Claude Code's Caveman Plugin Benchmarked Against 'Be Brief' — Real Token Savings of 22-87% Depending on Prompt

Max Taylor / Hacker News·medium signal

A developer benchmarked the viral Caveman Claude Code skill (560+ GitHub stars) that strips filler words from AI responses to save output tokens. Real-world benchmarks show 61-68% reduction on discursive text (about 25% of total session), with thinking/reasoning tokens untouched. Multiple intensity levels from 'lite' to 'ultra'. A March 2026 paper found brevity constraints actually improved accuracy by 26 percentage points on certain benchmarks. 87 points, 64 comments on HN.