caveman

why use many token when few do trick

A skill that makes AI coding agents respond in compressed caveman-style prose — cutting ~65% of output tokens while keeping full technical accuracy.

Install

tessl install juliusbrussee/caveman

Works with Claude Code, Cursor, Codex, Gemini, Copilot, and more.

Before / After

Normal Claude (69 tokens):

"The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object."

Caveman Claude (19 tokens):

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

Same fix. 75% less word. Brain still big.

Intensity levels

• Lite — terse professional. Drop filler, keep articles.
• Full (default) — caveman. Fragment sentences, drop articles, compress all.
• Ultra — maximum compression. Abbreviate everything.
• 文言文 — classical Chinese compression. Three variants: lite, full, ultra.

Switch anytime with /caveman lite, /caveman ultra, etc.

Benchmarks and evals

Same quality — same brain

38 Tessl task eval scenarios test whether caveman degrades technical correctness: 35 coding problems across 10 languages (JS, TS, Python, Go, Rust, Java, CSS, SQL, HCL, YAML) + 3 negative cases. Each scored against weighted technical checklists — zero style points, only facts.

19 independent runs across 4 agents:

Delta never exceeds 1.1 percentage points. On some scenarios caveman scores higher than baseline — brevity forces model to focus. Fewer word, same brain, as found by research that shows brevity constraints improved accuracy by 26 percentage points on certain benchmarks.

Full results on Tessl.

Reproduce:

tessl eval run skills/caveman --agent claude:claude-sonnet-4-6 \
  --variant without-context --variant with-context

...but for significantly less tokens

Range: 22%–87% savings across prompts. Caveman only affects output tokens — thinking/reasoning tokens are untouched.

More

Full documentation, additional install options, and sub-skills (caveman-commit, caveman-review, caveman-compress) at github.com/JuliusBrussee/caveman.

Quality evals contributed by Baruch Sadogursky using Tessl eval infrastructure.