loads only when executing the skill. Templates, reference data, example outputs, validation scripts, and configuration files. These are the heaviest resources in the skill system, often running 2,000 to 10,000 tokens each. The progressive disclosure architecture ensures they are never loaded unless the skill is actively executing, keeping the token cost at zero during normal conversation.

Linked files are referenced by path in the core instructions. "Use the template at .claude/skills/templates/proposal-template.md as the output format." "Validate the output against the schema at .claude/skills/schemas/csv-schema.json." "Reference the checklist at .claude/skills/checklists/qa-checklist.md for the quality verification step." Claude reads these files just-in-time during execution.

The most effective linked file pattern is the golden example — a complete, correct output that demonstrates exactly what the skill should produce. Instead of describing the output format in prose, show Claude a finished example. "The output should match the structure and tone of .claude/skills/examples/competitor-analysis-example.md." Claude is a pattern-matching engine. A golden example is worth fifty lines of format specification.

Organize linked files in a consistent directory structure: .claude/skills/[skill-name]/templates/, .claude/skills/[skill-name]/examples/, .claude/skills/[skill-name]/schemas/. This convention makes skills self-contained and portable. A skill directory contains everything the skill needs — front matter, core instructions, and all linked resources. Moving a skill to another project means copying one directory.