Caveman Plugin Gains Popularity for Making LLMs Speak Like Cavemen to Save Tokens
Caveman Plugin Gains Popularity for Making LLMs Speak Like Cavemen to Save Tokens
The Caveman plugin optimizes LLM responses by eliminating unnecessary phrases, saving up to 75% of tokens without quality loss.
It configures the model to exclude greetings, polite phrases, lengthy transitions, definitions for beginners, and other introductory phrases, leaving only the essence.
According to the author, up to 75% of tokens can be saved without quality loss (there are various compression modes, from light to ultra-telegraphic).
The project on GitHub (github.com/JuliusBrussee/caveman) has already received nearly 80,000 stars. Initially, it was just a skill for Claude Code, but now the author has created a standalone plugin from Caveman and even released a full-fledged coding agent Caveman Code.
The author notes that the skill is used even by employees of OpenAI, Nvidia, GitHub, and others. At Legrand, employees are reportedly recommended to use Caveman to avoid exceeding token limits.
Interestingly, the project has been supported by OpenAI's CTO Shane Sweeney, who personally contributed to the repository by adding Codex support.
Why it matters
AnalysisCaveman showcases an innovative approach to optimizing token usage in LLMs, which can significantly reduce computational costs. It also highlights the importance of efficiency in AI solution development.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
MIT researchers show LLMs can read and edit neural network weights without training
MIT CSAIL researchers introduced Weight-Space Programming, where an LLM directly reads and modifies model weights; Astra improved Qwen3.5-0.8B math scores without gradient descent or fine-tuning.
Part I. AI Agent Architecture: The Foundations
The eleven foundational building blocks of modern autonomous agents: model taxonomy, the agentic runtime loop, advanced prompting, distillation, skills architecture, memory systems, function calling, and FinOps.
Part II. AI Agent Architecture: From Model to Memory and Tools
A deep engineering dive: tokenization, KV-cache, sampling determinism, the training stack, quantization, context engineering, the safety stack, production RAG, ReAct, and multi-agent swarm systems.
Frontier Models
The most powerful class of artificial intelligence at the forefront of global research (Claude 3.7 Sonnet, OpenAI o3/GPT-4.5, Gemini 2.0 Pro), defining the limits of modern reasoning, autonomy, and coding capabilities.