Caveman

🪨 why use many token when few do trick — a Claude Code skill/plugin (also Codex, Gemini, Cursor, Windsurf, Cline, Copilot, 30+ more) that makes the agent talk like a caveman: drop articles, filler, and pleasantries, keep substance, use fragments. Cuts ~65% of output tokens on average (range 22–87%) with 100% technical accuracy preserved. Brain still big. Mouth small.

🛠️ How to use

Quick-start, start to finish:

1. Install (one-time). Needs Node ≥18, takes ~30s, idempotent (safe to re-run), skips agents you don’t have.

   # Windows (PowerShell 5.1+) — my setup
   irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex

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

2. Turn it on. Type /caveman or say “talk like caveman”. On Claude Code a session hook writes a flag file so the agent talks caveman from message one — no command needed (also auto-activates each session on Codex and Gemini).
3. Pick a level (sticks until session end): /caveman lite (drops filler) → /caveman full (default caveman) → /caveman ultra (telegraphic) → /caveman wenyan (classical Chinese, shortest).
4. Use the sub-commands as needed — /caveman-commit, /caveman-review, /caveman-stats, /caveman-compress <file> (see 🧩 What you get below for the full list).
5. Turn it off. Say “normal mode” — do this for code, commits, and security writing where terseness hurts clarity.

Pair it with Headroom to also compress input/context tokens: caveman shrinks the mouth, Headroom shrinks what the brain reads. See Token Optimization for Claude Code for the broader toolkit.

Links

Description

• Repo: https://github.com/JuliusBrussee/caveman (MIT)
• Author: Julius Brussee

Download or use

# macOS / Linux / WSL / Git Bash
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash
 
# Windows (PowerShell 5.1+)
irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex

Needs Node ≥18. ~30s, idempotent (safe to re-run), skips agents you don’t have. Trigger: type /caveman or say “talk like caveman”. Stop with “normal mode”.

🧩 What you get

• /caveman [lite|full|ultra|wenyan] — compress every reply; levels stick until session end. lite drops filler, full is default caveman, ultra is telegraphic, wenyan uses classical Chinese (shortest).
• /caveman-commit — Conventional Commit messages, ≤50 char subject, why over what.
• /caveman-review — one-line PR comments (L42: 🔴 bug: user null. Add guard.).
• /caveman-stats — real session token usage + lifetime savings + USD; tweetable line via --share.
• /caveman-compress <file> — rewrites a memory file (e.g. CLAUDE.md) into caveman-speak, cutting ~46% input tokens every session; code/URLs/paths byte-preserved.
• caveman-shrink — MCP middleware that wraps any MCP server and compresses tool descriptions.
• cavecrew-* — caveman subagents (investigator/builder/reviewer), ~60% fewer tokens than vanilla so the main context lasts longer.
• Statusline badge — [CAVEMAN] ⛏ 12.4k (lifetime tokens saved); silence with CAVEMAN_STATUSLINE_SAVINGS=0.

For Claude Code, a session hook writes a flag file so the agent talks caveman from message one — no /caveman needed. Auto-activates each session on Claude Code, Codex, Gemini.

Reasoning for

Direct, repo-level lever for the same goal as Token Optimization for Claude Code, but attacking output tokens (and, via caveman-compress, input tokens) instead of context-window plumbing. Complements progressive-disclosure context design. Key caveat (from the README): caveman only affects output/visible tokens — thinking/reasoning tokens are untouched, so it makes the mouth smaller, not the brain. Biggest win is readability and speed; cost savings a bonus.

Useful here as an always-on mode for routine agent chatter in this vault and in Qamera AI / PLSoft work, while keeping normal mode for code, commits, and security writing.

🧪 Benchmarks

Real Claude API token counts, averaged over 10 prompts:

	Normal	Caveman	Saved
Average output	1214	294	65%
caveman-compress memory files	898	481	46%

Three-arm eval harness (baseline / Answer concisely. / skill) lives in evals/, so the delta is measured against an explicit terse baseline, not the verbose default. The README also cites a March 2026 paper, “Brevity Constraints Reverse Performance Hierarchies in Language Models”, reporting brevity constraints improved accuracy by 26 points on some benchmarks.

Caveman ecosystem

Five composable tools, one philosophy (agent do more with less):

• caveman — output compression (this note).
• caveman-code — whole terminal coding agent, caveman top to bottom; ~2× fewer tokens than Codex; MIT.
• cavemem — cross-agent memory compression.
• cavekit — spec-driven build loop.
• cavegemma — Gemma 4 31B fine-tuned on caveman pairs.

Alternatives considered

• Token Optimization for Claude Code — broader catalog of token-reduction tools (proxies, context sandboxes, code graphs); caveman fits its “output/memory compression” slot.
• Graphify — token reduction via knowledge-graph retrieval rather than terse phrasing.
• Agent Skills / Vercel Skills — the skill-packaging ecosystem caveman ships through.

Resources

• INSTALL.md — full install matrix, all flags, per-agent detail
• benchmarks/ + evals/ — raw data + reproduction harness
• Related: Awesome Claude Code, Building Claude Skills Guide, Karpathy Skills, Superpowers, gstack
• Extending Claude Code — Tools for Its Blind Spots — compiled article featuring this tool

---

Template: tool