Caveman: Cut AI Response Tokens by 75% Without Losing Accuracy

What Caveman Does#

It makes AI coding assistants shut up — not by making them dumber, but by stripping filler.

Before Caveman, Claude Code explains a React re-render like this:

“The reason your React component is re-rendering is likely because you’re creating a new object reference on each render cycle. This causes React to see the prop as changed, triggering a re-render…”

After Caveman:

“New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo.”

Same fix. 69 tokens down to 19. 72% reduction.

Setup#

30 seconds.

# macOS / Linux / WSL
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash

# Windows PowerShell
irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex

Requires Node.js 18+. The install script auto-detects which AI coding tools you have and installs only for those. Type /caveman in Claude Code or say “talk like caveman” to activate. Say “normal mode” to stop.

Four Compression Levels#

Extra Features#

/caveman-commit — Generates Conventional Commit messages with subject lines under 50 chars, emphasizing “why” over “what.”

/caveman-review — Single-line PR review comments: L42: 🔴 bug: user null. Add guard.

/caveman-stats — Real-time token usage, lifetime savings, and dollar amounts. Add --share for a shareable stats line.

/caveman-compress <file> — Rewrites CLAUDE.md and project memory files. Average 46% input token reduction while preserving code, URLs, and paths verbatim.

What Surprised Me#

Compression is better than expected. Official benchmarks show 65% average reduction across 10 tasks, and my experience matches. The React re-render example went from 1,180 to 159 tokens — 87% reduction.

It only affects output tokens. Thinking/reasoning tokens are completely untouched. A 2026 paper found that constraining models to be concise actually improved accuracy by 26 points on some benchmarks. Less talk, more correct.

Things to Watch Out For#

Not all 30+ supported AI agents have identical feature sets. Main ones (Claude Code, Cursor, Codex, Windsurf, Cline, Copilot) are fully supported. The status bar badge [CAVEMAN] ⛏ 12.4k in Claude Code updates on each /caveman-stats run. Set CAVEMAN_STATUSLINE_SAVINGS=0 to disable.

What’s Next#

Try different compression levels to find your daily sweet spot. Pair with caveman-shrink (MCP middleware) for compressed tool descriptions and cavecrew sub-agents for reduced sub-task token usage.