Reject Modernity: Forcing Claude to Speak 'Caveman' to Save 75% Tokens

Sup guys. Have you ever wanted to punch your screen because an LLM starts with "I'd be happy to help you with that" for the 100th time today? "Here is the code you requested...", "Let me summarize...". Shut up and give me the code! When you're burning midnight oil on a tight deadline, watching an AI type out polite fluff while eating up your tokens is pure agony.

Today, I'm bringing you a magical artifact that forces your ai tools to skip the throat-clearing, get straight to the point, and more importantly... save your hard-earned API credits.

TL;DR: Reject modernity, embrace Monke

Here's the quick rundown for you lazy scrollers: A mad lad named Julius taught Claude to talk like a literal caveman. What does that mean? Instead of giving you a full-blown thesis, it throws raw keywords at your face. The result? It grabbed 24.9K stars on GitHub and became the most useful meme in developer tooling today.

Let's check out the unga-bunga features:

• Token Diet: Cuts output tokens by ~75%. Less tokens to generate = speed go brrr (~3x faster).
• 4 Grunt Levels: From Lite, Full, Ultra, down to... Classical Chinese mode. Absolutely unhinged!
• Terse Reviews: Spits out 1-line PR comments like: L42: 🔴 bug: user null. Add guard. No BS.
• Input Compression: Automatically rewrites your CLAUDE.md into caveman-speak, saving ~46% of input tokens every session.
• Works Everywhere: One-line install for Claude Code, Cursor, Windsurf, Copilot, you name it.

Here's a before-and-after to show you the massive difference:

• 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..."
• Caveman Claude (19 tokens): "New object ref. Inline prop = re-render. Wrap in useMemo."

What the PH and Reddit tribes are saying

Naturally, the internet has opinions on this "de-evolution":

• The Hype Train: Most devs are loving it. One guy commented "BSB - Brain still big." Another dude mentioned he already hardcoded it into his AGENTS.md to load the caveman skill on the very first message.
• The Math Nerds: Some pointed out that injecting the system prompt initially increases input tokens. But don't panic—thanks to prompt caching, you're saving massive tokens in the long run.
• The Paranoid Seniors: Asking the real questions here: "At Ultra intensity, does it drop context that actually matters, like edge case safety warnings?" Another dev worried about "meaning drift" across sessions when the AI compresses instruction files.
• The Pragmatists: People are relieved because the tool is actually smart: it gives critical security warnings in full sentences, and keeps the rest strictly unga-bunga.

The C4F Verdict: Talk is cheap, tokens aren't

Let's be real. LLMs are verbose by default to please the general public. But for us devs, "Talk is cheap, show me the code." Forcing your AI to strip the polite fluff doesn't just save you money and rate limits; it actually helps you focus on the logic without drowning in text.

There's even a wild stat (citing a supposed March 2026 paper—time travelers, eh?) claiming that brevity constraints actually improve AI accuracy by 26% on certain benchmarks. Verbose is not always better.

Bottom line: This is a must-have survival tool for anyone hitting usage limits. Let the AI do the heavy lifting, and tell it to shut up about it.

Sauce: Product Hunt