The Model Eats the Harness — But Only the Middle
Anthropic deleted 80% of Claude Code's system prompt with no measurable loss. A four-layer anatomy of the agent harness, three tests for what should have been a script, and a worked example where a well-written skill dissolves into a git hook.