# Caveman AI Skill **Caveman is an open-source ([[MIT License]]) agent skill that makes your coding agent answer in terse fragments instead of prose, to cut output tokens.** Its own tagline: "why use many token when few do trick". It works across [[Claude Code]] and 30+ other agents. The rule it injects is narrow on purpose. Drop filler, keep substance, use fragments — but never touch code, commands, or error messages. It compresses style, not content, and it keeps your language: write Portuguese, get grunted Portuguese back. ## Key aspects - Install via one script (`install.sh` / `install.ps1`), Node ≥18, detects and configures every agent on the machine - Activate with `/caveman` or just "talk like caveman" - Compression levels: lite, full, ultra, wenyan - `/caveman-stats` for savings, `/caveman-commit` for conventional commits, `/caveman-compress` to shrink memory files (reported ~46% input token reduction) ## The savings number is smaller than advertised The repo claims 65% output reduction. [[JetBrains]] ran an independent test on 82 paired agent tasks and measured **8.5%**. The gap isn't dishonesty, it's workload. The 65% figure comes from chat-style Q&A, where filler prose is most of the output. In a real coding agent the token stream is dominated by code, diffs and tool calls — exactly the things Caveman refuses to touch. So it compresses the small slice. Quality held: 8 tasks better, 10 worse, 64 tied, p = 0.82. No measurable difference. Two caveats worth carrying. The 8.5% only applies when the skill is forced on; letting the agent self-trigger saves less. And a ~10% cost saving is routinely wiped out by the variance of a single hard task. JetBrains' early 10-task run showed *-30%* savings before the sample got big enough to mean anything, which is a good reminder about small-sample benchmark claims generally. My read: worth installing because it costs nothing and doesn't degrade output, but do not budget around it. See [[Token Budget]]. ## Related - [[Ponytail (AI)]] — same shape (cross-agent skill, intensity levels, MIT), constrains *how much code* the agent writes rather than how much prose - [[Claude Code]] - [[AI Agents]] - [[Agentic Engineering]] - [[Context Engineering]] - [[Token Budget]] - [[MIT License]] ## References - Source code: https://github.com/juliusbrussee/caveman - JetBrains measurement: https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens