Skills aren't the innovation
Skills are markdown files with optional packages attached. The file format isn’t the innovation. Progressive disclosure is.
I keep seeing the same question: how do I adopt skills in my framework? How do I use them in Mastra, LangChain, AI SDK?
Wrong question. The right question: how do I implement progressive disclosure?
In Claude Code, skills load when invoked. The agent sees a registry of skill names and descriptions. It doesn’t see the actual instructions until it decides to use one. Context stays lean until the moment it’s needed. That’s progressive disclosure: hide information from the LLM for as long as you can, reveal context only when needed.
This is Search → View → Use applied to agent capabilities. Search the registry. View the full instructions. Use the capability.
You don’t need Anthropic’s file format to implement this:
- Define capabilities as separate instruction sets
- Give the agent a registry (names and descriptions only)
- When the agent invokes something, inject the full instructions
- Execute
Anyone using any framework can implement this in an afternoon.
Skills are part of a larger wave. Anthropic is pushing ideas (MCP, Claude Code, skills) and everyone is adopting, just like everyone adopted OpenAI’s tool calling. Frameworks like Mastra and LangChain are downstream. It’s not on them to tell you how to adopt skills. The pattern is framework-agnostic.
There isn’t much to skills as a file format. But there’s a lot to progressive disclosure. That’s the idea worth adopting.