Author Skills for progressive disclosure: split rarely-co-used paths into separate files to keep context lean

Anthropic Engineering·medium signal

Anthropic's Agent Skills guidance formalizes three-level progressive disclosure: YAML frontmatter (name + description) costs ~100 tokens per skill and is all that loads until a skill triggers; the SKILL.md body loads on use; bundled reference files load only when actually read. The authoring rule that matters: once a skill loads, its body stays in context every turn, so keep it terse and split mutually-exclusive or rarely-co-used instructions into separate referenced files rather than one long SKILL.md. Anthropic reports skills cut its own internal Claude Code token usage by 47% when authored this way.