How Claude Code is dealing with prompts, two cases analysed
I intercepted every API call Claude Code makes and compared two sessions: a simple greeting (8 calls, 22 output tokens) vs. a full security review (81 calls, multi-agent orchestration). The same 19K-token system prompt powers both. Here's what I found about routing, caching, subagents, and the real cost of AI-powered development tools.