Recently, I’ve been learning about Agent Skills, and this is how I understand them so far:
It has three levels:
Level 1: Metadata – similar to a system prompt, always loaded
Level 2: Instructions – loaded on demand to save context window
Level 3: Resources – References and Scripts, also loaded on demand. I see References ≈ lightweight RAG. Scripts ≈ Tools
Overall, agent skills feel like Agent + Router. Instead of one-off System Prompts, it uses progressive disclosure: load instructions and resources only when needed.
What I find interesting is the lower learning curve compared to building full agents — at the end of the day, it’s just a well-structured Markdown file.
Still learning, but this feels like a practical step toward more scalable AI systems with less overhead.
Reference: Agent Skills - Claude Docs