If your Claude Code API bill feels high, the cause is rarely too many turns. It is a handful of measurable patterns: re-reading the same files, oversized tool output, and cache misses. The data is in the transcripts Claude Code already writes to ~/.claude/projects/. Full guide and a free tool: https://github.com/Ludoonus/claude-token-report
These practices are covered in depth in The Claude Code Operator's Handbook — 18 chapters on running AI coding agents safely and efficiently. Read a free sample or get it ($29).
Top comments (2)
Excellent analysis on token waste! The point about re-reading files and oversized tool output is spot-on.
One token sink I've noticed that isn't covered here: during brainstorming, Claude tends to waste tokens on premature tool calls — reading files, editing, running bash — while you're still thinking through ideas. I put together a small plugin to block that (Brainstorm-Mode under mehmetcanfarsak on GitHub) that prevents tool execution during brainstorming. Just chat with the agent without burning tokens on unnecessary file reads. Deactivate it when you're ready and it works normally. Has helped cut token waste on exploratory tasks.
For anyone curious about that, the install is straightforward — it's on GitHub under mehmetcanfarsak/Brainstorm-Mode. It's a drop-in plugin, activate it when you're exploring ideas, and it blocks Write/Edit/Bash while letting Read/Search pass through. Pretty minimal setup.