caveman-token-optimizer（穴居人令牌优化器）

Caveman Token Optimizer

Skill by ara.so — Daily 2026 Skills collection.

A Claude Code skill and Codex plugin that makes AI agents respond in compressed caveman-speak — cutting ~65% of output tokens on average (up to 87%) while keeping full technical accuracy. No pleasantries. No filler. Just answer.

What It Does

Caveman mode strips:

• Pleasantries: "Sure, I'd be happy to help!" → gone
• Hedging: "It might be worth considering" → gone
• Articles (a, an, the) → gone
• Verbose transitions → gone

Caveman keeps:

• All code blocks (written normally)
• Technical terms (exact: useMemo, polymorphism, middleware)
• Error messages (quoted exactly)
• Git commits and PR descriptions (normal)

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

Install

Claude Code (npx)

npx skills add JuliusBrussee/caveman

Claude Code (plugin system)

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

Codex

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

Install once. Works in all sessions after that.

Manual / Local

git clone https://github.com/JuliusBrussee/caveman.git
cd caveman
pip install -e .

Usage — Trigger Commands

Claude Code

/caveman          # enable default (full) caveman mode
/caveman lite     # professional brevity, grammar intact
/caveman full     # default — drop articles, use fragments
/caveman ultra    # maximum compression, telegraphic

Codex

$caveman
$caveman lite
$caveman full
$caveman ultra

Natural language triggers

Any of these phrases activate caveman mode:

• "talk like caveman"
• "caveman mode"
• "less tokens please"
• "be concise"

Disable

/caveman off
# or say: "stop caveman" / "normal mode"

Level sticks until changed or session ends.

Intensity Levels

Benchmark Results

Real token counts from Claude API (reproducible via benchmarks/ directory):

Important: Caveman only affects output tokens. Thinking/reasoning tokens are untouched. Caveman make *mouth* smaller, not brain.

Reproducing Benchmarks

git clone https://github.com/JuliusBrussee/caveman.git
cd caveman/benchmarks

# Set your Anthropic API key
export ANTHROPIC_API_KEY=your_key_here

# Run benchmark suite
python run_benchmarks.py

# Compare normal vs caveman responses
python compare.py --task react-rerender
python compare.py --task auth-middleware
python compare.py --all

Code Examples — What Caveman Mode Changes

Before (normal, 69 tokens)

The reason your React component is re-rendering is likely because
you're creating a new object reference on each render cycle. When
you pass an inline object as a prop, React's shallow comparison
sees it as a different object every time, which triggers a
re-render. I'd recommend using useMemo to memoize the object.

After (caveman full, 19 tokens)

New object ref each render. Inline object prop = new ref = re-render.
Wrap in `useMemo`.

Code blocks stay normal — caveman not stupid

# Caveman explains in grunt, but code stays clean:
# "Token expiry check broken. Fix:"

def verify_token(token: str) -> bool:
    payload = jwt.decode(token, SECRET_KEY, algorithms=["HS256"])
    # Was: payload["exp"] < time.time()
    # Fix:
    return payload["exp"] >= time.time()

What Caveman Preserves vs. Removes

# Tokens caveman REMOVES (waste):
filler_phrases = [
    "I'd be happy to help you with that",   # 8 tokens gone
    "The reason this is happening is because", # 7 tokens gone
    "I would recommend that you consider",   # 7 tokens gone
    "Sure, let me take a look at that",      # 8 tokens gone
    "Great question!",                        # 2 tokens gone
    "Certainly!",                             # 1 token gone
]

# Things caveman KEEPS (substance):
preserved = [
    "code blocks",          # always normal
    "technical_terms",      # exact spelling preserved
    "error_messages",       # quoted verbatim
    "variable_names",       # exact
    "git_commits",          # normal prose
    "pr_descriptions",      # normal prose
]

Integration Pattern — Using in a Project

If you want caveman-style compression in your own Claude API calls:

import anthropic

client = anthropic.Anthropic()  # uses ANTHROPIC_API_KEY env var

# Load the caveman SKILL.md as a system prompt addition
with open("path/to/caveman/SKILL.md", "r") as f:
    caveman_skill = f.read()

response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=1024,
    system=f"{caveman_skill}\n\nRespond in caveman mode: full intensity.",
    messages=[
        {"role": "user", "content": "Why is my React component re-rendering?"}
    ]
)

print(response.content[0].text)
# → "New object ref each render. Inline prop = new ref = re-render. useMemo fix."
print(f"Tokens used: {response.usage.output_tokens}")  # ~19 vs ~69

Session Workflow

# Start session with caveman
/caveman full

# Ask technical questions normally — agent responds in caveman
> Why does my Docker build take so long?
→ "Layer cache miss. COPY before RUN npm install. Fix order:"
[code block shown normally]

# Switch intensity mid-session
/caveman lite

# Turn off for PR description writing
/caveman off
> Write a PR description for this auth fix
→ [normal, professional prose]

# Back to caveman
/caveman

Troubleshooting

Caveman mode not activating:

# Verify plugin installed
claude plugin list | grep caveman

# Reinstall
claude plugin remove caveman
claude plugin install caveman@caveman

Savings lower than expected:

• Caveman only compresses *output* tokens — input tokens unchanged
• Tasks with heavy code output (like Docker setup) see less savings since code is preserved verbatim
• Reasoning/thinking tokens not affected — savings show in visible response only
• Ultra mode gets maximum compression; switch if full mode feels verbose

Need normal mode for specific output:

/caveman off   # for PR descriptions, user-facing docs, formal reports
/caveman       # re-enable after

Benchmarking your own tasks:

cd benchmarks/
export ANTHROPIC_API_KEY=your_key_here
python run_benchmarks.py --task "your custom task description"

Why It Works

Backed by a March 2026 paper "Brevity Constraints Reverse Performance Hierarchies in Language Models": constraining large models to brief responses improved accuracy by 26 percentage points on certain benchmarks. Verbose not always better.

TOKENS SAVED          ████████ 65% avg (up to 87%)
TECHNICAL ACCURACY    ████████ 100%
RESPONSE SPEED        ████████ faster (less to generate)
READABILITY           ████████ better (no wall of text)

Key Files

caveman/
├── SKILL.md          # the skill definition loaded by Claude Code
├── benchmarks/
│   ├── run_benchmarks.py   # reproduce token count results
│   └── compare.py          # side-by-side comparison tool
├── plugin.json             # Codex plugin manifest
└── README.md

Links

• Repo: https://github.com/JuliusBrussee/caveman
• Homepage: https://juliusbrussee.github.io/caveman/
• Also by Julius: Blueprint — spec-driven dev for Claude Code

*One rock. That it.* 🪨