Where my 2.2 billion tokens went, and what I changed

What you need

• Claude Code or Cowork. The audit has to read the usage logs on your computer, so a regular chat won't work.
• Steps 2 and 3 are Claude Code settings.

Run the audit

Paste this in, and swap [X] for the number on your usage page:

Claude says I used [X] tokens this week. That seems way too high. Read my usage logs, show me where the tokens actually went, and help me fix it.

Cap how big a chat gets

Open ~/.claude/settings.json and add this:

{
  "env": {
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "400000"
  }
}

Once a chat hits about 400K tokens, Claude summarizes it instead of re-sending the whole thing every turn. If the file already has settings in it, add the line inside the existing "env" block instead of pasting a second one. Any number from 100000 to 1000000 works.

Make it ask before launching a bunch of agents

Add this to ~/.claude/CLAUDE.md (your global instructions):

Before using the Workflow tool or launching more than 3 subagents, stop and ask me first. State roughly how many agents it will launch and why a single-agent pass is not enough. For reviews and security sweeps, default to a single-agent pass.

For me, one "do a full security sweep" prompt launched 321 agents and used 471M tokens.

Start fresh chats

When a feature is done, or the next day, start a new chat. Every message in an old chat re-sends everything before it.