Caveman Prompting: Cut LLM Output Tokens by 65%

Your AI agent just used 1,200 tokens to tell you about a race condition in PostgreSQL. The fix was three lines. The other 1,100 tokens were filler. Pleasantries. Hedging. "I'd be happy to help you with that."

Caveman would have said it in 232 tokens. Same fix. Same accuracy. 81% less noise.

#What Is Caveman

Caveman is a Claude Code skill created by Julius Brussee that forces AI agents to communicate in terse, compressed language while preserving 100% technical accuracy. Think of it as a semantic constraint engine for agentic communication.

The tagline says it all: "Why use many token when few token do trick."

It works by stripping out articles (a, an, the), filler words (just, really, basically, actually, simply), pleasantries (sure, certainly, of course, happy to help), and hedging language. Code blocks, error messages, and technical terminology pass through untouched.

The result is output that reads like a telegram from someone who genuinely respects your time.

#The Numbers Tell the Story

Caveman achieves an average 65% token reduction across standard software engineering tasks. But the range is what makes it interesting.

A React re-render diagnosis that normally takes 69 tokens drops to 19. That is 87% savings on a single debugging exchange. Auth token expiry fixes go from 704 tokens to 121. PostgreSQL race condition debugging drops from 1,200 to 232. Even a Git rebase vs merge explanation, which requires actual teaching, still saves 58%.

The median task saves you roughly two-thirds of your output tokens. At current API pricing where output tokens cost significantly more than input tokens, this is not a style preference. It is a cost engineering decision.

#How It Actually Works

Caveman operates at three intensity levels, each progressively more compressed.

Lite mode keeps grammar intact but removes all fluff. Professional terseness. No "I'd be happy to help" preambles, no "Let me explain" throat-clearing. Just the answer.

Full mode is the default. Sentence fragments. Dropped articles. Complete caveman energy. "Bug in auth middleware. Token expiry check use less-than not less-than-or-equal. Fix:" followed by the code block.

Ultra mode goes telegraphic. Maximum compression. Abbreviated everything. For developers who have seen enough "Sure!" responses to last a lifetime.

There is also a wenyan (文言文) mode that applies Classical Chinese literary compression patterns to English output. Three sub-variants, each increasingly terse. This is not a joke. It works.

Activation is simple: say /caveman, type $caveman in Codex, or just tell the agent "talk like caveman" or "less tokens please." It sticks until you say "stop caveman" or "normal mode."

#The Science Behind the Grunt

Here is where Caveman gets genuinely interesting. A March 2026 paper by MD Azizul Hakim evaluated 31 open-weight models across 1,485 problems and found something counterintuitive: brevity constraints improved accuracy by 26 percentage points on certain benchmarks.

The paper identifies a mechanism called "spontaneous scale-dependent verbosity." Larger models tend to over-elaborate, reasoning themselves into wrong answers. When you constrain them to be brief, they cut the overthinking and land on correct solutions more reliably.

This means Caveman is not just saving tokens. It may be making your agent smarter by preventing it from talking itself into errors.

The paper also found that brevity constraints reduced performance gaps between large and small models by up to two-thirds. So your smaller, cheaper model might perform closer to the flagship when you stop letting it ramble.

#Three Skills, One Philosophy

Caveman ships with three specialized skills beyond the core communication constraint.

caveman-commit generates Conventional Commits-compliant messages with subjects under 50 characters. It focuses on the "why" over the "what" because good commit messages matter more than most developers think.

caveman-review produces one-line PR review comments. Instead of a paragraph about a potential null pointer, you get: "L42: bug: user null. Add guard." Same information. Fraction of the noise.

caveman-compress is the most powerful skill. It recompresses memory files like CLAUDE.md, project notes, and session context by 36 to 60%. Code, URLs, and technical content stay intact. Only the surrounding prose gets crushed. It creates a .original backup so you can always recover the verbose version.

This last skill is significant for anyone running long agentic sessions. Context windows fill up. Memory files bloat. Caveman-compress lets you reclaim window space without losing information.

#Installation Is One Line

Caveman runs on virtually every AI coding agent available today.

For Claude Code, it auto-activates via SessionStart hooks after a marketplace install. For Codex, it uses repo-local .codex/hooks.json. Gemini CLI gets a native extension. Cursor, Windsurf, Cline, Copilot, and over 40 other agents install through npx skills.

The fact that a single skill can work across this many platforms speaks to how the AI agent ecosystem is converging on shared conventions. Skills are becoming the portable unit of agent behavior.

#Why This Matters for Context Engineering

Context engineering is the practice of designing what goes into and comes out of an LLM's context window. Every token in that window has a cost: financial, latency, and cognitive. Wasted output tokens compound across multi-turn agent sessions because each response becomes part of the context for the next turn.

Caveman attacks the output side of this equation. By compressing responses at the generation layer, it reduces what flows back into context on subsequent turns. The savings multiply.

Consider a 10-turn debugging session. Normal Claude might produce 8,000 output tokens across those turns. Caveman produces 2,800. That is 5,200 fewer tokens flowing through the context window, which means the agent has more room for actual code, more room for reasoning, and lower latency on every subsequent response.

This is not about being cheap. It is about being efficient with a finite resource.

#The Viral Moment

Caveman's GitHub stars jumped from dozens to 500 in half a day, eventually passing 20,000. It hit the front page of Hacker News. Hackaday covered it. Yahoo Tech reported on it. Chinese tech media called it "the most powerful prompt skill of 2026."

The resonance is obvious. Every developer who has watched Claude open with "Sure! I'd be happy to help you with that!" before answering a yes-or-no question felt something when they saw Caveman. It validates a frustration that has been building since GPT-3: AI is too verbose, and we have been paying for the privilege.

#Should You Use It

If you run AI agents in production or spend meaningful time in Claude Code, Codex, or Cursor, Caveman is worth a try. The savings are real, the installation is trivial, and the worst case scenario is you type "stop caveman" and go back to normal.

For teams with API budgets, the math is straightforward. A 65% reduction in output tokens at scale translates directly to reduced costs and faster response times. The research suggesting improved accuracy is a bonus.

For solo developers, Caveman just makes the experience better. Less scrolling. Less noise. More doing.

Caveman not dumb. Caveman efficient. Caveman say what need saying. Then stop.

#Resources

GitHub: JuliusBrussee/caveman

Landing page: juliusbrussee.github.io/caveman

Research: "Brevity Constraints Reverse Performance Hierarchies in Language Models" (arXiv:2604.00025, March 2026)