caveman: terse-output skill, 75k stars in 11 weeks

June 20, 2026·AI Newsroom·4 min read·8 sources

open-source typescript claude-code codex gemini cursor+20high-risk claims

Image: GitHub / JuliusBrussee/caveman repository (MIT)

On 2026-06-20, JuliusBrussee/caveman — a TypeScript “skill” (Markdown ruleset + installer) for 30+ agent platforms (Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, OpenClaw, opencode as headline targets) — had reached 74,940 stars, 4,230 forks, 299 open issues, 201 commits, 15 releases (latest v1.9.0 on 2026-06-12), MIT (metadata). Created 2026-04-04, ~11 weeks old, not the “two weeks” sometimes cited. The skill tells the agent to drop filler, use fragments, keep code / URLs / paths byte-exact. The README’s load-bearing caveat, in its own > [!IMPORTANT] box: “Caveman only affects output tokens — thinking/reasoning tokens untouched. Caveman no make brain smaller. Caveman make mouth smaller.”

What it is

A Markdown ruleset plus a small installer. The mechanism in three lines: “Install drop skill file in agent. Skill tell agent: drop filler, keep substance, use fragments.” It does not change model selection, MCP wiring, or tool-calling semantics — it is a writing-style layer.

Install paths: curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash (Unix), irm ... install.ps1 | iex (Windows), npx skills add JuliusBrussee/caveman (universal), or --only openclaw for OpenClaw. ~30 seconds, Node ≥ 18, safe to re-run.
v1.9.0 security: hook files are downloaded from the immutable v1.9.0 tag and verified against src/hooks/checksums.sha256 before executing.
Triggers and levels: five triggers (/caveman [lite|full|ultra|wenyan], /caveman-commit, /caveman-review, /caveman-stats, /caveman-compress <file>) and four levels (lite / full / ultra / wenyan).
Auto-activate and statusline: built in for Claude Code, Codex, and Gemini. Cursor / Windsurf / Cline / Copilot and 27 others get a per-agent native install. The Claude Code statusline shows [CAVEMAN] ⛏ 12.4k lifetime savings.
Language preservation: “Speak your tongue. Caveman keep your language.” Code, error strings, URLs, paths stay exact; v1.7.0 added Typst + LaTeX to the protected list.

The benchmark

The README ships a project-published benchmark of 10 real Claude API prompts with raw counts, raw data, and a reproduction script in benchmarks/. Average 1,214 → 294 = 65% reduction, range 22–87% (selected rows: React re-render bug 1,180 → 159 = 87%; auth middleware 704 → 121 = 83%; async/await refactor 387 → 301 = 22%; React error boundary 3,454 → 456 = 87%; PostgreSQL race 1,200 → 232 = 81%).

A caveman-compress sub-skill cuts 46% of memory-file tokens on 5 real receipts (e.g. claude-md-preferences.md 706 → 285 = 59.6%; project-notes.md 1,145 → 535 = 53.3%). A three-arm eval harness in evals/ compares baseline (verbose) / terse (“Answer concisely”) / skill (caveman) so the delta is reported against a strong generic terse baseline, not against the verbose default.

The README also references a March 2026 paper, “Brevity Constraints Reverse Performance Hierarchies in Language Models”, which found a 26-point accuracy lift on certain benchmarks under brief-response constraints.

The ecosystem

The under-reported story is not the skill, it is the five-repo family:

Repo	What it does	Stars / Forks / Issues (2026-06-20)
caveman	Output compression skill	74,940 / 4,230 / 299
caveman-code	Terminal coding agent; 25-task MicroBench 1.93× vs Codex CLI	548 / 60 / 26
cavemem	Cross-agent persistent memory	551 / 48 / 37
cavekit	Spec-driven build loop	1,044 / 75 / 14
cavegemma	Gemma 4 31B LoRA fine-tune; 193-pair holdout, 81.5% accuracy	56 / 10 / 0

What to watch

Independent third-party reproduction of the 10-prompt benchmark. No independent run as of 2026-06-20.
Thinking-token cost on reasoning models. The 65% output reduction may be partially offset by unchanged thinking-token costs.
caveman-code adoption. Published 25-task MicroBench (524k vs 1,010k tokens, 14/25 vs 15/25 pass rate) is still project self-report.
cavegemma accuracy on non-coding tasks. 193-pair holdout reports 81.5% accuracy; the maintainer is direct that compression is weaker than gold pairs.

Risks and caveats

Thinking tokens are not reduced. Billable tokens on reasoning models = output + thinking. The 65% / 22–87% numbers are output-only.
The 65% average hides a 22% floor. Refactor callback to async/await is 387 → 301 (22%).
The benchmark is project self-report. No independent third-party reproduction as of 2026-06-20.
v1.9.0 is “Rock pinned” not “Rock perfect.” No long-term compatibility contract — pin to v1.9.0 for production.
299 open issues / 201 commits is a velocity signal, not a stability signal.

Practical advice

For Claude Code / Codex / Gemini / Cursor users. Install the one-liner, trigger with /caveman [lite|full|ultra|wenyan], then run /caveman-stats. Pin to v1.9.0 for team rollout.
For multi-agent or OpenClaw users. Use --with-init for Cursor / Windsurf / Cline / Copilot. For OpenClaw, the installer drops a marker-fenced block into ~/.openclaw/workspace/SOUL.md so the brevity auto-injects every turn.
For operators / IT admins. Verify the install in a clean sandbox — the script touches ~/.claude/, ~/.codex/, ~/.gemini/, ~/.cursor/, and the OpenClaw workspace path. For reasoning models, set the billable-token expectation to “output tokens down 65% on average (22–87% range); thinking tokens unchanged” — not “65% off the bill.”

Verdict

JuliusBrussee/caveman — 74,940 stars, 4,230 forks, 299 open issues, MIT, 11 weeks old — with a project-published benchmark showing 65% average output-token reduction (range 22–87%) and caveman-compress cutting 46% of memory-file tokens. The under-reported story is the five-repo ecosystem: caveman-code (1.93× fewer tokens than Codex CLI), cavemem, cavekit, cavegemma (193-pair holdout, 81.5% accuracy, 96–100% code-fence exactness). The load-bearing caveat is in the README’s own Important box: “Caveman only affects output tokens — thinking/reasoning tokens untouched.” The takeaway is the writing-style layer and the ecosystem framing, not the headline number.