🪨 Caveman：59.5k Stars 的 Claude Code 技能，一句话省 65% Token
「为什么用很多 token，当很少 token 就能 trick」——这句魔性口号来自 JuliusBrussee/caveman，一个在 2026 年 4 月 4 日发布、两周冲到 59.5k Stars 的开源 Claude Code 技能/插件。它不做别的：让 AI Agent 说人话——不是"人话"那种客套，是真的像原始人一样精简。
几个数字
| 场景 | 正常 Claude | Caveman | 节省 |
|------|:-----------:|:-------:|:----:|
| 解释 React 重渲染 bug | 1180 tokens | 159 tokens | 87% |
| 修复 auth 中间件 token 过期 | 704 | 121 | 83% |
| 实现 React Error Boundary | 3454 | 456 | 87% |
| 全部 10 个 benchmark 平均 | 1214 | 294 | 65% |
数据不掺水。基准测的是 Answer concisely. 的简洁模式 vs Caveman，不是 vs 默认啰嗦模式。
怎么用
安装一句话：
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash
装好后，对 Agent 说 talk like caveman 或输入 /caveman 激活。想要更狠的可以指定等级：
/caveman lite — 去掉废话/caveman full — 默认原始人模式/caveman ultra — 电报体/caveman wenyan — 文言文（比 ultra 还短）恢复原状：说 normal mode。
不只有输出压缩
Caveman 生态有三个兄弟：
| 工具 | 作用 |
|------|------|
| caveman (this) | 输出压缩 — 减少 Agent 说的话 |
| cavemem | 跨 Agent 记忆 — 减少 Agent 忘的事 |
| cavekit | Spec-Driven 开发循环 — 减少 Agent 猜的错 |
还有个 caveman-shrink MCP 中间件，包装任意 MCP 服务器自动压缩工具描述。以及 cavecrew-* 子 Agent（investigator/builder/reviewer），比 vanilla Agent 省约 60% token，主 Agent 的上下文窗口能撑更久。
原理不复杂
安装脚本在 Agent 配置目录丢一个 skill 文件，通过 system prompt 告诉 Agent：去掉 filler，保留 substance，用碎片句。它不改 thinking/reasoning token——脑子还是那个脑子，只是嘴变小了。
有个有趣的佐证：2026 年 3 月的论文《Brevity Constraints Reverse Performance Hierarchies in Language Models》发现，约束大模型用简短回答反而在某些 benchmark 上提升了 26 个百分点的准确率。话多不一定对。
兼容哪些 Agent
Claude Code、Codex、Gemini（内置自动激活），Cursor / Windsurf / Cline / Copilot / OpenClaw 等 30+ 个。安装脚本自动检测已安装的 Agent 并逐个装好。
一句话总结：花大价钱租了最聪明的脑子，就别再为废话付钱。
🪨 Caveman: 59.5k Stars Claude Code Skill That Saves 65% Tokens by Talking Like a Caveman
"Why use many token when few token do trick" — that's the tagline of JuliusBrussee/caveman, a Claude Code skill that dropped on April 4, 2026 and hit 59.5k stars in six weeks. It does exactly one thing: makes AI agents shut up and get to the point.
The Numbers
| Task | Normal Claude | Caveman | Saved |
|------|:-------------:|:-------:|:-----:|
| Explain React re-render bug | 1180 tokens | 159 tokens | 87% |
| Fix auth middleware token expiry | 704 | 121 | 83% |
| Implement React Error Boundary | 3454 | 456 | 87% |
| 10-benchmark average | 1214 | 294 | 65% |
Honest three-arm eval: baseline vs Answer concisely. vs skill. Even against terse mode, Caveman still cuts 65%.
Install
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash
~30 seconds. Needs Node ≥18. Safe to re-run. The script auto-detects your installed agents (Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, OpenClaw, 30+ more) and installs for each.
Trigger with /caveman or just say "talk like caveman". Stop with "normal mode". Four levels of grunt:
/caveman lite — drop filler words/caveman full — default caveman/caveman ultra — telegraphic/caveman wenyan — classical Chinese (even shorter)Beyond Output Compression
The Caveman ecosystem ships three tools:
| Tool | What |
|------|------|
| caveman | Output compression |
| cavemem | Cross-agent memory |
| cavekit | Spec-driven development loop |
Plus caveman-shrink — an MCP middleware that wraps any MCP server and compresses tool descriptions. And cavecrew-* sub-agents (investigator/builder/reviewer) that use ~60% fewer tokens than vanilla agents.
How It Works
The installer drops a skill file into your agent's config directory. The skill tells the agent: drop filler, keep substance, use fragments. It only affects output tokens — thinking/reasoning tokens are untouched. Brain still big. Mouth small.
Compatibility
Claude Code, Codex, Gemini (auto-activate), Cursor, Windsurf, Cline, Copilot, OpenClaw, and 30+ more. The installer handles detection and per-agent configuration.
Bottom line: You paid for the smartest brain available. Stop paying for fluff.