If you’ve been using Claude or any other large language model for serious work, you’ve probably felt that sting when the token counter runs up faster than expected. Token costs add up, especially when you’re generating long outputs or handling complex conversations.
Someone on GitHub came up with a clever trick: make Claude sound like a primitive human. Short sentences, minimal vocabulary, no fluff. The result? A dramatic drop in token usage. The project is called something like “Claude Primitives” (I’ll link it below), and it’s basically a set of system prompts that strip down language.
Why does this work? Large models tend to produce verbose responses by default. They like to add context, rephrase, explain. But if you tell Claude to “speak like a caveman,” it will output things like “Me see bug. Fix now.” instead of “I’ve identified a software defect in the login module and will proceed to patch it using the following steps.” The latter uses maybe three times more tokens.
The project provides several prompt templates optimized for different tasks: code review, debugging, summarization, even creative writing. Each one enforces a specific level of brevity. For example, a “caveman” mode might only allow 3–5 words per sentence, while a “concise” mode still feels natural but cuts out redundancies.
I tested it with a sample code review. Normally, Claude would write something like “The variable name data is too generic and does not convey its purpose. Consider renaming it to userData for clarity.” With the primitives prompt, it became “Rename data to userData. Clearer.” That’s a solid 60% token reduction.
But there’s a trade-off. You lose nuance. If you need detailed explanations or creative flourishes, this isn’t for you. It’s best for internal tools, automated reports, or situations where speed and cost matter more than eloquence.
The repo is pretty straightforward. Clone it, pick a prompt file, and either prepend it to your conversation or set it as system prompt in the API. Works with Claude (Anthropic’s models) but should also work with other LLMs that support system messages. No dependencies, no installation. Just text files.
One thing I appreciate: the author didn’t over-engineer it. No fancy UI, no complex configurations. It’s a pragmatic solution to a real problem. And it’s open source, so you can tweak the prompts to fit your exact needs.
If you’re burning through tokens and looking for a quick win, give it a shot. You might save enough to cover a few more API calls.
[Link to the project – search “claude primitives” on GitHub]