Anthropic's progressive disclosure framework proves that treating skills as loader specifications, not static prompts, can reduce context costs by 3x. The original 1,200-line SKILL.md consumed 20% of the context window before any task execution. By restructuring to load only essential components at the right time, the same instructions achieved identical results with just 7% context usage. This architectural shift isn't about rewriting instructions - it's about rethinking how and when information enters the model's working memory.

The kitchen metaphor explains why this matters. Skills function like a chef's recipe system: frontmatter (YAML metadata) acts as the pinboard with one-line summaries always visible. The procedural instructions (SKILL.md body) come down only when needed, and references/scripts are accessed on demand like specific recipe pages or tools. Misplacing content violates this logic - adding frontmatter to reference files turned them into full skills, cluttering the agent's mental workspace. The fix required strict adherence to levels: 100-token frontmatter for routing, 500-line bodies for active execution, and unbounded references for deep dives. This structured approach prevents silent failures like context drift or version sensitivity, which plague prompt-based systems.

The practical impact is undeniable. Developers using Claude Code, Kiro, or Codex CLI can now build complex workflows without sacrificing performance. The key lies in architectural discipline - a 180-line SKILL.md pointing to external files outperformed a monolithic 1,200-line version because it paid for information only when executed. This matters for teams scaling AI tools, as context window efficiency directly affects cost and responsiveness. The lesson extends beyond Anthropic's ecosystem: any agent harness adopting skill.md as an open standard benefits from this architectural clarity. The kitchen analogy holds because both systems require precise organization to avoid chaos under pressure.