Agent Skills standard — progressive disclosure for coding agent context¶
Insight: The Agent Skills standard establishes three distinct context mechanisms for coding agents: Skills (progressively loaded context folders scanned by name/description, loaded fully only when relevant — "lazy loading for context"), Rules (always-active constraints loaded every session), and Commands (user-triggered actions like slash commands). The key architectural insight is that skill descriptions serve as routing metadata — they must be keyword-dense for discovery, while skill bodies should be procedural checklists, not reference wikis.
Detail: Skills use a standardized folder structure: .provider/[skill-name]/SKILL.md with metadata and optional extras (scripts, templates, reference docs). The progressive disclosure pattern prevents context pollution — agents don't read every skill upfront, just scan names and descriptions. This mirrors web development lazy-loading patterns. Builder.io implements this standard alongside Claude's native format, and skills placed in .claude/ directories are also discoverable.