How AI agents & Claude skills work (Clearly Explained)

Synopsis

Ras Mic explains how AI agents actually work: the anatomy of a context window (system prompt, agent.md, skills, tools, code, conversation, ~250k token limit), why agent.md files usually waste tokens, and how skills use progressive disclosure to keep just name and description in context until needed. He walks through building skills by walking the agent through a workflow, capturing the run, then recursively feeding failures back into the skill file.