Caveman Mode prompt
Set up the free "caveman" plugin in my Claude Code so you (Claude Code) reply in short "caveman speak" - dropping greetings, filler and recaps - to stop long answers burning through my usage limits. caveman is a real, free, open-source Claude Code plugin (github.com/JuliusBrussee/caveman, MIT). It only shortens your OUTPUT - it never changes my code, my commands, or my error messages. Do the steps in order and change nothing until STEP 1. Safety first: check whether it is already installed - run `claude plugin list` and look for "caveman". If it is already there, tell me, change nothing, and skip to VERIFY. (If a bare `claude ...` command is blocked inside this session, re-run just that command with CLAUDECODE cleared: `CLAUDECODE= claude ...` on macOS/Linux, or `$env:CLAUDECODE=''; claude ...` in PowerShell.) STEP 0 - Prerequisite: Node 18 or newer (the plugin's installer runs on Node). Run `node -v`; if it is missing or older, stop and tell me to install or update Node first. STEP 1 - Install it for Claude Code ONLY (this does not touch Cursor, Codex, or any other tool). Run these two commands: claude plugin marketplace add JuliusBrussee/caveman claude plugin install caveman@caveman That is the whole install - no API key, nothing beyond Claude Code. STEP 2 - Restart so it loads. Plugins and their hooks only load at startup, so tell me to fully quit and reopen Claude Code now (approve "caveman" if it asks). Until I restart it is not active, so hand this back to me and wait. VERIFY (after I restart) - run `claude plugin list` and confirm "caveman" shows as installed. caveman auto-activates from the first message (it installs a SessionStart hook), so your next answer should already be short caveman speak - no filler, no recap. As a quick test, answer my next question tersely so we both see it is on. If it does not seem active, tell me to run `/caveman`, or to re-run `claude plugin install caveman@caveman` and restart. Finally, tell me how to drive it: `/caveman lite` (just drop the filler), `/caveman` or `/caveman full` (full caveman, the default), `/caveman ultra` (most terse), and "normal mode" to turn it off. And note honestly that it trims your spoken OUTPUT (roughly 60-65% fewer output tokens on average), not your hidden thinking - so it saves usage on long answers, it does not shrink reasoning.
Claude Code is quietly verbose. Ask it a simple question and you get a paragraph of "Certainly! I'd be happy to help", a restatement of what you just asked, a plan, the answer, and then a summary of what it already told you. Every one of those extra words is tokens - and tokens are exactly what your usage limits are counted in. So you hit your cap faster, for answers three times longer than you needed.
caveman fixes that. It is a free, open-source Claude Code plugin (github.com/JuliusBrussee/caveman) that makes the assistant reply in short "caveman speak": no greetings, no hedging, no recaps, just the substance, often in fragments. On average it cuts Claude's output by around 60-65%. The important part is what it does NOT touch: your code, commands, and error messages are kept exactly, and it does not make Claude any less capable - as the project puts it, it makes the mouth smaller, not the brain.
The prompt below installs it for Claude Code in about 30 seconds. It needs Node, checks whether you already have it, and installs it for Claude Code only so it leaves your other tools alone. Once it is in, caveman turns itself on from the first message of every session, and you can dial it with /caveman lite, full, or ultra (or say "normal mode" to switch it off). One honest caveat: it trims the answer you see, not Claude's hidden thinking, so it saves usage on long replies rather than heavy reasoning.