Andrej Karpathy’s Method To 10X Your Claude Skills 🧠
You built the Skills. Now here's the step-by-step system that makes them dramatically better while you sleep 🤖
👋 Hey, Linas here! Welcome to another special issue of my daily newsletter. Each day, I focus on 3 stories that are making a difference in the financial technology space. Coupled with things worth watching & the most important money movements, it’s the only newsletter you need for all things when Finance meets Tech. If you’re reading this for the first time, it’s a brilliant opportunity to join a community of 370k+ FinTech & AI leaders:
This is the guide I wish had existed when I finished building the 12 startup Skills from The One-Person Unicorn. If you haven’t read The Ultimate Guide to Claude Skills yet, start there - it covers how to build Skills from scratch.
This guide picks up where that one left off: how to take Skills you’ve already built and make them reliably excellent.
In February, I shared a system of 12 interconnected Claude Skills that covers the entire founder journey - from idea validation to board management. Hundreds of thousands of views, tens of thousands of downloads, and lots of positive feedback.
But I had a problem I didn’t write about.
My fundraising Skill would produce a pitch deck structure that followed the Sequoia/YC hybrid format about 70% of the time. The other 30%, it would drift: missing the traction slide, burying the ask on slide 9, or opening with a market size number instead of the problem. My sales Skill nailed MEDDIC qualification about two-thirds of the time and produced something vague and generic the rest. The idea validation Skill sometimes ran the Mom Test framework perfectly and sometimes gave the kind of advice you’d get from asking ChatGPT “how to validate a startup idea.”
I was manually reviewing and fixing every output. That’s not an AI operating system. That’s a fancy autocomplete with extra steps.
Then Andrej Karpathy released something that solved the problem entirely.
Not a new model. Not a prompt trick. A method.
He called it autoresearch, and 42,000 people starred it on GitHub in the first week alone. Karpathy built it for optimizing ML training code, but the pattern underneath is universal. It works on any file where you can define what “good” looks like and measure it.
I pointed it at my startup Skills. The fundraising Skill went from producing correct pitch deck structures 70% of the time to 94%. The sales Skill’s MEDDIC qualification accuracy jumped from 65% to 91%. Each optimization run took a few hours, mostly unattended, and cost less than a coffee.
This guide shows you exactly how to do the same thing. You’ll get:
The method explained in practical terms (no ML knowledge needed)
A step-by-step walkthrough with a real worked example
Eval templates for each of the 12 startup Skill categories
The two downloadable files that make the whole thing work seamlessly, and
The mistakes that waste entire runs, so you can avoid them.
If you’re building with Claude Skills, this is the difference between Skills that sort of work and Skills that work 99% every time.
To make Claude even more powerful, I’m also sharing an end-to-end guide on how to Turn Claude in Excel Into Your Senior Financial Analyst, and how I turned Claude Cowork Into My Personal COO that does real work while I sleep. It has everything you need to make AI your most powerful resource & leverage, and automate your life 🤖