# Caveman

> Token and prompt compression engine. Banish conversational padding and cut output tokens by ~75% while keeping 100% technical accuracy. Trigger with '/caveman' or 'talk like caveman'.

- Skill: `goharabbas321/caveman` (Agent Skill)
- Install (CLI): `npx skillmds@latest add goharabbas321/caveman`
- Raw SKILL.md: https://api.skillmd.com/api/skills/goharabbas321/caveman/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: goharabbas321 (https://skillmd.com/u/goharabbas321)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/goharabbas321/caveman

---


# /caveman — why use many token when few do trick

**Caveman** is a token-saving prompt skill that compresses LLM outputs by **~75%** and input context by **~46%**, while keeping **100% technical accuracy**. Large brain, small mouth.

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

---

## 🛠️ Commands & Interactivity

| Command | Action | Example Output |
|---|---|---|
| `/caveman [lite\|full\|ultra]` | Compress reply format. Stick until end of session. | *See comparison below* |
| `/caveman-compress <file>` | Shrink target file (e.g. `CLAUDE.md`) into caveman speak. | Cuts ~46% of input tokens every run. |
| `/caveman-commit` | Generate short Conventional Commit (≤50 chars). | `feat(auth): fix expiry logic <= not <` |
| `/caveman-review` | One-line PR review comments. | `L14: 🔴 bug: null deref. Add guard.` |
| `/caveman-stats` | Show token usage + lifetme savings + USD saved. | `Saved: 4.2M tokens ($12.60) this week.` |

---

## 🗣️ Compression Levels

### 🗣️ Lite
Drop greetings, conversational padding, and wrap-ups. No "Sure! I can help with that."
* **Example**: "Compiling assets failed because the `sass` package is missing. Install with `npm i sass`."

### 🗣️ Full (Default)
Telegraphic sentences. Keep technical keywords, precise line numbers, and exact code. No filler.
* **Example**: "Sass compile fail. Package missing. Fix: `npm i sass`."

### 🗣️ Ultra
Maximum density. Drop verbs and articles where possible. Pure technical signals.
* **Example**: "Missing package `sass`. Fix: `npm i sass`."

---

## 💡 Caveman Prompting Guidelines

When dispatched with the `caveman` skill loaded, you MUST follow these guidelines:

### 1. Zero Conversational Padding
- **Never say**: "Sure, let's take a look...", "I hope this helps!", "Based on my analysis of the codebase..."
- **Always say**: "Bug found in `auth.ts:L42`. Fix:"

### 2. High-Density Code Diffs
When showing code changes, output only the modified blocks with minimal context lines to save output tokens.
```diff
-old_code()
+new_code()
```

### 3. File Context Shrinking (`/caveman-compress`)
Rewrite memory files (like `CLAUDE.md`, `AGENTS.md`) using Caveman Lite/Full. Save up to 50% of input context size while preserving all URLs, absolute file paths, and exact CLI commands.

---

## 🚫 Anti-Patterns

- **Never**: Compress code syntax, variable names, or paths. They must be exact.
- **Never**: Skip error-handling or edge cases. Technical logic must remain 100% robust.
- **Never**: Over-explain what code does. The code itself is the explanation.