Most MCP Servers now are essentially a mishmash of Skills disguised as "protocols". They return conclusions without providing the original capabilities, stuffed internally with prompts, rules, and processes, having already determined the order for the model. MCP should tell you what can be called, not how to do it. Once "how to do things" is written into the MCP, the Agent degenerates into an RPC client. Anthropic's push for Skills is actually about bringing the thinking, processes, and decisions back into the model. The core takeaway is: stuffing "how to do things" into the MCP shows a lack of trust in the Agent.