Caveman Token Optimizer

This strips the pleasantries and filler from Claude's responses to cut output tokens by 65-75% while keeping technical accuracy intact. It drops articles, hedging phrases, and conversational fluff but preserves code blocks, exact error messages, and technical terms verbatim. Three intensity levels from lite (just drop filler) to ultra (telegraphic fragments). The benchmarks are real: a React debugging response went from 1180 tokens to 159 tokens with the same fix. Install once with npx or the plugin system, toggle it per session with /caveman commands. Useful when you're burning through API costs on long coding sessions and don't need Claude to say "I'd be happy to help" before every answer.