Constrain output style to cut tokens — and brevity may raise accuracy, not lower it

GitHub·medium signal

The Caveman skill (JuliusBrussee/caveman) instructs the agent to answer in terse fragments with no preamble while keeping code, commands, error strings, and symbols byte-for-byte exact, averaging 65% output reduction across 10 prompts (range 22–87%). The non-obvious part is the auto-clarity rule: it reverts to normal prose for security warnings, irreversible-action confirmations, and multi-step sequences where fragment ambiguity risks a misread. It cites a March 2026 paper, 'Brevity Constraints Reverse Performance Hierarchies in Language Models,' reporting up to 26-point accuracy gains from brevity by shrinking the surface area for hallucination.