Preview

What it is.
caveman is a Claude skill that changes how your coding agent replies. Each answer comes first, with no greeting or recap, and code and error messages stay exactly as written. It stays on for the whole session until you say stop caveman.
That matters because every word the agent writes is billed as an output token, and you read it before you act. A paired test by JetBrains across 86 coding tasks found caveman cut output tokens by 8.5%, with no measurable loss in quality.
It switches back to full sentences for security warnings and irreversible actions, and anything saved outside the chat, such as commits and docs, stays in normal prose. The skill is one part of a larger toolkit, whose full install also adds a proxy and model routing.
What you get.
- Replies that lead with the answer and end with the next step.
- Code blocks, commands, file paths and error messages left exactly as written.
- Full sentences for security warnings and before any irreversible action.
- Normal prose in anything saved outside the chat, such as commits and docs.
- Two stricter modes, /ultracave and /megacave, for even shorter replies.
- A one-line way out: say stop caveman or normal mode.
How to set it up.
Install the skill on its own with npx skills add JuliusBrussee/caveman -g.
Or, inside Claude Code, run /plugin marketplace add JuliusBrussee/caveman, then /plugin install caveman@caveman.
Type /caveman if it doesn't start on its own, or $caveman in Codex.
Use /ultracave when you want the shortest replies on routine work.
Say stop caveman or normal mode to go back to normal replies.
Leave out the full npx @caveman-ai/cli install unless you also want its proxy and model routing.
Pricing Plans
Free and open source under Apache-2.0.
Use cases
Long Claude Code sessions
You spend an afternoon building a lead-scoring pipeline in Claude Code. Replies stay short, so you scroll less and pay for fewer output tokens.
Debugging a broken workflow
You paste an n8n error into Claude Code. The reply names the cause and the change, with the error quoted exactly.
Working in a small terminal
Claude Code running over SSH on a laptop screen leaves little room. Short replies put the command near the top.
Best for
GTM engineers
They build workflows and scripts in Claude Code and want the agent's replies out of the way.
RevOps leads who script
They read agent output in a terminal and care about the exact command more than the explanation.
Teams paying per token
Shorter replies cost less when the bill counts output tokens.