1. Placement decides whether subagents see shared content and how often you pay for it

    Shared content has four locations in Claude Code: agent body, rules files, skills, and Read calls. Two reach subagents. Two do not. The table, the decision tree, and the audit steps.

  2. Prompt caching, from the four fields up

    Every Claude API response returns a four-field `usage` block that tells you exactly what cached, what did not, and where your bill came from. Here is what the fields mean, what invalidates them, and how to diagnose a bad bill without leaving the response.

  3. Agent token cuts, 63-87%

    Reduce Claude Code agent token costs 63–87% by auditing where content lives across the four system-prompt layers. How to run the persona-vs-data audit, what frontmatter fields enforce structurally, and the subagent memory gap that re-bills rejected work.