Breaking

Caveman Skill for Claude Code Cuts Output Tokens by 65% on Average

June 30, 2026 at 13:16 EDT

Software Dev & Coding
AI Agents
Open Source

An open-source skill/plugin called "Caveman" for Anthropic's Claude Code coding tool, which slashes output token usage by forcing extremely terse replies, has been gaining traction among developers. Released on GitHub around April 2026, the skill strips filler and verbose explanations from AI responses, returning only the essentials in a "caveman"-style shorthand to reduce the tokens consumed during generation.

April 2026 · Caveman for Claude Code

Talk Like a Caveman, Burn Fewer Tokens

An open-source skill forces Claude Code to drop filler and reply in terse "caveman" shorthand — cutting output tokens by an average of 65% so coding sessions last longer and costs stay down.

65%

Average output-token reduction across 10 tasks

87%

Max cut (React error boundary task)

~2×

Effective usage before hitting rate limits

Output tokens per task — Normal vs Caveman

Same unit: output tokens. Taller column = more tokens generated.

3454

Normal
(React fix)

456

Caveman
(87% cut)

1214

Normal
(10-task avg)

294

Caveman
(65% cut)

Same fix, fewer words

Normal

"The reason your component is re-rendering is that a new object reference is created on every render..."

Caveman

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

Why developers like it

Rate limits hit far slower — roughly double the usage
Works across 30+ agents (Cursor, Copilot, Gemini CLI…)
lite / full / ultra compression levels via /caveman
Some tests saw more focused, higher pass-rate answers

Where it falls short

Only trims output — little help when input context is the cost driver
Terse style may slightly dent complex reasoning quality
Optimizing the whole input often matters more than output alone

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.