Caveman is a token-trimming skill for AI coding agents: a set of instructions that tells the model to drop filler and answer in short, clipped sentences. It is aimed at developers who pay for Claude Code or a similar agent by the token and who notice how much of the bill goes to preamble, restated questions and closing summaries nobody reads. The repository puts the saving at 65 percent of tokens. Its slogan — "why use many token when few do trick" — is also, more or less, its entire method.
What it does
Caveman does not change what your agent does, only how it writes. The README frames the problem plainly: you pay for AI by the token, and your agent writes like it knows that. The fix is to make the agent stop padding. Same brain, fewer words, smaller bill. The video built around the project shows the project's own before-and-after example — the same bug, the same fix, far fewer wasted words around it.
Beyond the core skill, the README advertises a few things through its badges:
- The skill is described as working with more than thirty agents, with details in a separate install document.
- A set of wrap profiles covers ten agents natively.
- The repository is licensed under a combination of MIT and BSL.
The scale is unusual. GitHub records about 102,900 stars and about 5,980 forks, and the project crossed 100,000 stars the week the video was made. It was created in April 2026 and was still being pushed to in September 2026, so this is a young repository that grew very fast rather than an old one coasting.
How it works
There is no model change here, no fine-tune and no proxy rewriting your traffic. Caveman is delivered as a skill — text the agent loads alongside your own prompt, which then shapes every reply it writes. That is why the savings land on the output side of the exchange: the agent still reads whatever context you hand it, but it stops spending tokens narrating what it is about to do and recapping what it just did.
The compatibility claim follows from the same design. Because the payload is instructions rather than an integration, it can be dropped into any agent that reads a skill or system prompt, which is how a single small project reaches thirty-odd tools. The wrap profiles appear to cover the cases where an agent needs to be launched or configured rather than simply handed a file. GitHub lists Go as the repository's main language, though the README excerpt available here does not spell out what that Go code covers.
Getting started
Installation is the short part. The README links a dedicated install document and the project is also listed on a skills directory, and the video reports that a single command is enough to get running. The recommended route is the project's own quickstart rather than any third-party write-up, since the instructions change as the skill evolves. The README's own navigation mirrors that path: see an example, install, check what you save, then read where your tokens actually go.
When to use it / when not
This is worth turning on if you run long agent sessions, pay for them yourself, and mostly want answers rather than essays. Debugging loops, repetitive refactors and any workflow where you already skim the agent's prose are the obvious fits.
It is a worse fit in a few situations. If the agent's output is going to other people — design documents, onboarding notes, code review comments a teammate reads — clipped caveman phrasing costs you the clarity you were paying for. If your spend is dominated by large context windows, huge file reads or long tool transcripts, trimming replies attacks the smaller half of the bill. And the 65 percent figure is the project's own number, presented alongside a before-and-after example, not an independent benchmark; treat it as an order of magnitude and measure your own usage before and after.
One more practical note: the license badge shows MIT combined with BSL rather than plain MIT, so anyone planning to bundle or resell this should read the terms rather than assume.
Caveman deserves attention from anyone whose team is running coding agents at volume and watching the invoice grow, and from anyone curious about how far a pure prompt-level change can move real costs. It is not infrastructure and it is not clever engineering; it is one instruction, packaged well, with an install path short enough that trying it costs almost nothing. Judge it by your next bill. If the terse style annoys you, uninstalling is as cheap as installing was, and that asymmetry — tiny cost to try, measurable saving if it sticks — is the real reason a joke about cavemen ended up with six figures of stars.