Labsco
Labsco/Repos/juliusbrussee/caveman
REPO PACKAGE

juliusbrussee caveman

COMMUNITY
juliusbrussee · publisher85,036 repository starsMIT · Freegithub.com/juliusbrussee/caveman
7skills
3groups
6ready to use

The package is caveman, a Claude Code / Codex / Gemini / Cursor skill (plus 30+ other agents via its own installer) that swaps an agent's normal explanatory prose for compressed "caveman-speak" while keeping code, commands and error text byte-for-byte exact. All 7 skills here belong to that one idea: caveman itself is the toggle, with six levels from lite to a classical-Chinese wenyan mode, sticky for the session; caveman-commit, caveman-compress and caveman-review apply the same compression to commit messages, memory files like CLAUDE.md, and PR review comments; caveman-stats reports the session's measured savings; and cavecrew is a set of three subagents (investigator, builder, reviewer) that delegate the same way to keep the main context window lighter.

This suits someone reading a lot of agent output during a session — coding work in Claude Code, Codex or Gemini CLI, where terse, code-accurate answers speed up iteration. It is not aimed at anyone who needs the agent's longer explanations for onboarding or an unfamiliar codebase, and levels above lite intentionally drop connective reasoning from the visible answer, not just filler words.

READ THE FULL ANALYSIS

The 65% figure is the author's own benchmark, and the README states its own limits in the same breath. It counts output tokens only; input and reasoning tokens are untouched, and the skill's instructions add roughly 1,000-1,500 input tokens per turn — so on an already-terse workload the net effect can go negative rather than positive. The 65% average itself comes from 10 prompts the author ran and committed to the repo's benchmarks/ folder, not from an independent test.

What you install here is one piece of a five-repository family, and only this piece. The README promotes caveman-code (a full terminal coding agent), cavemem, cavekit and a Gemma fine-tune called cavegemma, plus a separate caveman-shrink npm package and a second skills repository, JuliusBrussee/skills, with four unrelated skills of its own (grill-me, interface-kit, junior-to-senior, loop-factory). None of those are part of the 7 skills counted on this page.

Our source README was truncated at 12KB of a 14.9KB file, cut off partway through a "Caveman 2" section that describes a forthcoming, more rigorously measured successor as not yet shipped — today's /caveman-stats numbers remain the author's own local estimates.

7Grouped into 3 sets by the authors.
6Work with nothing else to set up.
85,036Stars on the GitHub repository, at last check.