Skill Progressive Disclosure
Skill architecture pattern formalized by Anthropic in 'Equipping Agents with Agent Skills' article: a skill progressively exposes more details to the agent based on context. Level 1: short description (1 line) always in context. Level 2: full SKILL.md loaded if agent deems skill relevant. Level 3: resource files (templates, scripts) loaded on demand. Avoids saturating context with useless details while keeping full capability accessible.
Strengths
- Massive context savings: 50 installed skills ≈ 5KB in context instead of 500KB
- Pedagogically structured: skill becomes readable like a book (overview → details)
- Official Anthropic pattern, natively supported by Claude Code and Skills SDK
Limitations
- Requires writing discipline: structuring a skill in levels demands info hierarchy thinking
- Level 1 (1 line) must be very well written, it's the only moment agent decides to invoke or not
Best for
- Design system manager shipping internal skills (ds-audit, component-create, token-update) to team
- Team exceeding 5 installed skills and noticing slowdown / context overflow