Why MCP Beats Skills for LLM Tool Access

David Mohl lays out a practical case for treating MCP as the default interface layer for LLM tools, while limiting skills to knowledge transfer - essentially documentation, not execution. His friction list against skills is concrete: they depend on CLI access that ChatGPT and Perplexity can't use, they dump entire instruction files into context windows, and managing secrets in ephemeral environments is a mess. MCP avoids most of this through remote deployment, OAuth, and selective tool loading. He proposes renaming skills to something like "LLM_MANUAL.md" and treating MCPs as connectors. The distinction matters more now that Claude Code, Cursor, and ChatGPT are all converging on tool ecosystems but drawing the line between prompts and integrations in very different places.