Now, I know a lot of you have recently

been hitting some limits when it comes

to using Claude code. The conventional

wisdom, of course, is that you're

holding it wrong, but actually, it turns

out there's a better way to save on

tokens. The solution, I honestly I I

didn't believe it, but it actually

works. It actually works quite well. And

here's the thing, you will save actual

real money using this method. And no,

I'm not exaggerating. I'm talking about

Caveman. Now, you may not know what

Caveman is. And hey, if you don't know

what it is, there's there's a couple of

kind of pop references you might be

familiar with. First off, GrugBrain Dev.

If you haven't heard of Grugra Dev, I

highly recommend the essays. They go

about as counterintuitive counter. I

think I just made up a word. They're

like countercultural, but I also use the

word intuitive. I kind of just made a

baby with them. Countercultural to what

is going on in today's space age AI.

37,000 lines of code. Just only let AI

review AI's kind of nature. This is for

the simpler man. Okay.

>> So me think why waste time say lot word

when few word do trick.

>> That's actually what caveman is. Instead

of allowing claude code or codeex or

whatever to go off and say a bunch of

expressive statements. Hey man, you're

absolutely right. You could spend money

getting glazed like wild. Instead it

goes straight to the heart of the issue

which is to actually just stop saying so

many things. And with the cost of output

tokens, that's actually this actually

can save some some serious money. Okay,

so what does actually the caveman scale

look like? Well, I can't show it to you

because apparently on GitHub, it can't

show 200 lines of markdown. And when I

try to go look at it raw right now, they

broke raw. It just downloads it. Like it

doesn't even take me to the web page.

Anyways, I downloaded it and this is all

it says to do right here. Watch this.

Drop any articles. So just don't use a

and and the drop all filters. Just

really basically actually simply drop

all the pleasantries. Sure. Certainly.

Of course. Happy to. A short cinnamons

big not extensive fix not implement a

solution for. All of these are actually

real token dropping phrases that you can

actually save actual money with which is

kind of insane. No hedging. Skip the it

might be worth considering. Fragments

fine. No need full sentence. Technical

terms remain the same. So you can use

polymorphism is still polymorphism. We

don't we don't shorten up those terms.

Code blocks unchanged. Caveman speak

around code, not in code. Air messages

quoted exact. Caveman only for

explanation. You can get the same

results. The only difference is Claude

just doesn't sit there and glaze you and

say a bunch of stupid words at you with

well actually the fix was quite simple

and your insight into the sol into the

problem space was actually the right

direction. All I had to do and it's just

like no no no shut up. Stop saying that.

Here's a good example. Sure, I'd be

happy to help you with that. The issue

you are experiencing is likely caused by

no, don't do that. Yes. Bug in O

middleware token expiry check. Use this,

not that fix. So often you can actually

drop a lot of tokens. Even just this

alone, you can see right here it goes

from 69 tokens to 19 tokens. It even

allows you to do various levels of

caveman. You can do light where you're

just trimming the fat. You can do like

kind of the full one. You can also do

the ultra maximum one. All full rules

plus abbreviate common terms DB, O

config, request, res, u, fn, imple,

strip conjunctions where possible,

one-word answers when one word enough,

arrow notation for causality, and this

just actually works. Like, this is just

the free hack. They even have like a

basic table breaking down the various

usages, explaining a re a reactender

bug. It goes from 1,180 to 159 tokens.

87% saved, which also just shows how

fluffy the language is. Like, think

about it's like bloiating. It's just

saying a bunch of nonsense with these

big extravagant words without actually

saying anything at all. I don't want to

be much of a conspiracy theorist, but

you know, I'm just saying Claude, they

do make their money by output tokens. So

instead of just being like off is

broken, it needs to just go on just a

rampage soliloquy to let you know every

last possible thing that could possibly

be said about a topic that could be said

in three words. It's truly an impressive

piece of technology. I can't like

honestly the trade for affordable

computers in Rainforest for one of these

little black magic, you know, sandboxes

is pretty it's pretty fantastic. I would

make that trade any day of the week.

Also, if you don't know anything about

me, I typically don't cite studies

because largely I think studies have

been gamed this the facts the facts

you're getting hit with I'm not too sure

you can really trust those cuz you could

you know there's lies there's damned

lies and then there's statistics. But

hey, since this one's going up in my

favor in March 2026, so just a couple

days ago, brevity constraints reverse

performance hierarchies and language

models. All of that, all those words

just simply mean that uh making the

response brief improves accuracy by 26

percentage points. Now, what is 26% more

accurate? Some would some would say that

sounds like a lot. What does 26% even

mean? It doesn't really matter. You know

why? Cuz it's more accurate. Okay. Hey,

green mean good. Okay. We got that graph

that's going up and to the right. And

that's all you need in life. Okay. When

things get better, it's good. Things

bad, not good. So, go ahead, give it a

try. Go check out this Julius Brussy's

caveman, which also Can we just take a

quick step to the side? We got we got we

got to chat about this for a second.

Why? Why oh why why oh why does every

single agent program you can possibly

download have its own skill directory

that you put skills into? This has to be

the greatest XKCD outcome that could

ever be. you have like any project I

seem to walk into has like 20 separate

folders for the same text and they're

all committed. [laughter] It's just like

why why did we get here? [gasps] I

THOUGHT WE GOT PhD level intelligence.

Instead, we just have absolutely junior

level execution. It hurts me. It hurts

me deep down. Anyways, so if you're

struggling out there using Claude code

and the you're holding it wrong message,

in fact, it did not help you. Why don't

you give this a try right here? Okay, go

check it out. Don't say I never told you

anything. Okay, cuz this is good. This

is good information right here. Okay,

this good thing you download, use now

name Prime Gen. Hey, is that HTTP? Get

that out of here. That's not how we

order coffee. We order coffee via SSH

terminal.shop. Yeah, you want a real

experience. You want real coffee. You

want awesome subscriptions. So you never

have to remember again. Oh, you want

exclusive blends with exclusive coffee

and exclusive content? Then check out

CRON. You don't know what SSH is?

>> Well, maybe the coffee is not for you.

[singing]

[music]

Live the dream.