Caveman

Caveman is an open-source token compression skill for Claude Code and Codex that strips conversational padding from agent output while preserving code, technical terminology, and accuracy. Uses a semantic constraint engine to remove hedging, sycophantic phrases, and filler — achieving up to 87% token reduction with a claimed 26-point accuracy gain on key benchmarks. Built for developers running agentic CLI workflows who want faster, cheaper responses without losing precision.