The AI That Finally Learned to Shut Up
AI coding assistants are incredible tools, but they have a verbosity problem. They talk like they're trying to hit a word count on a high school essay, burying the critical line of code you need under a pile of conversational fluff and apologetic preambles. We've all seen it: "Certainly! I'd be happy to help you with that. The issue you're encountering is a common one..." Just give me the fix. Julius Brussee's caveman is the answer to that prayer. It's a simple, brilliant plugin for over 30 AI agents that makes them talk like, well, a caveman. The project's motto says it all: "why use many token when few do trick".
Caveman solves the verbosity problem by stripping away the filler, leaving only the dense, technical core of the answer. It's not a different AI; it's a filter that makes your existing AI's mouth smaller, not its brain. The result is a dramatic reduction in output tokens—the project claims an average of 65%—which translates to faster answers and lower API bills.
Less Talk, More Rock
The genius of Caveman lies in its simplicity and effectiveness. After a single-command installation that auto-detects the agents on your machine, you can toggle the mode with a simple /caveman command. The effect is immediate. A long-winded explanation about a React re-render bug becomes a terse, scannable directive: "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo." All the technical accuracy is preserved, but the token count plummets from 69 to 19.
This isn't just a party trick; the project backs it up with reproducible benchmarks. Across ten common developer tasks, from fixing an auth bug to implementing a React error boundary, Caveman consistently slashes output tokens, with savings ranging from 22% to a staggering 87%. For teams running hundreds of thousands of agent interactions a month, these savings can be substantial.
What's particularly impressive is the project's intellectual honesty. A prominent "Honest number warning" in the README clarifies that Caveman isn't a silver bullet for cost savings. The skill itself adds about 1-1.5k tokens to the input on every turn, and it only compresses output. This means for very short conversations, you could end up with a net-negative token saving. The author rightly points out that the real win is "readability and speed," with cost savings as a bonus. This transparency is rare and builds immense trust in the project.
Beyond the core compression, Caveman offers a surprisingly deep ecosystem of tools. You can choose from different levels of conciseness, from lite to ultra, and even a wenyan mode that uses classical Chinese for maximum information density. The tool also provides specialized commands like /caveman-commit for generating conventional commit messages and /caveman-review for creating one-line PR review comments. There's even a middleware package, caveman-shrink, to compress tool descriptions for other AI agents, and a suite of cavecrew-* subagents that operate with the same token-frugal philosophy. This isn't just a prompt; it's a whole workflow built around the idea of brevity.
Where the Cave Leaks
Despite its brilliance, Caveman isn't for everyone. The most obvious group who should steer clear are junior developers or anyone learning a new technology. The verbose, step-by-step explanations from a standard AI agent are a feature, not a bug, for learners. Caveman strips away that helpful context, which could leave beginners more confused than enlightened. The terse style is also ill-suited for any situation where an AI is used to generate client-facing text, documentation, or formal reports. "Bug in auth middleware. Fix:" is great for a developer, but not for an email to your boss.
The project's popularity is also a double-edged sword. With over 95,000 stars, it's clearly struck a chord, but it also has 456 open issues. While the release cadence is decent (the latest was about a month ago), it appears to be largely the work of a single developer. This creates a potential maintenance bottleneck. AI agents and their plugin ecosystems are constantly changing, and a tool like Caveman, which sits as a layer on top of them, is vulnerable to breaking changes from upstream. Users are betting on the author's continued passion and availability to keep the project compatible with a rapidly evolving landscape.
Finally, the "Honest Number" caveat is a real limitation to consider. If your workflow involves many rapid-fire, short questions, the overhead from Caveman's own prompt could easily outweigh the output savings. It shines brightest on complex queries that typically generate long, multi-paragraph answers.
The Verdict
So, should you use Caveman? For any experienced developer who uses AI coding assistants daily, the answer is an enthusiastic yes. It's a tool that respects your time and attention. It makes AI interactions faster to read, less distracting, and cheaper to run. The setup is trivial, the usability is excellent, and the author's transparency about its limitations makes it easy to trust.
Caveman is more than just a cost-saving utility; it's an opinionated statement about how human-AI interaction should work. It trades manufactured conversational politeness for raw, efficient information transfer. It won't replace the need for detailed explanations when you're learning, and it's not a magic wand for your cloud bill in every scenario. But for the daily grind of coding, debugging, and reviewing, it turns your chatty AI partner into the silent, efficient professional you always wanted it to be. It's a sharp, focused tool that does one thing exceptionally well. Oog.