Descriptions:
Cole Medin demonstrates how to implement Anthropic’s “skills” concept — also known as progressive disclosure — in any AI agent framework, independent of Claude Desktop or Claude Code. The core problem skills solve is context window bloat: rather than loading all capabilities upfront as MCP servers do, progressive disclosure lets an agent discover and load tools only when a conversation actually requires them.
The video includes a working GitHub template built with Pydantic AI that illustrates a three-layer disclosure pattern. The system prompt contains only brief capability descriptions; when the agent needs a skill, it loads a markdown file with full instructions; for complex API usage, an optional reference document is available as a third layer. A live demo shows the agent dynamically pulling in a recipe finder skill and a weather skill based on the user’s query, with logs confirming only the relevant tool enters the context window.
Medin also covers evals and observability as a way to verify the agent is genuinely leveraging skills rather than ignoring them. The approach is explicitly framework-agnostic — the same pattern works with LangChain, Crew AI, Agno, or no framework at all — and the template is designed to be dropped into existing projects with minimal modification.
📺 Source: Cole Medin · Published January 29, 2026
🏷️ Format: Tutorial Demo
