The Claude Code Skill That Makes Your AI 70x Cheaper (And Smarter)

I just watched Jack Roberts break down the top 5 Claude Code skills on GitHub — tools that have collectively racked up 100,000+ stars — and one in particular stopped me in my tracks.

If you're using Claude Code for any serious development work, there's a very good chance you're burning through tokens in the most inefficient way possible. Jack's breakdown explains why — and what to do about it.

The real problem with Claude Code isn't the AI. It's that every new session starts from scratch. That's the expensive part.

Skill #1: Karpathy Graphify — Your Codebase as a Subway Map

The first skill Jack covers is inspired by Andrej Karpathy's approach to AI knowledge bases. Called Graphify, it turns your codebase into a queryable knowledge graph — and it's reportedly 70x cheaper when asking questions about your project inside Claude Code.

Here's how Jack explains the concept: think of your codebase as a city. Every file is a subway station, every import is a subway line, and every community of related files is a neighborhood. The important nodes — the files everything else connects through — are your "Grand Central Stations."

The problem Claude Code has without this: it walks the streets door by door. With Graphify, it rides the subway lines directly to what it needs.

Every time you start a new Claude Code session, the context resets. You're paying token costs to re-establish what the AI "knows" about your project each time. Graphify shortcuts that massively by giving Claude a map of how everything connects — so it can answer questions in far fewer tokens.

Setup is a single command. It clones from GitHub, builds the graph, and you're querying against your own codebase knowledge base within minutes.

Where it really shines: repos with 500+ files. According to Jack, for anything under 30 files, the overhead can negate the benefit. But for larger projects, the savings compound fast.

In the video, he demonstrates asking about how a specific RAG system works in one of his projects — and Graphify routes the AI directly to the relevant code without the usual token-hungry exploration. He says it's now a permanent part of every project he works on.

Skill #2: Firecrawl — Clean Web Data for AI Agents

The second skill Jack covers is Firecrawl — an AI agent tool designed to solve a specific headache: scraping the web for data that's actually clean and usable.

The problem is familiar to anyone who's built AI pipelines that pull from the web. Websites aren't designed for AI — they're designed for humans. Some pages are JavaScript-heavy, others pure HTML, and the raw output is often messy, low-quality, and requires extensive cleanup. You end up with worse data than you needed, costing time and tokens to fix.

Firecrawl handles the conversion, delivering structured, AI-ready content from messy web pages. For anyone building agents that research, monitor, or pull live data, this is a serious efficiency upgrade.

The Bigger Point

What I love about Jack's approach is the focus on efficiency over raw capability. Both of these skills don't make AI do things it couldn't do before — they make it do the same things at a fraction of the cost.

In any serious AI workflow, tokens are the recurring cost. Knowledge graph routing and clean data pipelines both attack that problem directly. That's the right way to think about AI tooling: not "what can it do?" but "how much does it cost to do it, and how can I drive that cost down?"

The best AI setups aren't the ones with the fanciest models — they're the ones with smarter architectures underneath.

The Bottom Line

Jack Roberts' Top 5 Claude Code Skills video is worth 22 minutes of your time if you're serious about building with Claude Code. Graphify alone changes the math on what it costs to work with larger projects. If you're hitting token limits faster than expected, this is exactly the kind of tool that fixes the underlying problem.