Blog

Technical writing on full stack development, AI, and open source

How Claude Code is dealing with prompts, two cases analysed

May 22, 2026 · 14 min read

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.

claude-code llm api-analysis prompt-engineering caching developer-tools anthropic ai-agents

Claude Code planning mode

May 21, 2026 · 14 min read

An analysis of Claude Code's two-layer planning system — plan mode and the Plan subagent — based on intercepted API calls, system instructions, and real session data from my inspector tool.

claude-code ai-tools prompt-engineering developer-tools planning llm-internals