After the launch of Claude Skill, Claude took advantage of the situation to optimize the original shortcomings and limitations of MCP: A large number of inefficient llm requests consume a lot of tokens and the accuracy cannot be guaranteed.
The essential reason is that the mcp client at that time essentially used the prompt word engineering idea to guide the interaction between mcp and llm, while claude skill It is based on the idea of context engineering to interact with large models.
One of the most effective techniques is to offload the context to the file system.
The context here can be a specific executable scirpt file.
This is also This is the essential reason why the first mcp client can save a huge amount of tokens and greatly improve the success rate of mcp.
