---
title: "Skill-Caveman-Concise replies save tokens"
description: "Caveman is an open-source AI Agent Skill whose motto is \"why use many token when few token do trick\". It makes AI assistants reply in ultra-short caveman-style sentences, compressing verbose output into short phrases — reportedly saving around 65% of token usage."
url: "https://aaronchan.com/kb/skill-caveman/"
updated: "2026-10-07"
category: "AI-Skill (en)"
---

# Skill-Caveman-Concise replies save tokens

> 📅 Last updated: 2026-10-02
> ⚠️ Star counts and version info may change over time; refer to the official GitHub repo

---

## 1. What is it?

Caveman is an open-source AI Agent Skill whose motto is "**why use many token when few token do trick**". It makes AI assistants reply in ultra-short "caveman-style" sentences, compressing verbose output into short phrases — reportedly saving **around 65% of token usage**.

| Item | Detail |
| --- | --- |
| **Author** | JuliusBrussee |
| **Repo** | [https://github.com/JuliusBrussee/caveman](https://github.com/JuliusBrussee/caveman) |
| **Stars** | ~109k (at time of writing) |
| **License** | Apache-2.0 (fully re-licensed from v3.0.0) |
| **Nature** | Skill + Proxy + Middleware trio |
| **Official site** | [https://caveman.so](https://caveman.so) |

---

## 2. Core principle

Caveman isn't just a prompt — it is split into three layers:

| Component | Role |
| --- | --- |
| **Skill (what it says)** | instructs the model to reply in ultra-short sentences |
| **Proxy (what it reads)** | a middle-layer proxy that compresses input sent to the model |
| **Middleware (what your app sends)** | lets your own application apply compression too |

> In other words, it compresses both "what the model says" and "what the model reads", saving tokens on both ends.

---

## 3. Usage

The `/caveman` command offers several intensity levels:

| Command | Effect |
| --- | --- |
| `/caveman lite` | light compression |
| `/caveman full` | standard compression |
| `/caveman ultra` | extreme compression |
| `/caveman wenyan-lite / full / ultra` | Classical Chinese (wenyan) version, various levels |
| `/caveman off` | turn off |

---

## 4. Supported platforms

Caveman supports several major coding agents and CLIs:

| Platform | Installation |
| --- | --- |
| **Claude Code** | `claude plugin marketplace add JuliusBrussee/caveman` |
| **Gemini CLI** | `gemini extensions install https://github.com/JuliusBrussee/caveman` |
| **Oh My Pi (OMP)** | `npx -y github:JuliusBrussee/caveman -- --only omp` |
| **Qwen Code** | `caveman qwen` |
| **Codex / Cursor / Windsurf / Cline** | `npx skills add JuliusBrussee/caveman --skill '*'` |

---

## 5. Real-world results

JetBrains actually tested the effect of "talking to AI like a caveman", and the results matched the official claim — **roughly 65% token savings**. This made Caveman go viral in the coding-agent community, becoming one of the most-starred tools of its kind.

---

## ⚠️ Notes

| Note | Description |
| --- | --- |
| Reduced readability | compressed sentences are terser and may affect reading flow |
| Not for every scenario | avoid when formal, detailed output is required |
| Adjustable intensity | pick lite / full / ultra as needed; no need to go extreme |
| Third-party open source | community-developed, not official Anthropic output |

---

## 🔗 References

| Source | Link |
| --- | --- |
| GitHub Repo | [https://github.com/JuliusBrussee/caveman](https://github.com/JuliusBrussee/caveman) |
| Official site | [https://caveman.so](https://caveman.so) |
| JetBrains test article | [https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/](https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/) |

---

*Sources: GitHub and public test articles (October 2026)*