Caveman 2 cuts input tokens by 33%

The Caveman skill — an MIT-licensed add-on for Claude Code, Codex, Gemini, Cursor, Windsurf and 30+ other coding agents — shipped version 1.10 this week, branded Caveman 2 in the project README. The tagline captures the move in two lines: the original skill made agents say less; Caveman 2 makes them read less too.

65%fewer output tokens on a 10-task Caveman benchmark — and the v2 proxy adds a 33.2% cut on inputs.

The original trick is the simplest possible. The skill injects a short instruction set that turns an agent’s replies into terse caveman English while keeping code, commands and error messages byte-for-byte identical. The maintainer’s pinned task set, published in the repo, averages a 65% cut on output; individual tasks range from a 22% trim on already-tight prompts to an 87% cut on the verbose ones.

The new move is Caveman Proxy, a local process that sits between the agent and the model. It compresses what the agent reads — tool schemas, file dumps, MCP tool results (structured output from Model Context Protocol plugins), logs, repeated file contents — before the next provider call, and stores the originals on disk for retrieval. In a pinned 54-run Claude Code benchmark, the proxy reports a 33.2% cut on provider-reported input tokens while passing all 18 exact-answer checks.

The skill and the proxy are separable. The skill needs no account, no proxy and no code changes — one install command and a slash command, and the agent quiets down.

A parallel tack on the training-data side

The same idea is showing up at the training layer, in the Caveman Reasoning Compression collection on Hugging Face. The release note reads like a thesis: compressed reasoning traces train models to reason in about 2-3x fewer tokens at equal or better accuracy on the standard reasoning benchmarks.

The Gemma 4 12B and Qwen3.5 4B fine-tunes both train on compressed reasoning traces — see the box for per-model numbers. Terse reasoning is becoming a training objective, not just a prompt trick.

This is the strategic bit. Caveman-the-skill is a prompt-time intervention — your existing agent, with a thinner voice. The Hugging Face work is models trained to be terse natively — fewer tokens is the model’s property, not a setting. Two teams have arrived at the same destination from different directions.

The numbers behind the headline

Caveman is unusually upfront about its limits. The README has a section called Honest number warning:

In the surrounding prose, the maintainers spell out that input and reasoning tokens are untouched and that the skill itself adds around 1-1.5k input tokens per turn, so whole-session savings run smaller than the 65% headline. On already-terse workloads, they say, the skill can run net-negative. The maintainers publish docs/HONEST-NUMBERS.md and ask you to measure against your own workload before declaring victory.

That matters for a UK small team weighing it. On a Claude Pro or Max subscription at £18-20 a month, the bill is fixed and the savings arrive as faster replies and a less-clogged context window, not a smaller invoice. On usage-based API billing, the output saving lands on the meter — but reasoning tokens (which dominate agent workloads) are untouched by the skill. Whole-session savings run smaller than the headline, and the proxy is the only layer that touches inputs.

What to do this afternoon

The skill is the easy win. It needs Node.js 18+, finds every supported agent on your machine and is safe to rerun. Grab the one-line installer from the project’s INSTALL.md — the macOS/Linux and Windows PowerShell scripts are pasted there ready to run. After install, activate with the /caveman slash command. Expect the next reply to come back in tight caveman-speak — terse sentences, no filler, code and errors unchanged; if it goes too far, /caveman off restores the full voice. On a flat subscription, don’t expect a bill change — the win is a quieter agent and a less-clogged context window, not a smaller invoice.