SkillsIntermediateClaude Code
Skills: long reference goes in separate files
A short SKILL.md with a link to the heavy reference does not consume context on tasks that do not need it.
Claude works in three steps: it scans the name and description, loads the body of SKILL.md if it is relevant and only then pulls in the linked files (templates, specs, scripts). That is why the overhead is proportional to relevance, not to the number of installed Skills. You can also force loading by invoking it by name (/skill-name).
Example
api-docs-skill/ ├── SKILL.md <- format + 1 full example └── spec-template.yaml <- full schema, loaded on demand # Inside SKILL.md: Read ./spec-template.yaml for the full endpoint schema.