Get the full token-saving Agent OS setup inside the AI Profit BoardroomJoin
I ────── The Caveman Command Engine™
Tested on Fable 5 · July 2026
The Caveman Command Engine™ + Caveman.
Same brain. Smaller mouth. I tested it on Fable 5 — 69% fewer output tokens, 37% cheaper, and every answer was still right.
Here's the problem with the smartest models.
Fable 5 charges $50 for every million tokens it SAYS.
And by default, it says a lot. Long intros. Polite padding. Three paragraphs where one line would do.
Caveman is a free little skill that fixes the talking, not the thinking.
You install it once. Your agent answers like a smart caveman: short, blunt, and exactly right. The code stays byte-for-byte perfect.
I didn't take the README's word for it — I ran my own test against Fable 5. The receipts are below.
−69%output tokens (my real test)
−37%total cost on Fable 5
5/5answers still technically correct
30+agents it installs into
II ────── How it works, in simple words
Same brain. Smaller mouth.
Here's the whole idea, simple enough for anyone.
When an AI answers you, you pay for every word it says. Those are "output tokens" — and on the smartest models they're the most expensive thing you buy.
Caveman is a set of instructions your agent reads once, at the start of every chat.
The instructions say: drop the filler words. No "Sure! I'd be happy to help." No "basically" and "actually". No three-paragraph warm-ups.
Say the thing. Give the fix. Stop talking.
Three rules keep it safe:
1. Code is sacred. Code blocks, commands, file paths and error messages are never touched. Byte-for-byte exact.
2. It knows when to speak up. For security warnings and dangerous actions — like deleting a database — it automatically drops caveman-speak and writes a full, clear warning. Then goes back to grunting.
3. It speaks your language. Write in Portuguese, it grunts in Portuguese. It compresses the style, never the meaning.
The whole trick in one picture — the thinking never changes, only the talking shrinks.
"Won't short answers be dumber answers?"
The opposite, sometimes.
A March 2026 research paper tested 31 models and found forcing brief answers IMPROVED accuracy by up to 26 points on some benchmarks.
And in my own 5-prompt test, every caveman answer contained the same fix as the long version. Less word, same brain.
III ────── Exactly how it works, step by step
What Caveman actually is — and what happens on every message.
Let's kill the mystery completely. No magic here. Here is the exact mechanic, step by step.
Caveman is just a text file of rules. That's the whole secret. It's not an app. Not a new model. Not a middleman server your messages pass through. It's about 2,000 tokens of written instructions — a "skill" — that tell your agent HOW to speak. Nothing more.
The installer copies that file into your agents. When you run the one-line install, it looks at your machine, finds every AI agent you have — Claude Code, Codex, Gemini, Cursor, 30+ others — and drops the rules file into each one's skill or plugin folder, in the format that agent expects. That's why one command covers everything.
Your agent reads the rules at the start of every chat. AI agents always read their instruction files before answering you — that's how skills work. So from message one, the caveman rules are sitting in the model's context, right next to your question.
The rules tell it what to kill. The actual file says, almost word for word: drop articles (a, an, the), drop filler (just, really, basically, actually), drop pleasantries (sure, certainly, happy to help), drop hedging. Answer in the pattern: thing → action → reason → next step. Fragments are fine.
The rules tell it what to protect. Code blocks: unchanged. Commands: unchanged. File paths, API names, error messages: byte-for-byte exact. And no fake shorthand either — the rules ban made-up abbreviations like "cfg" or "impl", because the tokenizer splits those into the same number of tokens anyway. Zero saving, harder to read. Full word wins.
The model then WRITES short — it never edits after. This is the key point. Nothing trims the answer after the fact. The model simply chooses fewer, plainer words as it writes, because its instructions said to. The thinking is identical. Only the word choice changes. That's why nothing technical goes missing.
Every shorter reply is billed smaller. You pay per token the model outputs. A reply that would have been 1,349 tokens comes out as 324 — my real React test — and your bill shrinks by exactly that difference, on every single reply, forever.
The safety valve watches every turn. Before answering, the rules make it check: is this a security warning? An irreversible action? A multi-step sequence where dropped words could cause a misread? If yes, it switches to full clear sentences for that part, then goes back to caveman. Short never beats safe.
One message's journey — the rules ride along as input, the model writes short on the way out. No editing, no middleman, no magic.
So when someone asks "but what IS it?" — the honest one-line answer:
Caveman is a 2,000-token instruction file that rides along with every message and tells the model to skip the filler while writing.
The model was always capable of answering short. Nobody had ever told it to.
✦ ✦ ✦
IV ────── Straight from the source
The official sources. Read it and run it yourself.
From the repo itself: "Caveman no make brain smaller. Caveman make mouth smaller." — it shrinks what the agent says, not what it knows.
V ────── My story · why this matters
I was you. Then I put my agents on the caveman diet.
Before
I moved my hardest work onto Fable 5 — the best model I've ever used.
And then I watched it burn tokens explaining things I already understood.
Every answer opened with "Sure! I'd be happy to help you with that."
Every fix came wrapped in three paragraphs of padding.
At $50 per million output tokens, the padding was the expensive part.
Then I installed Caveman, and the padding died.
After
Now the same model answers in tight, blunt lines: "Bug in auth middleware. Fix:"
My test showed 69% fewer output tokens — with every answer still correct.
The total bill dropped 37%, even counting the skill's own overhead.
And honestly? The short answers are easier to read than the long ones ever were.
You can have this too. One command. Free.
VI ────── The receipts
Real people. Real wins. Inside the Boardroom right now.
3,600+ Founders inside AIPB
400k YouTube subscribers
38 Countries · live members
163k X / Twitter followers
29k+ Udemy students
I'm not going to paste invented quotes here.
The wins are real and written by the members themselves — agency owners, ecom founders, course creators, solo operators across 38 countries.
Read them in their own words.
You've seen the numbers above. They're from my own test, run this week, against the most expensive model I use.
The next ten minutes show exactly how it works and how to install it.
So here's the deal.
If you're reading this — promise yourself one thing right now. You're going to finish this guide AND run the one install command before you sleep tonight. Just one command. Because the moment your agents stop wasting words, every single reply you get for the rest of your life costs less and reads faster.
The people sitting still are paying full price for padding. The people implementing today are the ones who'll look back in six months and say "that was the moment."
Be one of those people.
Commit to the transition. Commit to taking action today. This changes everything about your workflow.
✦ ✦ ✦
VII ────── I tested it — here's the proof
My real test: 5 prompts, Fable 5, both ways.
I didn't trust the README. I ran the same 5 real questions against Fable 5 twice — once normal, once with the caveman skill loaded — and read the exact token counts off the API.
Prompt
Normal out
Caveman out
Saved
Cost: normal → caveman
React re-render bug
1,349
324
−75%
$0.068 → $0.037
Database connection pooling
1,655
473
−71%
$0.083 → $0.044
Python KeyError fix
1,144
390
−65%
$0.058 → $0.040
Git rebase vs merge
1,091
355
−67%
$0.055 → $0.038
Block AI crawlers (robots.txt)
1,052
415
−60%
$0.053 → $0.041
Total
6,291
1,957
−69%
$0.316 → $0.200
Same five questions. Same model. 69% fewer output tokens, 37% off the total bill — and that total already includes the skill's own overhead (about 2,000 extra input tokens per turn, which I'll explain below).
Here's what the answers actually looked like, on the git question:
Normal Fable 5 — 1,091 tokens
"# Git Rebase vs Merge — Both integrate changes from one branch into another, but they do it very differently. ## Merge … Creates a merge commit that ties the two branches together. History looks like…" (and on for pages)
full headings · diagrams · long walkthrough
Caveman Fable 5 — 355 tokens
"Merge: keep both branch histories, tie together with merge commit. History show true parallel work. Rebase: replay your commits on top of target branch, linear history, no merge commit. Merge pros: safe, never rewrite history, good for shared branches…"
same advice · same commands · third of the words
I also installed the plugin straight into Claude Code with one command — it's running inside my Agent OS right now:
claude plugin marketplace add JuliusBrussee/caveman
claude plugin install caveman@caveman
My real Fable 5 test — output tokens per prompt, measured from the API. Every bar shrank 60–75%.
"Did you check the short answers were still right?"
Yes — that was the whole point of testing.
All five caveman answers contained the same core fix, the same commands and the same warnings as the long versions.
The React answer still said useMemo. The git answer still gave the golden rule about shared branches. Nothing technical was lost — only the padding.
VIII ────── The Fable 5 math
Why this matters MOST on Fable 5.
Fable 5 has a lopsided price tag: $10 per million input tokens, but $50 per million output tokens.
The words it SAYS cost five times more than the words it READS.
So a tool that shrinks the saying — and only the saying — hits the expensive side of the bill.
Now the honest part, because there always is one.
The caveman skill itself is about 2,000 tokens of instructions, and the agent re-reads them every turn.
That's 2,000 extra INPUT tokens each reply — about 2 cents on Fable 5.
But in my test it saved roughly 870 OUTPUT tokens per reply — about 4.3 cents.
Cheap side up a little, expensive side down a lot. Net: 37% off the total.
One warning from the tool's own docs, which my testing confirms: if your workload is already terse — short answers, mostly code — the savings shrink and can even go negative.
I proved this on my own Agent OS memory file: it's already written tight (tables, commands, no filler), and caveman-compress could only shave 0.3% off it.
The tool's own tests show ~46% savings on wordy, prose-heavy memory files.
Lesson: caveman kills filler. If you already write like a caveman, you've already banked the savings.
Fable 5's lopsided pricing — output costs 5× input. Caveman trades a little of the cheap side for a lot of the expensive side.
"Doesn't running Agent OS burn a fortune in tokens?"
No — that's the biggest myth about it, and this guide is literally another layer of the answer.
Agent OS runs the everyday 90% on a free local model (on your own machine, nothing leaving it), free APIs slot in for more, and for the frontier work it drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, and Agent OS plugs straight into it, so you're not paying twice.
Now add Caveman on top and the paid replies themselves shrink 60–75%.
And inside the AI Profit Boardroom there are full token-optimisation tutorials, so you cut usage to the bone and never think about it again.
IX ────── The framework
The Caveman Command Engine™.
Four layers. Each one shrinks a different part of your AI bill — and together they run my whole Agent OS on a fraction of the tokens.
i.
The Mouth
You stop paying for padding — the caveman skill cuts what your agent says by 60–75%, with code kept byte-perfect. One install covers Claude Code, Codex, Gemini, Cursor and 30+ agents.
ii.
The Memory
You stop re-paying for wordy memory files — caveman-compress rewrites prose-heavy files like CLAUDE.md into tight caveman-speak (~46% smaller on the tool's own tests), and that saving repeats every single session. Write tight files from day one and you bank it forever.
iii.
The Crew
You stop burning context on chatty helpers — cavecrew subagents (investigator, builder, reviewer) run ~60% leaner than vanilla ones, and caveman-shrink squeezes the tool descriptions of any MCP server you plug in.
iv.
The Command
You put the whole thing under Agent OS command — free local models take the everyday 90%, the paid CLIs you already own take the frontier work, and Caveman makes those paid replies 37% cheaper. Every layer compounds.
The Caveman Command Engine™ — four layers of shrink, all commanded from one Agent OS dashboard.
X ────── Old way vs new way
Old way vs new way.
Old way$0.316 per 5 answers
Every reply opens with "Sure! I'd be happy to help you with that"
Three paragraphs of setup before the actual fix
You scroll and skim to find the one line that matters
Fable 5 bills you $50/M for every padded word
Memory files full of polite prose reload every session
Subagents chat away your main context window
New way$0.200 per 5 answers — same fixes
"Bug in auth middleware. Fix:" — answer first, zero warm-up
69% fewer output tokens on my real Fable 5 test
You read the reply in one glance — faster than skimming
Code, commands and errors stay byte-for-byte exact
One free install covers 30+ agents, on from message one
"I'm not technical. Can I actually install this?"
One command. Paste it in a terminal, press enter, done in 30 seconds.
It finds every AI agent on your machine by itself and installs for each one.
If anything breaks, you open your agent and literally tell it: "Read the caveman repo and install it for me." The agent fixes its own brain.
That's it. On Claude Code, Codex and Gemini it's on from the first message.
Say "normal mode" any time to switch it off, or /caveman to bring it back.
You can also pick how hard it grunts:
lite — drops the filler, keeps full sentences. Professional but tight. full (default) — "New object ref each render. Wrap in useMemo." Classic caveman. ultra — "Inline obj prop, new ref, re-render. useMemo." One word when one word enough.
And the bonus commands: /caveman-commit for tight commit messages, /caveman-review for one-line PR comments, /caveman-stats to see your lifetime savings in dollars, and /caveman-compress to shrink a memory file forever.
The grunt ladder — same advice at every level, the sentence just keeps shrinking.
"What if a short answer is dangerous — like deleting my database?"
Built in: auto-clarity.
For security warnings, irreversible actions and anything where compression could cause a misread, caveman automatically switches to full, clear sentences — writes the complete warning — then goes back to grunting after the dangerous part is done.
✦ ✦ ✦
Get the whole operating system
Want the full token-saving stack, not just the mouth?
Caveman shrinks the replies. The Agent Operating System shrinks the whole bill. When you join the AI Profit Boardroom you get everything:
The Caveman Command Engine setup — this guide's stack, installed and configured
Token-efficiency playbooks — the full tutorials on cutting usage to the bone
Free local models — the everyday 90% of work at $0 on your own machine
Every CLI you already pay for — Claude, Codex, Gemini, Kimi, GLM, Grok in one dashboard
Hermes Astros — the 24/7 YouTube competitor watcher that writes your titles
The Hermes Oracle — its sibling that watches X and drafts your posts
The Video Director — topic in, finished video out
The memory vault — an Obsidian brain your agents actually read
Agent Kanban — Planner → Builder → Reviewer teams that ship real work
3,600+ founders + me — daily tutorials, weekly calls, new tools added the week they ship
You're not buying a tool. You're getting the whole operating system I run a seven-figure business on — as a zip, with coaching calls where we set it up together.
Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Set up in an afternoon · used in 38 countries · new tools added every week
XII ────── Three beliefs to drop
What's holding you back.
Wrong: "Long, detailed answers mean the AI is being thorough."
Right: Long answers mostly mean padding. My test kept every fix, every command and every warning while cutting 69% of the words — and research shows brevity can actually IMPROVE accuracy. Thorough is about substance, not length.
Wrong: "Token costs are just the price of using the best models."
Right: Token costs are the price of using the best models CARELESSLY. Free local models for the everyday work, the CLIs you already pay for on the hard work, and caveman on the replies — same output quality, a fraction of the bill.
Wrong: "A tool claiming 65% savings is probably overselling it."
Right: Healthy instinct — so test it. I ran my own 5-prompt benchmark and got 69%. And the tool's own docs openly list where it LOSES (already-terse workloads) — my compress test on a tight file saved just 0.3%, exactly as they warn. Honest tools survive honest tests.
Don't take my word for it
158 pages of members who already broke through these exact beliefs. Their stories — real businesses, real wins — are documented here.
Run the one-line install.curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash — it finds every agent on your machine.
Open Claude Code and ask it anything. On Claude Code, Codex and Gemini, caveman is on from message one. Feel the difference immediately.
Pick your level./caveman lite if you want full sentences, /caveman ultra if you want grunts. Full is the sweet spot.
Test it on YOUR work. Ask the same question in normal mode and caveman mode. Check the fix survives. It will.
Compress your wordiest memory file./caveman-compress CLAUDE.md — prose-heavy files shrink ~46%, and it keeps a readable backup. Skip files that are already tables and commands (mine only saved 0.3% — tight files are already done).
Check your savings./caveman-stats shows real session tokens and lifetime dollars saved, right in the terminal.
Point your expensive model at it. The pricier the output tokens — Fable 5 at $50/M is the extreme — the more each shrunk reply is worth.
Stack it with the Agent OS. Free local models for the everyday 90%, your existing CLIs for the frontier work, caveman on every paid reply. That's the full Caveman Command Engine.
XIV ────── Recap
What you gain.
You stopped paying for padding. 69% fewer output tokens on my real Fable 5 test — the expensive side of the bill, cut by two thirds.
You stopped skimming. "Bug in auth middleware. Fix:" — the answer arrives first, and you read it in one glance.
You kept every fix. All five test answers stayed technically correct. Code, commands and errors are never touched.
You stayed safe. Auto-clarity drops the grunting for security warnings and dangerous actions, automatically.
You shrank your memory forever. Wordy memory files compress ~46% once, and that saving repeats every session after.
You covered every agent. One free install works across Claude Code, Codex, Gemini, Cursor and 30+ others.
You know the honest limits. Already-terse workloads save little (my tight file: 0.3%) — the wins come from killing filler, and now you know where the filler lives.
You commanded the whole engine. Free local models + the CLIs you already own + caveman on top — the full stack, one dashboard.
"Why use many token when few do trick?"
Your move
Six months from now, this is normal.
The people who wire in token efficiency today are the ones running ten agents for the price of one next year — while everyone else stares at their usage dashboard wondering where the money went.
The Agent OS gives you the whole system: the Caveman Command Engine from this guide, the free local models that handle the everyday 90%, every CLI you already pay for in one dashboard, Hermes Astros watching YouTube, the Oracle watching X, the Video Director, the Kanban teams and the memory vault — one zip file, with coaching calls where we set it up together, step by step.
Daily tutorials. A 30-day roadmap. 3,600+ founders across 38 countries, someone online whenever you get stuck.
And every new tool that ships — like this one — I test it, I break it, and I add what works to the OS the same week.
