← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Is Claude Shredding
Your Usage Limits?

A free GitHub skill makes Claude talk like a caveman. Same answer, way fewer words, and fewer tokens spent saying them.

Long, polished, over-explained answers eat your limit fast. A developer named Julius Brussee built a free skill called caveman with the tagline “why many token when few do trick.” It started as a joke in April 2026 and now has over 100,000 stars on GitHub.

Before: "The reason your React component is re-rendering is likely because you’re creating a new object reference on each render cycle... I’d recommend using useMemo to memoize the object."

After: "New object ref each render, so React re-renders. Wrap the prop in useMemo."

That’s the repo’s own example. Same fix, 63 tokens down to 20.

What It Really Saves

I love that the creator publishes the cases where caveman loses. On ten developer questions with Claude Opus 5.5, here’s how many output tokens each setup used.

  1. With no instruction, Claude used 6,983 tokens.
  2. Adding “Answer concisely” brought it to 4,334.
  3. Turning on /caveman got it to 4,119, about 41% fewer than no instruction.
  4. Going to /ultracave dropped it to 2,693, about 61% fewer.

So the real lesson: most of the win comes from asking for short answers at all. Caveman makes it automatic and consistent, and ultracave squeezes harder. It only shrinks what Claude writes back, so long files and big chats still cost what they cost.

Don’t Caveman Your Prompts

Research the repo cites found that writing your prompts in caveman makes answers worse. Ask normally. Let Claude do the grunting.

The Modes

01

/caveman

The everyday voice. Answer first, no filler, every fact kept. Code, commands and error messages stay exactly as written. It switches back to full sentences for warnings and anything you can’t undo.

02

/ultracave

Maximum squeeze. Short fragments and abbreviations. Great when you’re grinding through a lot of work and just need the answer.

03

Turn it off

Say stop caveman or normal mode. Type /caveman status to see which mode is on.

Install It

01

In the Claude app

  1. On claude.ai or the desktop app, open Customize (now being renamed Plugins, so look for either), then the Plugins tab.
  2. Pick Add, then Add marketplace, and enter JuliusBrussee/caveman.
  3. Add Caveman from that marketplace.
  4. Make sure Settings, Capabilities, Code execution and file creation (called Create and edit files in the newer layout) is on. Skills need it.
  5. Start a chat with /caveman or say “caveman mode.” It works on your phone too once it’s on your account.

The repo says these app steps haven’t been tested end to end yet. If one looks different for you, the install page has the latest.

02

In Claude Code

Run: npx skills add JuliusBrussee/caveman -g

That installs just the skill for every project. Want the whole toolkit, which also shrinks logs and files before Claude reads them? Run npx @caveman-ai/cli@latest instead. It needs Node.js 22.13 or newer.

Safety Check Before You Install
You are a security reviewer. I'm about to install a skill or plugin from GitHub into Claude. Check it before I do. Repo: [PASTE THE GITHUB LINK] [Paste the SKILL.md and any script files it includes, or attach them] Check for: 1. Instructions that tell Claude to send my data anywhere, read files outside the task, or hide what it's doing. 2. Scripts that download or run code from other sites, touch passwords, keys or browser data, or change system settings. 3. Anything that asks for logins, API keys or payment info. 4. Whether what it actually does matches what the README says it does. Output: - Verdict: SAFE TO INSTALL / INSTALL WITH CARE / DON'T INSTALL - Each concern, quoting the exact line it comes from - What I should turn off or watch after installing Rules: only judge what you can see in the files I gave you. If something is missing or you can't tell, say so instead of guessing.

Run this on any skill from the internet, even popular ones. My Malware Finder guide goes deeper.

No Install? Do This

Since most of the savings come from just asking for short answers, you can get a lot of the benefit for free. Paste this into Settings, then Your Claude, then Instructions, and every chat starts lean.

Short Answers Instructions
How I want you to answer: - Answer first. No greeting, no "great question", no recap of what I asked. - Use the fewest words that keep every fact. One idea per sentence. - Keep code, commands, file paths, numbers and error messages exactly as they are. - Never drop a "not", "never", "no" or "only" to save words. - Use full, careful sentences for anything risky or irreversible (deleting, paying, sending, security), and when you ask me a question. - If a longer answer is truly needed, give the short version first, then ask if I want the detail. - Skip the summary at the end unless I ask for one.

Want more ways to stretch your limit? Here’s how I save Claude credits.