caveman: Why Use Many Token When Few Token Do Trick
A Claude Code skill called caveman is at 103k GitHub stars, +545 today, on one joke executed with complete seriousness: make the agent talk like a caveman, save 65% of output tokens. The tagline is the whole product philosophy — why use many token when few token do trick.
Under the joke is real engineering. The skill itself is MIT-licensed instructions that push the agent into terse communication patterns; measured across ten benchmark tasks with real Claude API calls, output tokens dropped 65%, with individual tasks ranging from 22% to 87%. Then there's a proxy layer (BSL-1.1, converting to Apache later) that compresses data before the agent reads it, keeping originals recoverable locally — 33.2% fewer provider-reported input tokens in testing. The docs are honest that full-session savings run lower once you count rule overhead, which is more candor than most benchmark pages manage.
The reason this lands now: the frontier labs just spent a week competing on exactly this axis. Fable 5.1's cache-price cut, Gemini 3.8's promo pricing, Muse Spark bragging about 20% fewer tool calls — the cost of an agent loop is the new battleground, and caveman is the community's entry: don't wait for the lab to discount the tokens, just emit fewer of them. Verbosity was never intelligence.
Repo at https://github.com/JuliusBrussee/caveman.
← Back to all articles
Under the joke is real engineering. The skill itself is MIT-licensed instructions that push the agent into terse communication patterns; measured across ten benchmark tasks with real Claude API calls, output tokens dropped 65%, with individual tasks ranging from 22% to 87%. Then there's a proxy layer (BSL-1.1, converting to Apache later) that compresses data before the agent reads it, keeping originals recoverable locally — 33.2% fewer provider-reported input tokens in testing. The docs are honest that full-session savings run lower once you count rule overhead, which is more candor than most benchmark pages manage.
The reason this lands now: the frontier labs just spent a week competing on exactly this axis. Fable 5.1's cache-price cut, Gemini 3.8's promo pricing, Muse Spark bragging about 20% fewer tool calls — the cost of an agent loop is the new battleground, and caveman is the community's entry: don't wait for the lab to discount the tokens, just emit fewer of them. Verbosity was never intelligence.
Repo at https://github.com/JuliusBrussee/caveman.
Comments