token-squasher

Compress verbose system prompts into dense, token-efficient instructions without losing logic.

Features

• Token-efficient compression - Reduces prompt length by removing filler while preserving all logic
• Interactive REPL - Multi-line prompt input with command support
• Real-time statistics - Token count reduction and compression percentage
• Colored output - Syntax highlighting for better readability
• Claude Sonnet integration - Uses Anthropic's latest model for intelligent compression
• Reasoning display - Shows what was compressed and how (optional)
• Session management - Track multiple compression attempts

Prerequisites

• Node.js >= 18.0.0
• Anthropic API key

Installation

npm install -g @sidhxntt/token-squasher

Or run directly with npx:

npx @sidhxntt/token-squasher

Configuration

Set your Anthropic API key as an environment variable:

export ANTHROPIC_API_KEY=sk-ant-your-api-key-here

Usage

Interactive Mode

Start the REPL:

token-squasher

Commands

• :stats - Toggle token statistics display
• :verbose - Toggle compression reasoning display
• :clear - Clear screen and reset
• :quit or :q - Exit the application

Example Session

token-squasher v2.0.0
Compress verbose system prompts. Keep logic. Crush tokens.

✔ ANTHROPIC_API_KEY detected

Paste your system prompt and press Enter twice to squash.
Commands: :stats toggle stats  :verbose toggle reasoning  :quit exit
──────────────────────────────────────────────────────────

Prompt #1 — paste prompt then Enter twice to squash
› You are a helpful AI assistant. Please analyze the user's input carefully 
› and provide a comprehensive response that addresses all aspects of their 
› question. Make sure to be thorough and considerate in your analysis.
› 

  ⠋ Squashing tokens...

  ✔ Compressed successfully

Analyze user input comprehensively. Address all question aspects thoroughly.

───── TOKEN STATS ─────
Before   47 tokens  (187 chars)
After    12 tokens  (74 chars)
Crushed  74.5% ████████████████████░░░░
───────────────────────

Input Methods

1. Multi-line paste - Paste your prompt and press Enter twice
2. Line-by-line - Type each line, pressing Enter between lines, then Enter twice to finish

API Reference

The core compression function:

async function squashPrompt(promptText)

Parameters:

• promptText (string) - The verbose system prompt to compress

Returns:

• Object with squashed (string) and reasoning (string) properties

Model: Uses claude-sonnet-4-20250514 with max 4096 tokens

Project Structure

token_quasher/
├── index.js          # Main CLI application with REPL
├── package.json      # NPM package configuration
└── package-lock.json # Dependency lock file

Key Files

• index.js - Complete CLI implementation with Anthropic integration, REPL interface, and token estimation
• package.json - Defines the package as a global CLI tool with Node.js 18+ requirement

License

MIT