A skill that shortens agent responses; a CLI that compresses tool output too
A skill that makes agents speak briefly like a "caveman"; the CLI version also reduces logs, JSON, test, and web output, and offers optional model routing. Code, commands, and error messages are preserved verbatim. The only independently tested "winner" in this category: JetBrains found no quality loss in a test of 82 paired tasks, but output savings on real agent tasks were 8.5%, not the advertised 65%—even when forcibly enabled—because most agent output already consists of code and tool calls [1]. Elastic has also published its own tests [10].
Anyone who wants shorter, more readable agent responses; a low-risk but modest gain.
CLI telemetry is enabled by default, and in "Auto" mode, prompts and context may be sent to Caveman Cloud; use DO_NOT_TRACK=1 or install only the skill version. The rules add about 1,160 input tokens to every session, wiping out the savings on very short tasks.
npx skills add JuliusBrussee/caveman -g /plugin marketplace add JuliusBrussee/caveman /plugin install caveman@caveman