Why Use Lot Word When Few Word Save Token

We talk about the moment AI coding gets expensive enough to change how we build, from blown token limits to the stress of watching usage climb. We dig into a strange but smart idea: making Claude Code respond in “caveman” language to cut token costs while deciding when that simplicity backfires.

• setting token limits then regularly exceeding them while chasing “one more feature”
• how token pricing pressures choices for indie developers and pet projects
• Caveman as a Claude Code add-on that strips filler language to reduce output tokens
• keeping technical terms intact while removing boilerplate and flowery explanations
• when terse outputs help with obvious front-end bugs
• why QA and hard-to-reproduce cloud issues often need more context than minimal replies
• long-term token budgeting strategies for multi-year builds and maintenance

Josh

Jenna

Jack 

Randy 

Transcript

Hitting The Token Limit

SPEAKER_00 0:00

Gave myself a limit. Real that limit. Yeah. Oh. What's your limit? Every month I blow past that limit. And I'm like, I just need to add this one feature. It's critical to the core offering. It's high value. We should do it. And then I I blast it out thinking, okay, it's gonna think for 10 minutes or it's gonna work for 10 minutes. And you know, we'll be good. 40 minutes later, I'm like, what are you doing? Stop working. Like it just stop. Stop. And I'm watching it and I'm nervous and I'm watching just the meter go. And you gotta make decisions, man. It's forcing us to make choices. But that's neither here nor there. It has nothing to do with mythos, but I am so jealous that your work pays for your tokens because I'm doing it all out of pocket. Poor Randy was doing it out of pocket when tokens were like, I don't know, a million dollars a token. And it's been there since the beginning.

SPEAKER_01 1:06

Yeah, it's mental. I think that this whole thing is really just a push for more verbose token usage, honestly.

Caveman Mode For Claude Code

SPEAKER_01 1:16

It's funny that you bring that up because there's been sort of an open source, I don't really know if you want to call it a tool, but a quad code add-on called Caveman. Have you heard of this? I have not heard of Caveman. It is every single day that I'm reading AI news, I really think that I've seen it all in terms of how people are using this, abusing it, uh, maladapting it. And I really think this is the best part yet. So, Josh, have you ever watched the show uh uh The Office? Do you remember the episode where Kevin Malone, the accountant, starts speaking in sort of like archaic caveman language? Why use lot word when few words do trick in order to save time? Right? Do you remember this?

SPEAKER_00 2:01

Yeah.

SPEAKER_01 2:02

So somebody has built in a clawed code skill that in order to decrease their tokens and simplify their output, they run it through caveman to basically say, bug here, fix this, next step this, and effectively making it so that they remove all filler language so that their token output decreases. They're making their AI talk to them like a caveman to reduce their token costs.

How Token Cutting Works

SPEAKER_00 2:29

Okay, wait. I have so many questions. So if I talk to my AI like a caveman, is it using fewer tokens to process what I'm saying? That's part one. And part two is am I getting responses in a caveman way? I mean, that would make sense. They're using fewer words, so they're using fewer.

SPEAKER_01 2:51

That's what it is. It's it's the latter. Yeah. Okay.

SPEAKER_00 2:54

All right. So another question, follow-up. How does this apply to code?

SPEAKER_01 3:02

Exactly. So in the skill, there's a specific caveat that effectively says uh keep any sort of technical term like polymorphism uh or in invariance the same. It's all about how Claude will say, Oh my gosh, I I found the fix, and here it is. I had to change this. Whereas what it used to be was this, where now it's r removing all of the glazing and boilerplate veneer of it and saying fix, you were using greater than sign rather than greater than equals sign. And basically making it so that it's using as few words like it, it drops the and and and uh from all of the output in order to basically give you the minimum amount of information that you need to know that it's finding and solving the problem on each turn. And it's reducing token usage. It's crazy, right?

SPEAKER_00 4:05

Even if it reduces it by 10%.

SPEAKER_01 4:08

Yes, there's there's been multiple instances of it reducing something like a uh 2,000 token output to a hundred token output because it's just using uh like the arrow symbol to say X to Y, right? And it's keeping it that simple. So while you're abstracting away some of the context that you as a developer might want to know about what the bugs or changes that you're making in in your code base, you can cut down on a lot of the excessive flowery language that it adds, which is funny because based on Claude Code's leaked prompt for its agent harness, it already has told it to, instead of using three sentences to explain something, use one concise sentence. And yet apparently this is still too flowery for people, which I find to be hilarious.

SPEAKER_00 5:04

Webhook not work. Try again. Yes, yeah, exactly that. Webhook fail, try another thing. I don't know, but that makes sense because every little bit

Tradeoffs For QA And Debugging

SPEAKER_00 5:17

counts. We're entering this era where a good chunk of us are gonna be working on our pet project and the thing that is just our music that we're trying to get out. And I can see us to save cost, you know, you're not gonna do it all in one month, right? And it's not because it's the AI is incapable, it's because you're gonna run out of money. You got other stuff to do, right? Unless you're venture-backed or your company's paying for your tokens. Otherwise, most of us are gonna turn these into long-term projects where you know every paycheck that comes in, we're gonna put aside X amount of dollars, and we're gonna do this over two or three or four years. Some people will. I mean, even if you have not much money, even if you're making 30 grand a year, you can do that just by virtue of letting time work for you, right? Just to hope that in two or three years somebody doesn't hasn't already stolen your idea. But theoretically it can work. In that case, that kind of efficiency and cost cutting, even if you're cutting filler words over a one-year, two-year, three-year period, can really add up. Even for the people who are going to be maintaining their software for the next 10, 15 years, if you start doing that now, making your agent talk to you like a caveman, that's a hell of a lot of uh savings over a couple of years. So I'm all for it. I think it's cool. I'll try it.

SPEAKER_01 6:45

Have you tried it? I tried it for about 10 minutes, and I don't know. In in my role as QA, I need to be able to see a lot of the underpinning that AI already leaves out.

Long-Term Savings For Builders

SPEAKER_01 6:59

I think that if I was to use this in something like fixing a front-end bug manually or uh fixing like a very obvious side effect with very low necessity for nuance, I think I would easily use this. Um especially when looking at front-end issues or something that is uh pretty low complexity. When you start dealing with issues in environments that are very hard to reproduce, like cloud based issues, needing to comb through logs, I don't think that this is maybe the skill for you if you like to try to infer your own output from the AI. But that being said, if you know what the issue is, you know where it lives, why waste slot word when few word do trick to solve it?