The `caveman` skill claims 65% output-token savings by cutting prose, not precision.

Caveman is at 94,167 stars as a single skill for Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ others, stripping conversational filler while keeping code, commands, and error strings byte-for-byte exact. The README's worked example compresses a 69-token React re-render explanation to 19 tokens with the same useMemo fix. Savings are output-side only, which matters if your bottleneck is a subscription output cap rather than context. The 65% figure is the maintainer's own benchmark, unreproduced.