Summary created by Smart Answers AI
In summary:
- PCWorld reports that Claude AI users are adopting “caveman” prompting techniques to reduce token consumption by stripping filler words and articles from responses.
- This method can dramatically cut output tokens, with one neural network explanation dropping from 460 to just 80 tokens using simplified language.
- While effective for coding tasks, debates continue over whether caveman prompting compromises response quality for nuanced explanations.
It’s no secret that Claude gobbles up tokens like a Corvette guzzles gas—and just like gas, tokens cost money. That’s why the heaviest Claude users are always looking for ways to save on token usage. The latest trick? A wild and controversial new method: speaking like a caveman.
Here’s how it works: Using either a specialized system prompt or a full-on plug-in, users are instructing Claude to talk like a caveman, stripping away any and all fluff such as preambles, praise, asides, filler words (like “very” and “really”), redundancies, and even articles (“a,” “the”).
One of the many suggested “caveman” system prompts floating around Reddit include this one: “Speak primitive. Use nouns and verbs. No grammar filler (the, is, are, of). Keep words short. Save tokens. Be blunt.”
The results will vary wildly depending on the exact prompt you use. I tried a variation of that prompt, adding the specific articles to avoid while also asking it to chop the pleasantries and filler words. What I got from Claude wasn’t exactly true caveman speak, although I’m sure you could get that with proper prompting. Instead, I got super-clipped responses that were devoid of idle chit-chat and digressions.
For example, when asking a caveman-prompted Claude to explain the basics of neural networks, it came back with this:
Neural networks: layers of connected nodes. Input layer takes data. Hidden layers find patterns. Output layer gives result.
Nodes pass signals, weighted by importance. Training adjusts weights using backpropagation — minimize error, repeat millions of times.
Key types: CNNs (images), RNNs (sequences), Transformers (language). Modern LLMs are transformers at massive scale.
That’s a 50-word response, which comes out to roughly 80 tokens. (There are more tokens than words because an AI model must break down larger words, punctuation, and emojis into smaller units before processing them, and each of those units counts as a token.) Compare that to the 330-word, 460-token response from Claude when I didn’t use the caveman prompt.
So, case closed, right? Let’s all use caveman mode! Well, not quite.
Ever since the caveman prompt began gaining traction earlier this month, it has spawned heated debates on Reddit and elsewhere. Some users argue that this caveman method leads to dumbed-down responses. Others note that any token savings are only for the output tokens, not the tokens you send to Claude. Also, an elaborate “caveman” system prompt will cost you every time you send a new prompt.
It’s also possible that the Claude caveman method isn’t a jack-of-all-trades. While it could be well-suited for coding, a task where you’re typically expecting a cut-and-dried response, it’s probably not the best fit for “explainer” prompts that require more detail or nuance, such as my prior “tell me about neural networks” request.
Still, the Claude caveman method is another example of how users are getting creative about boosting their AI token efficiency, an issue that’s looming for even casual AI users as more and more powerful agentic tools (like Claude Cowork) hit the mainstream.
Indeed, the makers of the more popular Claude caveman plug-ins are quickly finding ways to stretch their Claude token use without sacrificing quality in the bargain.