caveman
Make your agent stop being polite: prose gets cut, code doesn't — 8.5% fewer output tokens in independent testing
Install
npx skills add https://github.com/juliusbrussee/caveman --skill cavemanOverview
A viral skill by solo developer Julius Brussee (★108K; skill is MIT, engine BSL-1.1), billed on the tagline "why use many token when few do trick" and #1 on GitHub Trending in July 2026. Once installed, the agent answers in the shortest sentence that still lands: code, commands, file paths and raw error messages are never touched, and security warnings or "are you sure?" prompts drop back to full sentences on their own. Three intensities — lite / full / ultra — plus a classical Chinese mode. The repo publishes all three measurements rather than one: JetBrains ran a paired A/B over 86 real coding tasks and found 8.5% fewer output tokens, roughly 10% cost, with no detectable quality change (sign test p = 0.82), noting that is the ceiling rather than the usual case — the advertised 65% is a different claim; Adobe Research's CAVEWOMAN paper measured 1.4–2.4× lower realized cost per model, up to 3×; the repo's own eval, against an "Answer concisely" control, is 50% fewer output tokens at the median. The same repo ships a local proxy that shrinks what the agent reads — logs, test output, diffs — saving 33.2% of input tokens across a pinned 54-run benchmark, where one HTML case came in at +9.9% and the author left that row in the table, red.
Key features
- Six intensity levels with instant exit via /caveman off; code, commands and error text are never compressed
- Full command set: /caveman-commit for one-line Conventional Commits, /caveman-review for one-line findings, /caveman-compress cuts memory files 46% on average
- cavecrew subagent presets (investigator / builder / reviewer) plus six work patterns — write less code, not just fewer words
- The big rock is a proxy between your agent and the provider: it compresses the read stream, 15,704 → 121 tokens on one browser-page case, byte-recoverable
Use cases
- Long sessions where logs, test output and diffs keep re-entering context on per-token billing
- Done with three paragraphs of throat-clearing before the actual answer
- CLAUDE.md / AGENTS.md growing out of control — /caveman-compress shrinks them with the original backed up
Only pays off on per-token billing: per-request pricing (Copilot premium requests, for one) saves nothing, and the ruleset rides along as ~1,000 input tokens per call, which can go negative on terse Q&A. The author's own advice: run the same task with and without it and compare the provider's billing page.
