🔥 Hot Repo: 16K Stars in 7 Days — Claude Talks Caveman, Costs Drop 65%
JuliusBrussee/caveman went from zero to 16,000 GitHub stars in 7 days by making Claude Code respond in caveman-speak — cutting output tokens by 65% on average and backed by a March 2026 paper showing brevity can actually improve accuracy.
By OMC Editorial on 2026-04-11
A developer named Julius Brussee noticed something infuriating: AI coding agents like Claude are chronically verbose. Ask Claude why your React component re-renders and you get 69 tokens of polite explanation. Ask a caveman and you get 19 tokens: "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo." Same answer. Same fix. Fewer than a third of the words.
That observation became JuliusBrussee/cavemanhttps://github.com/JuliusBrussee/caveman, a Claude Code skill/plugin that instructs agents to respond in terse, article-free, filler-stripped prose. The repo launched on April 4, 2026 and hit 16,000 GitHub stars by April 11 — seven days — while accumulating multiple front-page Hacker News threads and coverage from Hackadayhttps://hackaday.com/2026/04/06/so-expensive-a-caveman-can-do-it/ and Decrypthttps://decrypt.co/363440/devs-claude-talk-like-caveman-cut-costs-work-better.
What It Does
Caveman ships as a Claude Code plugin and as a skill installable via npx skills add for Cursor, Windsurf, Cline, Copilot, Codex, Gemini CLI, and 40+ other agents. Once installed, the /caveman command activates a prompt that drops articles, pleasantries, hedging, and filler synonyms — while leaving code blocks, error messages, file paths, and technical terms completely intact.
The project benchmarked 10 real coding prompts against the Claude API, measuring actual output token counts in both normal and caveman modes. Average savings: 65%. Range: 22%–87%. The biggest win came from an "Implement React error boundary" prompt: 3,454 tokens normally, 456 in caveman — an 87% reduction. Savings compound over sessions because shorter responses mean faster generation times, which Brussee claims produces a roughly 3× speed increase.
Critically, caveman only compresses output tokens — thinking and reasoning tokens are unchanged. As the README puts it: "Caveman no make brain smaller. Caveman make mouth smaller."
The Science
A March 2026 paper, "Brevity Constr