I think these system prompts reveal a couple of things about Claude Code and agentic tools in genereral.
1. Adding MCP/plugins creates a lot of overhead in the context window. For example you can see Figma taking up large base prompt and tool listings.
2. As agentic wrappers get more bloated, the system prompt grows. The memory facility is a good example. It's a wholly unimpressive mechanism taking up a large part of the context window.
3. Perhaps the most revealing thing is the security constraints that are just part of the prompt. The "auto mode" for permissions is just part of the system prompt.
As impressive as LLMs are, it seems that the only way to control this statistical beast is to add stuff to the base prompts. "IMPORTANT: Don't talk about goblins".
Claude Code wire trace reveals 13,000 words base prompt · HackerTrans
1. Adding MCP/plugins creates a lot of overhead in the context window. For example you can see Figma taking up large base prompt and tool listings.
2. As agentic wrappers get more bloated, the system prompt grows. The memory facility is a good example. It's a wholly unimpressive mechanism taking up a large part of the context window.
3. Perhaps the most revealing thing is the security constraints that are just part of the prompt. The "auto mode" for permissions is just part of the system prompt.
As impressive as LLMs are, it seems that the only way to control this statistical beast is to add stuff to the base prompts. "IMPORTANT: Don't talk about goblins".