@sourcetms: Me make a video. Me share with you. Me make skill. You install. You saved token. Thank you, Mr. Reddit. Comment caveman to get the skill. #ai #Tech #claude

tarhonyas.videok :

Ahh yes lets include a 2 page long system prompt in every prompt to save token use

Vijay K :

Why waste time say lot word when few word do trick?

5uper5hadowken :

“Please call the cavemnan agent and get their take on it” 😁

InternetJudge :

even better would be if we communicate with the AI in short hand for complex ideas like..." i want you to loop through this list of numbers" . Something like " for num in nums:"

sergio :

can we go lower than a caveman and have it talk like kevin from the office?

tony tells stories :

imagine, you've got a beautiful prompt and claude just says: no. 😂

Code :

I mean, caveman code can’t be any worse than what some of my coworkers produce…

Tired.Panda 🇨🇦 :

I have a great idea! why not communicate with it using thoroughly designed in a simplified logical syntax that transmits clear, organized and exact statements of what has to be done so the machine cannot make mistakes and... wait...

nick_shiva :

Bad idea! 1) The system instructions themselves and the whole conversation back and forth will be counted as token usage! 2) If you force LLM to give you an answer in a few tokens, you reduce the intelligence! The reasoning process is actually a hack that allows LLM to generate more tokens to “have more processing time” to reach a better response. Think of it as LLM thinks during token generation, and if you severely limit it, for example by saying, give me the final answer in your first sentence and also turn off the reasoning, you make the frontier model act like a local small model! So why bother? Downgrade the model to a cheaper one

Thomas - sourcetms | AI / Tech :

Guide and skill link: https://open.substack.com/pub/sourcetms/p/caveman-mode-for-claude?r=5frd5h&utm_medium=ios

R Jay :

bro you get infinite chats... Just start a new chat with percise context to continue. You shouldnt be completing an entire project on 1 chat 😂

Alex :

I tried something similar, but more extreme. We built essentially a DSL for the system we were working on, and then Claude could both respond to me in stuff like "HAT CTL DONE, NXT?". it was like talking in COBOL or something. it wasn't very useful. But it got interesting when I started multi agent teams, then they could communicate in the DSL, cutting down on loads of tokens. I suspect it could be an effective way of driving down token costs on long running agent team type activities, but it's almost impossible when a human is in the loop.

Fed :

caveman. me want skill

Omaiirr🏴 :

Api no work get new api😭

Marcin AI :

I been using this inside ChatGPT and Llms for a while but I never added this to my Claude code and Openclaw. Thanks!

TONY • Following :

Finaly i took 13 min break...

Steve :

My minimalistic traffic reduction prompt for ChatGPT: 1. No unnecessary small talk: Direct, structured answers without emojis, greetings, or closing remarks. 2. Personalization only when relevant: Use existing contextual information only when it is factually appropriate. 3. No discourse-driving topic extensions with suggestive follow-up prompts.

marian5331 :

cavemen

Jeidoz :

Lol. The next level is asian languages which are compact 😅

Jarrod Lee :

me use this. other ai's. save tokens.

dywa2fn9fjy6 :

wait when you ear that Chinese use less tokens

Depressed Nightmare Goblin Man :

i don't need AI because i can just use my brain, costs no tokens.

Allfather :

This is the best thing I've seen for AI in like 6 months. well done

Malcolm 🇦🇺 :

Claude keep token. Me think for self for free.😎

michael dejulia :