Caveman (Cut Token Usage)

Caveman Claude Code sounds like a joke, but it's both funny and practical.

I didn't think it would work when I first saw it. Then I tried it and started seeing the same thing others were reporting: much shorter responses and real output-token savings, often up to 75% in the right tasks.

That's the reason this took off. It's not just "lol Grug talk." It's a direct way to cut waste when Claude is verbose.

Why this matters

Claude Code has a token-consumption problem in everyday use. You ask for a simple function and get the code plus polite filler like "I'd be happy to help with that" or "Let me search the web for you."

Those pleasantries add up across many turns. Caveman is helpful because it strips that layer out and pushes Claude to do the task first, give the result, and stop.

A few examples from the discussion you shared make the impact concrete:

• normal Claude for a web-search task: ~180 tokens
• Caveman Claude for the same task: ~45 tokens
• "I executed the web search tool": ~8 tokens
• "Tool work": ~2 tokens

Each tiny wording change is small on its own. Across a full workflow, those swaps can save 50-100 tokens repeatedly.

Before and after Caveman Claude

How Caveman started

I first saw it on Reddit and scrolled past because it looked like another meme project. Later, it had around 10,000 upvotes and people were posting real usage results.

Julius Brussee packaged the idea into an installable skill at github.com/juliusbrussee/caveman. It quickly passed 5,000 stars and hit #1 trending.

Caveman's Github Repo's stars over time

There was even a link in the repo to a March 2026 paper suggesting concise response constraints can reduce token usage and sometimes improve accuracy. That gave the meme some technical weight.

Why Caveman cuts tokens

The mechanism is simple: remove ceremony.

Normal assistant style includes setup language, transition language, status narration, and optional explanation by default. Caveman rewrites the style policy so the model prioritizes:

• short sentences (often 3-6 words)
• no filler or preamble
• tool first, result first
• no explanation unless requested

Key insight: Caveman is best as a style optimization for verbose replies, not a complete cost optimization strategy.

This style constraint does not change pricing itself. It changes output length, and output length directly affects output-token spend.

What the Reddit thread got right

Here's the Reddit thread: reddit.com/r/ClaudeAI/comments/1sble09/taught_claude_to_talk_like_a_caveman_to_use_75

The auto-generated TL;DR across 400+ comments was basically: this is peak efficiency and hilarious.

People liked it because they are tired of wall-of-text answers. The famous line, "Why waste time say lot word when few word do trick?", captured the whole point.

But the caveats are important:

1. The 75% claim is mostly about output tokens.
   Total cost can still be dominated by input context (chat history, files, tool traces) re-read on every turn.
2. Forcing a caveman persona can reduce clarity in some cases.
   You can save tokens and still lose answer quality if the task needs nuanced reasoning.
3. It's a practical hack, not a magic bullet.
   Great for trimming response bloat, not a full fix for overall AI spend.

Where to start

1. Download the skill here: github.com/juliusbrussee/caveman. You can also download the skill file to your skills folder, or copy the contents of the skill and recreate it yourself.
2. Use Caveman on repetitive coding and execution-heavy tasks where verbosity slows you down. Just call /caveman in your prompt.
3. Keep normal style for tasks that need deep reasoning, nuance, or detailed teaching.
4. Track both token reduction and response quality so you know where the tradeoff is worth it.

Additional Reading

Here are some related guides to check out:

1. What is a Skill?
2. How to Setup Claude (10 Minute Setup Guide)
3. How to Create Your Own Custom Skill

How to download

Install from GitHub: github.com/juliusbrussee/caveman