ShresthSamyak

LLM Diet — AI skill for Claude Code

AI community

Cut AI coding tokens by 99%.

How to install LLM Diet

This entry records only its repository, not the path inside it, so there is no exact command to give. Open ShresthSamyak/LLM_DIET and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What LLM Diet does

Cut AI coding tokens by 99%. Deterministic context injection for Claude Code, Cursor, Windsurf.

Alternatives in AI

Jcodemunch MCP — Cut AI token costs 95%+ on code exploration 2.6k ★
Codesight — Universal AI context generator 979 ★
Skilldock — SkillDock is an AI skill manager and skill management desktop app for Claude Code, Cursor, Codex, Windsurf, Ge 490 ★

README

llm-diet

**Give Claude the right context upfront. Fewer turns, faster answers, lower cost.**

[![PyPI](https://img.shields.io/pypi/v/llm-diet)](https://pypi.org/project/llm-diet/) [![License: MIT](https://img.shields.io/github/license/ShresthSamyak/LLM_DIET)](LICENSE) [![Downloads](https://img.shields.io/badge/downloads-PyPI-brightgreen)](https://pypi.org/project/llm-diet/)

Deterministic context retrieval for AI coding tools. Parses your repo into a call graph, intercepts every file read Claude makes, and returns compressed versions — so Claude explores freely but cheaply.

The Problem

Every Claude Code session starts blind. Claude explores your entire codebase before answering — reading files, listing directories, running commands. That exploration costs tokens and time.

Without llm-diet:
  Claude reads 10 files × 8,000 tokens = 80,000 tokens consumed
  Cost: $0.19 for a simple bug fix session

With llm-diet:
  Claude reads 10 files × 300 tokens  = 3,000 tokens consumed
  Cost: $0.025 for the same session

How It Works

User prompt
│
▼
context-engine (call graph)
│  scores every function against your query
▼
Claude Code session opens
│
▼
Claude calls read_file("validators/amazon.py")
│
▼
llm-diet-shadow MCP server intercepts
│  returns compressed 872-token version
│  instead of raw 6,590-token file
▼
Claude answers — correctly — using compressed context

Claude thinks it explored. It did — but every read returned our compressed version, not the raw file.

Benchmark

**Tested on coupon-hunter-poc (40-node Python project)**

File	Original	Compressed	Reduction
validators/playwright_amazon.py	6,590 chars	872 chars	86%
orchestrator.py	10,492 chars	2,169 chars	79%
connectors/playwright_amazon.py	3,067 chars	631 chars	79%
openrouter_agent.py	2,860 chars	966 chars	66%
retailmenot_scraper.py	2,705 chars	960 chars