What's this about?
A Claude Code command almost nobody uses that clears out unused skills and MCP servers and identifies overlong CLAUDE.md files. One user reports killing 10K of permanent context in two minutes, and works the arithmetic: at 50 agent sessions a day, that is 500,000 tokens a day, 15 million a month — before counting the knock-on costs of repeated file reads, delayed compaction, and rework after quality degrades. dotey confirms: "it works."
Why does idle context cost so much?
Because it is resident. Every skill definition, every MCP server's tool schemas, and your CLAUDE.md are prepended to every request in every session. Ten thousand tokens you are not using is not a one-time cost — it is ten thousand tokens multiplied by every turn of every conversation, forever, plus the second-order effect that a fuller context compacts sooner and degrades faster.
Why does this matter to you?
- It is the highest-leverage thing in this digest that takes two minutes and costs nothing.
- MCP servers are the usual culprit — each one's tool definitions are permanently in context whether or not you call them.
- It pairs with disabling the 1M context window from yesterday: both are about what is resident rather than what is used.