caveman

why use many token when few do trick

Before/After • Install • Levels • Skills • Benchmarks • Evals

A Claude Code skill/plugin and Codex plugin that makes agent talk like caveman — cutting ~75% of output tokens while keeping full technical accuracy. Now with 文言文 mode, terse commits, one-line code reviews, and a compression tool that cuts ~45% of input tokens every session.

Based on the viral observation that caveman-speak dramatically reduces LLM token usage without losing technical substance. So we made it a one-line install.

Same fix. 75% less word. Brain still big.

Pick your level of grunt:

Same answer. You pick how many word.

┌─────────────────────────────────────┐
│  TOKENS SAVED          ████████ 75% │
│  TECHNICAL ACCURACY    ████████ 100%│
│  SPEED INCREASE        ████████ ~3x │
│  VIBES                 ████████ OOG │
└─────────────────────────────────────┘

• Faster response — less token to generate = speed go brrr
• Easier to read — no wall of text, just the answer
• Same accuracy — all technical info kept, only fluff removed (science say so)
• Save money — ~71% less output token = less cost
• Fun — every code review become comedy

npx skills add JuliusBrussee/caveman

npx skills supports 40+ agents — Claude Code, GitHub Copilot, Cursor, Windsurf, Cline, and more. To install for a specific agent:

npx skills add JuliusBrussee/caveman -a cursor
npx skills add JuliusBrussee/caveman -a github-copilot
npx skills add JuliusBrussee/caveman -a cline
npx skills add JuliusBrussee/caveman -a windsurf
npx skills add JuliusBrussee/caveman -a codex

Or with Claude Code plugin system:

claude plugin marketplace add JuliusBrussee/caveman
claude plugin install caveman@caveman

Codex:

1. Clone repo
2. Open Codex in repo
3. Run /plugins
4. Search Caveman
5. Install plugin

Install once. Use in all sessions after that. One rock. That it.

Trigger with:

• /caveman or Codex $caveman
• "talk like caveman"
• "caveman mode"
• "less tokens please"

Stop with: "stop caveman" or "normal mode"

Classical Chinese literary compression — same technical accuracy, but in the most token-efficient written language humans ever invented.

Level stick until you change it or session end.

Caveman make Claude speak with fewer tokens. Compress make Claude read fewer tokens.

Your CLAUDE.md loads on every session start. Caveman Compress rewrites memory files into caveman-speak so Claude reads less — without you losing the human-readable original.

/caveman:compress CLAUDE.md

CLAUDE.md          ← compressed (Claude reads this every session — fewer tokens)
CLAUDE.original.md ← human-readable backup (you read and edit this)

Code blocks, URLs, file paths, commands, headings, dates, version numbers — anything technical passes through untouched. Only prose gets compressed. See the full caveman-compress README for details. Security note: Snyk flags this as High Risk due to subprocess/file patterns — it's a false positive.

Real token counts from the Claude API (reproduce it yourself):

Range: 22%–87% savings across prompts.

A March 2026 paper "Brevity Constraints Reverse Performance Hierarchies in Language Models" found that constraining large models to brief responses improved accuracy by 26 percentage points on certain benchmarks and completely reversed performance hierarchies. Verbose not always better. Sometimes less word = more correct.

Caveman not just claim 75%. Caveman prove it.

The evals/ directory has a three-arm eval harness that measures real token compression against a proper control — not just "verbose vs skill" but "terse vs skill". Because comparing caveman to verbose Claude conflate the skill with generic terseness. That cheating. Caveman not cheat.

# Run the eval (needs claude CLI)
uv run python evals/llm_run.py

# Read results (no API key, runs offline)
uv run --with tiktoken python evals/measure.py

Snapshots committed to git. CI runs free. Every number change reviewable as diff. Add a skill, add a prompt — harness pick it up automatically.

If caveman save you mass token, mass money — leave mass star. ⭐

• Blueprint — specification-driven development for Claude Code. Natural language → blueprints → parallel builds → working software.
• Revu — local-first macOS study app with FSRS spaced repetition, decks, exams, and study guides. revu.cards

MIT — free like mass mammoth on open plain.