Progressive Disclosure in Agent Skills

Progressive disclosure is the architectural foundation of Anthropic's Agent Skills. The system loads skill information in three levels: Level 1 (metadata) provides just the skill name and description in the system prompt, helping Claude determine relevance without consuming context. Level 2 (core instructions) loads the complete SKILL.md file when Claude determines the skill applies to the current task. Level 3+ (supplementary resources) accesses bundled files like forms.md or reference.md only as specific scenarios require them. This hierarchical approach mirrors how you'd onboard a new team member—give them the overview first, detailed instructions second, and reference materials on demand.

This pattern solves the context window scaling problem. Agents with filesystem and code execution tools don't need to read the entirety of a skill into their context window—they load metadata cheaply at startup, then selectively access deeper content through tool invocation (like bash commands reading files). The result: Claude can maintain awareness of potentially unlimited skills while keeping context consumption minimal. For practical implementation details, see Building Agent Skills: A Practical Guide.