JuliusBrussee/caveman
Caveman is a Claude Code skill (also Codex, Cursor, and 30+ other agents) that makes the model answer in terse caveman-speak, claiming a 65% average cut in output tokens across 10 benchmark tasks.
**The gist** Caveman strips filler from agent replies — 'New object ref each render, wrap in useMemo'-style terseness — claiming **65% average output-token reduction** (range **22–87%**) across 10 benchmark tasks and **46%** on memory-file compression. It offers lite, full, ultra, and wenyan compression levels plus commit, PR-review, and stats sub-commands; a one-line installer needs **Node 18+**.
**Why it matters** Output tokens are the costly ones in long agent sessions, and this trims them without code changes. The repo cites a **March 2026 paper** suggesting brief responses can improve accuracy on some benchmarks, and the **/caveman-stats** command reports real session savings in USD so you can check the claim on your own workload.
**The gist** Caveman strips filler from agent replies — 'New object ref each render, wrap in useMemo'-style terseness — claiming **65% average output-token reduction** (range **22–87%**) across 10 benchmark tasks and **46%** on memory-file compression. It offers lite, full, ultra, and wenyan compression levels plus commit, PR-review, and stats sub-commands; a one-line installer needs **Node 18+**. **Why it matters** Output tokens are the costly ones in long agent sessions, and this trims them without code changes. The repo cites a **March 2026 paper** suggesting brief responses can improve accuracy on some benchmarks, and the **/caveman-stats** command reports real session savings in USD so you can check the claim on your own workload. **Watch out** Compression applies to **output only** — **thinking tokens** and input context are untouched — so savings on reasoning-heavy work fall short of the headline number, and gains shrink when the agent is already concise.