What Caveman Is and the Problem It Addresses
Caveman is an open-source project designed to reduce the number of tokens used when interacting with large language models, particularly in coding agent workflows. It functions as both a communication style guide and a technical proxy that modifies how AI agents generate and process text. The core idea is that by using a constrained, minimal phrasing style—inspired by controlled languages used in technical documentation—developers can achieve the same functional outcomes with significantly fewer input and output tokens. This matters because token usage directly affects cost, speed, and efficiency when using paid LLM APIs or running local models with limited context windows.
The project emerged from observations that verbose AI responses, while polite or explanatory, often contain redundant phrasing that does not add functional value in technical interactions. For example, when an AI explains a coding issue, much of the text may serve conversational tone rather than convey essential information. Caveman seeks to strip away this excess while preserving critical details like code snippets, numbers, file paths, and negations.
Why It Is Gaining Traction
Caveman has attracted attention from developers and researchers interested in prompt engineering and efficient LLM usage. Its rise coincides with growing awareness of token costs in production AI applications, especially as teams scale the use of coding agents like Claude Code or similar tools. The project gained visibility after being highlighted on platforms such as Hacker News and GitHub Trending, where it ranked highly in daily activity. External validation came from organizations like Adobe Research, JetBrains, and Elasticsearch Labs, which explored or referenced the approach in their own work. These citations suggest that the token-reduction claims are being tested in real-world or semi-formal settings, contributing to developer curiosity.
The appeal lies in its simplicity: rather than requiring complex retraining or fine-tuning of models, Caveman operates as a lightweight layer—either as a proxy that processes agent interactions or as a skill that guides how agents phrase their responses. This makes it easy to experiment with without altering underlying model infrastructure.
How It Works at a High Level
Caveman consists of two main components: a proxy (referred to as the “big rock”) and a communication skill (the “small rock”). The proxy sits between the user and the LLM agent, intercepting and modifying communication to reduce input token load. According to the project documentation, using the proxy results in 33.2% fewer input tokens across full sessions while maintaining answer correctness. This is achieved by preprocessing prompts to remove unnecessary phrasing before they reach the model, while preserving essential elements like code, numbers, and technical terms.
The skill component governs how the AI agent formulates its responses. It enforces a strict style: answers must lead with the core fact, avoid greetings or filler, limit sentences to one idea each, and retain all meaningful content such as negations, units, and code. The style draws from ASD-STE100, a controlled English standard used in aviation maintenance, which emphasizes clarity and brevity. Crucially, the project states that if Caveman phrasing does not reduce token count, the system defaults to plain language—ensuring no performance loss.
All original user prompts and code comments are preserved verbatim. The system checks each response before sending, removing only non-essential framing (like introductory phrases or summaries) while keeping every negation, number, and path intact. This ensures that informational fidelity is maintained even as verbosity drops.
Who It Is For and Real-World Use Cases
Caveman is aimed at developers and technical teams who use AI agents for coding, debugging, or technical documentation and are concerned with token efficiency. It is particularly relevant in scenarios where agents are invoked frequently—such as in automated code review, IDE-integrated assistants, or agent-based workflows—where small savings per interaction accumulate into meaningful reductions in cost and latency.
Use cases include reducing the token footprint of AI-generated explanations in pull request comments, minimizing input size when agents read large codebases or logs, and lowering costs when using paid LLM APIs at scale. The project notes that output costs can be 1.4 to 2.4× lower with Caveman-style responses, based on external testing across eight models and five datasets cited by Adobe Research. Elasticsearch Labs reported a 63.6% reduction in response tokens when adapting the approach for their MCP scenarios, with claims of zero information loss.
Because it requires Node.js 22.13+ and integrates with Claude-based workflows, it is most accessible to developers already operating in JavaScript/TypeScript or Node.js environments who have access to Claude Pro or Max tiers.
Caveats and Current Limitations
As an open-source project primarily driven by individual contributors, Caveman should be evaluated as an early-stage experiment rather than a production-guaranteed tool. The documentation does not specify long-term maintenance plans, formal support channels, or licensing details beyond its public GitHub presence. While external organizations have tested or referenced the approach, the project itself does not claim formal partnerships or endorsements.
The token reduction figures are based on specific test setups—such as 54 Claude Code runs with 18 of 18 answers correct—and may vary depending on the model, task complexity, or prompting style. The project emphasizes that results are not rounded up and that red rows in their tables indicate areas requiring caution, though the exact meaning of these indicators is not fully explained in the README.
Users should also consider that enforcing a strict communication style may not suit all interaction types. For instance, in cases where user empathy, clarification, or multi-step guidance is valued, the terse Caveman style might feel abrupt or insufficient. The project acknowledges this by limiting Caveman phrasing to cases where it actually reduces tokens, falling back to plain language otherwise.
Finally, while the project draws inspiration from established standards like ASD-STE100, it does not claim to be a certified implementation of that standard, nor does it guarantee compliance with any formal linguistic or safety framework.
Source: https://github.com/JuliusBrussee/caveman
