Caveman: Claude Code Token Compression Skill
JuliusBrussee/cavemanCaveman is a Claude Code token compression skill that strips the filler from your AI coding agent's answers, so replies come back as short fragments instead of paragraphs. It installs into Claude Code, Codex, Gemini, Cursor, and 30+ other agents from one command, and keeps code, commands, and error text byte-for-byte exact. The README reports it cuts about 65% of output tokens across its own benchmark set.
Reach for Caveman if your agent's long-winded explanations are burning output tokens and you'd rather read three-line answers than three paragraphs. Skip it if your prompts already get terse replies, or if input and reasoning tokens dominate your bill — the skill only compresses output, adds roughly 1 to 1.5k input tokens per turn, and the README openly admits it can go net-negative on already-short workloads.
The problem it solves
LLM coding agents pad answers: 'Sure, I'd be happy to help' preambles, restated context, and full-paragraph explanations of a one-line fix. On a metered API every one of those filler tokens costs money, and long replies are slower to skim. You either eat the cost or keep telling the model to 'be concise' on every prompt. Caveman exists to make brevity the default instead of a per-message chore.
What is it?
Caveman is a Claude Code token compression skill (also packaged as a plugin) that rewrites a coding agent's replies into terse, caveman-style fragments while leaving code, commands, file paths, and error text unchanged. It installs into Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ other agents from a single command. The README says it shrinks what the agent says, not what it knows, and reports roughly 65% fewer output tokens across a committed 10-prompt benchmark.
Why it's getting attention
Caveman sits at about 88,000 GitHub stars. The meme framing — 'why use many token when few token do trick' — travels well on social feeds, but the substance under the joke is real: it targets a cost that everyone paying for a coding agent feels, works across 30+ agents from one install, and ships benchmarks committed in the repo. The README's own 'honest number warning,' which spells out when Caveman loses, earned it credibility instead of hype-fatigue.
Key features
- ✓Six compression levels (normal, lite, full, ultra, wenyan) switchable mid-session with /caveman <level>; full is the default
- ✓Keeps code, commands, file paths, and error text byte-for-byte identical while compressing only the prose around them
- ✓/caveman-compress rewrites memory files like CLAUDE.md; the README reports ~46% fewer input tokens every session afterward
- ✓/caveman-stats reads your session log to estimate tokens and USD saved, with a lifetime counter on the Claude Code statusline
- ✓caveman-shrink MCP middleware wraps any MCP server and compresses its tool descriptions
- ✓cavecrew subagents (investigator, builder, reviewer) run terse so the main context window lasts longer
- ✓One installer detects and configures Claude Code, Codex, Gemini, Cursor, and 30+ agents; safe to re-run
Best use cases
- •Cutting output-token cost on chatty explain-and-review replies from a metered coding agent
- •Shrinking a large CLAUDE.md or memory file so every future session loads with less context
- •Generating short Conventional Commit messages with a capped subject via /caveman-commit
- •Producing one-line PR review comments with /caveman-review
- •Keeping subagent output terse so the main agent's context budget stretches further
How to install / try
Install with one command that scans your machine and sets up every agent it finds: ```bash curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash ``` On Windows PowerShell 5.1+: `irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex`. It needs Node 18 or newer, takes about 30 seconds, skips agents you don't have, and is safe to re-run. For just Claude Code as a plugin: `claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman`. The full per-agent matrix lives in the repo's INSTALL.md.
How to use
Turn it on by typing `/caveman` or saying 'talk like caveman'; on Claude Code, Codex, and Gemini a session hook makes it active from the first message with no command. Switch intensity with `/caveman lite|full|ultra|wenyan` and say 'normal mode' to turn it off. Caveman keeps your language — write Portuguese, get terse Portuguese — with wenyan as the deliberate exception, rendering in classical Chinese to pack the most meaning per token.
Strengths
- ✓One install wires up 30+ agents with no per-project setup, and re-running it is safe
- ✓Compresses style only, so code, commands, and errors stay exact — the README shows byte-preserved before/after output
- ✓Benchmarks are committed and reproducible in the repo rather than being marketing-only numbers
- ✓Honest about its own limits: a dedicated HONEST-NUMBERS doc explains when it wins and when it loses
- ✓No telemetry, accounts, or network calls after install — the skill is a prompt plus local hooks, per the README and SECURITY.md
Limitations & risks
- △Only shrinks output tokens; input and reasoning tokens are untouched, and the skill adds ~1 to 1.5k input tokens per turn, so whole-session savings run smaller than the headline 65% and can turn net-negative on already-terse work (the README says so)
- △The ~65% output and ~46% input figures are self-reported from the author's own 10-prompt benchmark, with per-task results ranging from 22% to 87% — treat them as estimates, not verified fact
- △Compressing prose can drop nuance: fragment answers read fast but lose the step-by-step hand-holding some users want, and the quality depends on the underlying model following the style instructions
- △The always-on behavior is specific to Claude Code, Codex, and Gemini; other agents need a manual /caveman toggle each session
- △/caveman-stats savings are local estimates, not audited receipts — the author positions the separate 'Caveman 2' to verify them
Alternatives
Who should try it — and who should skip
Anyone running a paid coding agent who reads a lot of long explanations and wants to trim output-token cost and reading time. It fits Claude Code, Codex, and Gemini users best, since it's active from message one there. Skip it if you already get short answers, or if input and reasoning tokens are the bulk of your bill — Caveman doesn't touch those.
Frequently asked questions
Caveman is a Claude Code token compression skill and plugin that rewrites your coding agent's replies into short caveman-style fragments while keeping code, commands, and errors byte-for-byte exact. It installs into Claude Code, Codex, Gemini, Cursor, and 30+ other agents from one command.
The README reports about 65% fewer output tokens across a committed 10-prompt benchmark, with per-task results from 22% to 87%. Those are self-reported numbers and only cover output — the skill adds input tokens and can be net-negative on already-short replies, which the README's honest-numbers note spells out.
The README's claim is that it shrinks what the agent says, not what it knows, and it leaves code, commands, and error text unchanged. It also cites a March 2026 paper arguing brief answers can improve accuracy on some benchmarks. Still, terse prose can drop nuance, so it's a readability/cost trade rather than a free win.
The README lists Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ others. On Claude Code, Codex, and Gemini it's on from the first message; other agents use a /caveman toggle.
Per the README and SECURITY.md, Caveman makes no network calls after install — no telemetry, analytics, accounts, or backend. The skill is a prompt, the hooks are local scripts, and /caveman-stats reads a log already on your disk.