ClaudeShrink

A Claude Code skill that shrinks massive prompts and files using LLMLingua to save tokens.

This skill enables Claude Code to handle massive files (logs, documentation, long traces) by compressing them using LLMLingua. It uses the gpt2 model locally on your CPU to reduce text size while maintaining semantic integrity, significantly saving context window tokens.

Estimated Token Savings

ClaudeShrink uses a dynamic adaptive sizing algorithm to decide how heavily to compress inputs. By default, it targets a ~70% reduction but respects hard floors and ceilings.

Input Type	Text Size	Original Tokens	Compressed Tokens	Reduction	Notes
Small Prompts	< 2,000 chars	~500	~500	0%	Meets 512-token safety floor; stays uncompressed
Large Texts	~15,000 chars	~3,750	~1,125	~70%	Targets 30% of original tokens
Massive Files	> 50,000 chars	> 12,500	4,096	> 70-90%	Hits the 4096-token hard safety cap

Installation

Run this one-liner in your terminal:

curl -fsSL https://raw.githubusercontent.com/g-akshay/ClaudeShrink/main/install.sh | bash

This will:

1. Clone the repo into ~/.claude/skills/ClaudeShrink
2. Create an isolated Python venv (no system pollution)
3. Install llmlingua, torch, transformers, and accelerate

Requirements: Python 3.9+ and git must be on your PATH.

• macOS: install Python via brew install python or python.org
• Linux (Ubuntu/Debian): sudo apt install python3 git — python3-venv is auto-handled by the installer

Note: The gpt2 model (~500 MB) is downloaded on first use, not at install time.

Updating

To update to the latest version, simply run the exact same command again. It is designed to be safe, additive, and will pull the latest code without destroying your environment:

curl -fsSL https://raw.githubusercontent.com/g-akshay/ClaudeShrink/main/install.sh | bash

System & Hardware Requirements

Because ClaudeShrink uses GPT-2 instead of massive local LLMs, it is incredibly lightweight and runs smoothly in the background while you code.

Component	Requirement	Explanation
Memory (RAM)	2GB+ (4GB recommended)	The local GPT-2 model is tiny. Standard developer machines will not notice it running.
Storage (Disk)	~4GB	torch and dependencies take ~3.5GB; the model weights are ~500MB.
Compute	Basic CPU	No GPU required. Runs lightning-fast on Intel, AMD, and Apple Silicon CPUs.
Software	Python 3.9+, Git	Standard environment prerequisites.

How to Trigger

Claude will automatically use this skill when it detects a request to process a large file. You can also explicitly trigger it by asking:

"Use the ClaudeShrink skill to read ./very_large_log.log and summarize the errors."

Configuration

Defaults in scripts/compressor.py:

Setting	Default	Notes
Model	gpt2	Lightweight, CPU-friendly
Device	cpu	Change to cuda if you have a GPU
Target Tokens	Auto (30% of input, 512–4096)	Adaptive — no manual tuning needed

Edit ~/.claude/skills/ClaudeShrink/scripts/compressor.py to override.

---

The "Why": Solving the AI Budget Burn

As Agentic AI and long context windows become the norm, token consumption has skyrocketed. Sending massive trace logs or full code repositories to LLMs on every query often leads to unseen budget explosions.

Recently, companies like Uber have made headlines for burning through their entire annual AI budgets in mere months. Giving engineers tools to push massive payloads to AI is powerful, but doing so without a compression layer is financially dangerous.

ClaudeShrink prevents this by:

1. Stripping redundant or low-information tokens exactly where humans easily miss them.
2. Handling the compression entirely on your local CPU for free before the payload touches the paid cloud API.
3. Giving AI assistants a standardized, enforced cap on how many tokens they consume per request.

Platform Support

Platform	Supported	Notes
macOS	✅	Intel + Apple Silicon
Linux	✅	Ubuntu, Debian (auto-installs python3-venv), Fedora, Arch
Windows	⚠️	Via WSL (Ubuntu recommended) only