Guyaga
WorkServicesAssetsWritingAbout
Start a project

Guy Aga

HomeWorkServicesAssetsWritingAbout
  • YouTube
  • LinkedIn
  • Instagram
  • Facebook

All writing · 14 Apr 2026 · 5 min read

The Caveman Trick: Cutting Claude Code Tokens by 75%

A viral skill makes Claude drop articles, pleasantries, and hedging — the 20k-star repo that teaches an LLM to grunt. I tested all four intensity modes on a buggy Python file. Here is what actually changed.

In this article

  1. Before and after
  2. Installing it
  3. Using it
  4. Intensity levels
  5. The real test
  6. When I'd actually use it
  7. Key takeaways

I saw the Caveman Claude Code skill on Reddit with 12,000 upvotes and almost scrolled past. Another joke project — teach Claude to talk like a Neanderthal, save tokens, move on. Then I actually installed it and watched my output tokens drop by two-thirds. Still a meme. Also working.

Reddit post: Taught Claude to talk like a caveman to use 75% less tokens

The premise is stupid on purpose. Claude Code burns tokens on pleasantries — "I'd be happy to help you with that," three-paragraph explanations you never asked for, hedging in every sentence. The caveman skill by Julius Brussee forces Claude to strip all of it: no articles, no filler, no "let me explain." The repo has 20.1k stars at the time of writing and hit #1 trending on GitHub and Hacker News.

github.com/JuliusBrussee/caveman, 20.1k stars

There's an actual research backbone under the meme — the March 2026 paper Brevity Constraints Reverse Performance Hierarchies in Language Models found that forcing concise output improved accuracy by 26 percentage points on certain benchmarks. Large models talk themselves into wrong answers through over-elaboration. Caveman is just that finding weaponised.

Caveman skill rock logo

Before and after

Ask normal Claude why a React component keeps re-rendering:

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. — 69 tokens

Ask caveman Claude:

New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo. — 19 tokens

Same answer. 72% fewer tokens. Code generation, tool calls, and internal reasoning are untouched — only the prose wrapping changes.

Side-by-side: Normal Claude 69 tokens vs Caveman Claude 19 tokens

Installing it

Two options. I used the plugin route.

Plugin install (recommended):

bash
claude plugin marketplace add JuliusBrussee/caveman

Terminal: adding JuliusBrussee/caveman marketplace

Then install the plugin itself:

bash
/plugin install caveman@caveman

Terminal: /plugin install caveman@caveman

Reload plugins and you should see caveman in the loaded list:

bash
/reload-plugins

Terminal: 16 plugins, 40 skills, 45 agents loaded

Standalone install (no plugin system):

bash
# macOS / Linux / WSL
bash <(curl -s https://raw.githubusercontent.com/JuliusBrussee/caveman/main/hooks/install.sh)

# Windows PowerShell
irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/hooks/install.ps1 | iex

Using it

Caveman registers a handful of slash commands:

Slash command list showing /init, /statusline, /add-dir etc.

The caveman namespace is where the interesting stuff lives — /caveman, /caveman-help, /caveman-commit, /caveman-review:

/cave command menu with caveman, caveman-help, caveman-commit, caveman-review

Trigger it any of these ways:

  • /caveman
  • "talk like caveman"
  • "caveman mode"
  • "less tokens please"

Stop with stop caveman or normal mode.

Caveman mode active. Full level. Ready. → stop caveman

Intensity levels

Caveman has three levels: lite, full, ultra. Switch by saying caveman lite or caveman ultra mid-prompt.

caveman lite command typed

ModeWhat it dropsFeel
NormalNothingDefault Claude
LiteFiller, hedging — full sentences stayPolite but blunt
FullArticles (a/an/the), long phrasesTerse, still readable
UltraAlmost everything non-essential, arrows replace wordsTelegraphic

The real test

I made a Python calculator with three intentional bugs and ran the same prompt — "find and fix the bugs in calculator.py" — through all four modes.

python
def divide(a, b):
    return a / b  # no zero check

def calculate_average(numbers):
    total = 0
    for i in range(len(numbers) + 1):  # off-by-one
        total += numbers[i]
    return total / len(numbers)

def get_percentage(value, total):
    return value / total * 100  # no zero check

calculator.py in VS Code with three bugs

Normal mode

Verbose, readable, finds all three:

code
Three bugs:

1. Line 14 — divide: no check for b == 0, raises ZeroDivisionError
2. Line 18 — calculate_average: range(len(numbers) + 1) goes one
   index past the end, raises IndexError
3. Line 23 — get_percentage: no check for total == 0, raises
   ZeroDivisionError

Normal mode response finding three bugs with line numbers

Normal mode editing calculator.py

Lite mode

Drops the padding, keeps cause-and-effect with arrows:

code
Three bugs found:

1. divide — no zero division check → ZeroDivisionError
2. calculate_average — range(len(numbers) + 1) is off-by-one →
   IndexError on last iteration
3. get_percentage — no zero check on total → ZeroDivisionError

Switched to caveman lite. No filler or hedging, but full sentences and articles stay.

Full mode

Articles gone, phrases tightened. Same bugs, shorter:

code
3 bugs found:

1. divide: no zero check → ZeroDivisionError
2. calculate_average: range(len(numbers) + 1) → IndexError on
   last iteration
3. get_percentage: no zero check → ZeroDivisionError

Caveman full mode active

Ultra mode

Stripped to primitives. range(len+1). last iter. Maximum compression:

code
3 bugs:

1. divide → no zero check → ZeroDivisionError
2. calculate_average → range(len+1) → IndexError on last iter
3. get_percentage → no zero check → ZeroDivisionError

caveman ultra command

Ultra active confirmation

All four modes found the same three bugs and applied the same fixes. The only variable is how much prose Claude wraps around the answer.

All modes comparison

When I'd actually use it

Honestly? Not all the time. Caveman is excellent when you're in a tight debugging loop — fast iteration, a lot of short questions, no need for explanations. It's a bad fit when you want Claude to teach you something, document a decision, or write a PR description someone else will read.

Pro tip: Pair caveman with a fresh /clear at the start of each debugging session. The token savings compound when your context stays small.

The paper's finding is the interesting part, not the meme. Forcing concise output made the model more accurate, not less — which lines up with every time I've watched Claude talk itself out of a correct answer by over-explaining. Caveman is a loud, unserious way to hit a real behavioural lever.

Key takeaways

  • ~75% fewer output tokens on prose-heavy responses — code and tool calls unchanged
  • Four modes: Normal, Lite, Full, Ultra — switch mid-conversation
  • Same answer quality in my test (3/3 bugs found at every level)
  • Best for: debugging loops, repeated queries, cost-sensitive pipelines
  • Worst for: teaching, documentation, anything a human will read later

Caveman skill end card


If you want to learn working with Claude Code better, visit the course page.

Claude Code · AI · Tools · Tutorial

Read next

Claude Managed Agents: Anthropic Just Shipped the Runtime

Newer

Claude Design: A Field Guide for Designers, Founders, and Engineers

Let’s build something.

Only I read it. Privacy

Or write to me

Guy@aga.digital
  • YouTube
  • LinkedIn
  • Instagram
  • Facebook

Guy Aga · Tel AvivPrivacy