Section 1
What it actually does
Claude has a default personality: helpful, polite, a little wordy. It opens with "Sure! I'd be happy to help you with that." It closes with "Let me know if you'd like me to elaborate on any of this." Every answer is padded with reassurance.
The Caveman Skill swaps that personality out. It tells Claude: drop the filler, keep the substance. Answer like a caveman who happens to have a senior engineer's brain.
The trick is that none of the technical content disappears. The fix, the reasoning, the code, the quoted error messages. All intact. Only the padding dies.
Before and after on the same question
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. 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."
Caveman Claude — 19 tokens
"New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."
Same fix. Same reasoning. Roughly 72% fewer output tokens. Across a real session, that adds up fast.
Section 2
Install it in 60 seconds
Open a Claude Code session. Paste this. Done.
One command
That's the whole install. It pulls the plugin from the marketplace, wires up the auto-start hook, and adds the statusline badge so you can see when it's active.
Restart your session. From that point on, Claude Code will caveman-speak in every session by default until you turn it off.
Not using Claude Code?
The same skill works on Gemini CLI, Cursor, Windsurf, Cline, Copilot, and 40+ other agents. Installs are one line each. Full list on the GitHub repo.
Section 3
Turn it on. Turn it off.
After install, caveman auto-activates every session. You don't have to remember to turn it on. But you will want to know how to steer it.
Activate manually
Type /caveman or just say "talk like caveman". Either works.
Switch intensity
Run /caveman lite, /caveman full, or /caveman ultra. The level sticks for the rest of the session or until you change it.
Turn it off
Say "stop caveman" or "normal mode". Claude snaps back to its default voice for the rest of the session.
Statusline badge
The plugin adds a [CAVEMAN] badge to your status bar so you always know if it's active. Shows the current level too, like [CAVEMAN:ULTRA].
Section 4
The 6 intensity levels
Three English levels plus three Classical Chinese (文言文) levels. Each one hits a different compression point. Same technical accuracy across all six.
Lite
Professional but no fluff
/caveman lite Sample output
"Your component re-renders because you create a new object reference each render. Inline object props fail shallow comparison every time. Wrap it in useMemo."
Use when
You want Claude to still sound like a person. No articles dropped. Filler cut. Grammar intact.
Full
Default caveman
/caveman full Sample output
"New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."
Use when
The daily driver. Drop articles, fragments OK, 75% token cut. What most people should run.
Ultra
Telegraphic grunt
/caveman ultra Sample output
"Inline obj prop → new ref → re-render. useMemo."
Use when
You know the domain cold and want the answer as a one-liner. Not for learning a new topic.
Wenyan-Lite
Semi-classical Chinese
/caveman wenyan-lite Sample output
"Grammar intact, filler gone. Light 文言文 compression."
Use when
You read Chinese and want even more compression than English caveman gives you.
Wenyan
Full classical Chinese
/caveman wenyan Sample output
"物出新參照，致重繪。useMemo Wrap之。"
Use when
Maximum classical terseness. Same technical answer, most token-efficient written language humans ever invented.
Wenyan-Ultra
Ancient scholar on a budget
/caveman wenyan-ultra Sample output
"Extreme classical compression. For when every token costs."
Use when
Novelty + maximum compression. Read Chinese, want the lowest token count possible.
Start on full. It's the default for a reason. Drop to lite if you find yourself asking Claude to re-explain things, push to ultra once you've built the muscle for reading telegraphic output.
Section 5
The 4 sub-skills you get for free
The main caveman mode is the headline, but the plugin also ships four targeted sub-skills. Each one applies caveman logic to a specific task.
caveman-commit
/caveman-commit Terse commit messages. Conventional Commits. ≤50 char subject. Why over what.
Example
caveman-review
/caveman-review One-line PR comments. Location, problem, fix. No throat-clearing.
Example
caveman-compress
/caveman:compress Compresses your CLAUDE.md into caveman format so Claude reads fewer tokens every session. Saves the original as a backup.
Example
caveman-help
/caveman-help Quick-reference card. All modes, skills, commands. One command.
Example
The quiet win here is caveman-compress. Your CLAUDE.md file loads on every session start. Compressing it cuts input tokens Claude reads by roughly 46% without losing any technical substance. Caveman makes Claude speak less. Compress makes Claude read less. Run both.
Section 6
Who this is for (and not for)
Install this if
- You run long Claude Code sessions and hit context limits mid-task
- You're paying for API usage and want to cut output-token costs
- You already know what Claude is about to say and don't need the preamble
- You want replies 2-3x faster because there are fewer tokens to generate
- Your PR reviews and commit messages have been getting out of hand
Skip this if
- You're brand new to Claude Code and learning what it can do
- Claude's output goes to non-technical readers (clients, team, stakeholders)
- Pair-programming flows where the narrative matters as much as the code
- You're using Claude to write blog posts, emails, or long-form copy
- You rely on Claude's explanations to teach yourself a new framework
The sanity check: caveman cuts the words, not the thinking. If you rely on Claude's full explanation to understand what it's doing, leave this alone. If you already know what you asked and just want the answer, turn it on today.
Section 7
Where caveman steps aside on its own
The skill is smart enough to not caveman-speak when clarity matters more than compression. It drops back to full sentences automatically for these cases:
Security warnings
If Claude is about to flag a credential in code or a risky permission, it writes the warning in full sentences so the risk is unambiguous.
Destructive operations
Anything that deletes files, drops a table, or force-pushes a branch gets explained normally. You need to understand the blast radius before confirming.
Multi-step sequences where fragment order matters
If shortening the answer would make the steps ambiguous, Claude writes them in order with full sentences.
Code, commits, PRs
The code Claude writes is never caveman-fied. Variable names, comments, commit bodies. All written normally. Only the conversation around the code gets compressed.
Section 8
My honest take after using it
I was skeptical. Cutting 75% of the output feels like it should cut 75% of the quality, and it doesn't. The outputs are still great. Same code, same reasoning, same fixes. I didn't have to re-ask a single question that I used to get in full paragraphs.
That was the surprise. Not that it compresses, but that the compression doesn't cost you anything in accuracy. The fluff really was fluff.
The second thing worth knowing: your first session back in normal mode after running caveman for a week feels weirdly slow. Like Claude is suddenly over-explaining everything. You don't miss the preamble once it's gone.
Install it. Leave it on for a few days. If you hate it, /caveman off and it's gone. But you probably won't.
Quick Reference
Every caveman command
/caveman /caveman lite /caveman full /caveman ultra /caveman wenyan-lite /caveman wenyan /caveman wenyan-ultra /caveman-commit /caveman-review /caveman:compress /caveman-help stop caveman normal mode Source and docs: github.com/JuliusBrussee/caveman. Built by Julius Brussee. MIT licensed.
Zero to Automated
Get the full AI operating system
Guides are a great start. But inside Zero to Automated, you get a done-for-you AI system deployed within 48 hours, plus everything you need to keep building:
- Claude Code 101 — 8 lessons, zero to automating real work by Sunday
- Ready-to-install Claude Skills — a growing library with 2+ new skills every week
- Done-for-you workflows — copy, paste, run. New ones every week.
- Weekly live builds — watch real projects get built, ask questions, follow along
- Small group coaching — direct access to both founders
Price locked for life at whatever you join at.