caveman-token-optimizer

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Caveman Token Optimizer

穴居人Token优化器

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.

技能作者：ara.so — 2026年度每日技能系列作品

一款Claude Code技能和Codex插件，可让AI Agent以压缩的穴居人风格话术回复，平均减少约65%的输出token（最高可达87%），同时完全保留技术准确性。没有客套话，没有冗余内容，只给答案。

What It Does

功能说明

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.

穴居人模式会删除以下内容：

客套话术："Sure, I'd be happy to help!" → 全部移除
模糊表述："It might be worth considering" → 全部移除
冠词（a、an、the） → 全部移除
冗长的过渡语 → 全部移除

穴居人模式会保留以下内容：

所有代码块（保持正常书写格式）
技术术语（完全保留：
```
useMemo
```
、
```
polymorphism
```
、
```
middleware
```
等）
错误信息（完全原文引用）
Git提交和PR描述（保持正常格式）

同样的解决方案，字数少75%，智能程度不打折。

Install

安装方法

Claude Code (npx)

Claude Code（npx方式）

bash

npx skills add JuliusBrussee/caveman

bash

npx skills add JuliusBrussee/caveman

Claude Code (plugin system)

Claude Code（插件系统方式）

bash

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

bash

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

Codex

Clone the repo
Open Codex inside the repo
Run
```
/plugins
```
Search
```
Caveman
```
Install plugin

Install once. Works in all sessions after that.

克隆仓库
在仓库目录中打开Codex
执行
```
/plugins
```
命令
搜索
```
Caveman
```
安装插件

安装一次即可，后续所有会话都生效。

Manual / Local

手动/本地安装

bash

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

bash

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

Usage — Trigger Commands

使用方式 — 触发命令

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

/caveman          # 开启默认（完整）穴居人模式
/caveman lite     # 专业简洁风格，保留语法
/caveman full     # 默认模式 — 移除冠词，使用短句片段
/caveman ultra    # 极致压缩，电报式表述

Codex

$caveman
$caveman lite
$caveman full
$caveman ultra

$caveman
$caveman lite
$caveman full
$caveman ultra

Natural language triggers

自然语言触发

Any of these phrases activate caveman mode:

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

以下任意短语都可以激活穴居人模式：

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

Disable

关闭方式

/caveman off

/caveman off

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

或者说："stop caveman" / "normal mode"


Level sticks until changed or session ends.


模式等级会保持生效，直到修改或者会话结束。

Intensity Levels

强度等级

Level	Trigger	Style	Example
Lite	`/caveman lite`	Drop filler, keep grammar	"Component re-renders because inline object prop creates new reference each cycle. Wrap in `useMemo` ."
Full	`/caveman full`	Drop articles, use fragments	"New object ref each render. Inline prop = new ref = re-render. Wrap in `useMemo` ."
Ultra	`/caveman ultra`	Telegraphic, abbreviate everything	"Inline obj prop → new ref → re-render. `useMemo` ."

等级	触发命令	风格	示例
Lite（轻度）	`/caveman lite`	移除冗余内容，保留语法	"Component re-renders because inline object prop creates new reference each cycle. Wrap in `useMemo` ."
Full（完整）	`/caveman full`	移除冠词，使用短句片段	"New object ref each render. Inline prop = new ref = re-render. Wrap in `useMemo` ."
Ultra（极致）	`/caveman ultra`	电报式表述，所有内容尽可能缩写	"Inline obj prop → new ref → re-render. `useMemo` ."

Benchmark Results

基准测试结果

Real token counts from Claude API (reproducible via

benchmarks/

directory):

Task	Normal	Caveman	Saved
React re-render bug	1180	159	87%
Auth middleware fix	704	121	83%
PostgreSQL pool setup	2347	380	84%
Git rebase vs merge	702	292	58%
Async/await refactor	387	301	22%
Docker multi-stage build	1042	290	72%
Average	1214	294	65%

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

来自Claude API的真实token计数（可通过

benchmarks/

目录复现）：

任务	正常模式	穴居人模式	节省比例
React重渲染bug修复	1180	159	87%
Auth中间件修复	704	121	83%
PostgreSQL连接池设置	2347	380	84%
Git rebase vs merge对比	702	292	58%
Async/await重构	387	301	22%
Docker多阶段构建	1042	290	72%
平均	1214	294	65%

重要提示： 穴居人模式仅影响输出token，思考/推理token完全不受影响。只是让AI"说话"更简洁，不是降低思考能力。

Reproducing Benchmarks

复现基准测试

bash

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

bash

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

Set your Anthropic API key

设置你的Anthropic API密钥

export ANTHROPIC_API_KEY=your_key_here

Run benchmark suite

运行基准测试套件

python run_benchmarks.py

Compare normal vs caveman responses

对比正常模式和穴居人模式的回复

python compare.py --task react-rerender python compare.py --task auth-middleware python compare.py --all

undefined

python compare.py --task react-rerender python compare.py --task auth-middleware python compare.py --all

undefined

Code Examples — What Caveman Mode Changes

代码示例 — 穴居人模式的变化

Before (normal, 69 tokens)

之前（正常模式，69个token）

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.

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)

之后（完整穴居人模式，19个token）

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

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

Code blocks stay normal — caveman not stupid

代码块保持正常 — 穴居人模式不会改动代码

python

undefined

python

undefined

Caveman explains in grunt, but code stays clean:

穴居人用简洁话术解释，但代码保持整洁：

"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()

undefined

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()

undefined

What Caveman Preserves vs. Removes

穴居人模式保留与移除内容说明

python

undefined

python

undefined

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 ]

undefined

undefined

Integration Pattern — Using in a Project

集成方案 — 在项目中使用

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

python

import anthropic

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

如果你想在自己的Claude API调用中使用穴居人风格压缩：

python

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)

with open("path/to/caveman/SKILL.md", "r") as f: caveman_skill = f.read()

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

undefined

print(f"Tokens used: {response.usage.output_tokens}") # ~19 vs ~69

undefined

Session Workflow

会话使用流程

undefined

undefined

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]

/caveman off

Write a PR description for this auth fix → [normal, professional prose]

Back to caveman

/caveman

undefined

/caveman

undefined

Troubleshooting

故障排查

Caveman mode not activating:

bash

undefined

穴居人模式未激活：

bash

undefined

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

claude plugin remove caveman claude plugin install caveman@caveman


**节省比例低于预期：**
- 穴居人模式仅压缩*输出*token — 输入token无变化
- 代码输出占比高的任务（比如Docker配置）节省比例较低，因为代码会完全保留
- 推理/思考token不受影响 — 仅可见回复部分有节省
- 极致模式可获得最大压缩效果，如果完整模式感觉还是冗长可以切换

**特定输出需要正常模式时：**

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


**为自己的任务做基准测试：**
```bash
cd benchmarks/
export ANTHROPIC_API_KEY=your_key_here
python run_benchmarks.py --task "your custom task description"

Why It Works

原理说明

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)

基于2026年3月的论文《Brevity Constraints Reverse Performance Hierarchies in Language Models》：限制大模型输出简短回复，在部分基准测试中准确率提升了26个百分点。冗长不一定更好。

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

Key Files

核心文件

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

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

链接

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. 🪨

代码仓库： https://github.com/JuliusBrussee/caveman
主页： https://juliusbrussee.github.io/caveman/
作者Julius的其他作品： Blueprint — spec-driven dev for Claude Code

一块石头，搞定。 🪨