Fun fact: I routed Claude Code to a local model and captured its system prompt. It is about 16K tokens long, compared with 6.5K tokens for Codex and 5.5K tokens for Gemini CLI. Some of the length is due to tool call descriptions, but I didn't have any additional MCP. No wonder why you can't effectively use local models with it unless you have a crazy machine that can run a bigger model capable of following lengthy instructions and handle very big context window. #LLM #AI gist.githubusercontent.com/chi…