It’s the story of how I spent one weekend letting an autonomous AI system run overnight to improve the very skills I use to build AI tools, and what happened when I checked the results in the morning.
Let me walk you through it…
The pizza analogy that actually makes sense
Before I get into the technical side, I want to tell you about Mickey.
Mickey is a restaurant owner. He makes pizza. The problem is, every time he tries to do everything himself, the result is inconsistent. The dough’s off, the chicken’s wrong, the vegetables aren’t prepped properly. It’s not that Mickey lacks effort. He just can’t be a master of everything at once.
So Mickey brings in three specialists. Tony, who’s spent 20 years grilling chicken and has never once overcooked it. Rosa, who runs the best vegetable stand in town (the farmers call her, not the other way around). And Chef Marco, who cooked in a three Michelin Star kitchen, retired, and still has the taste.
The pizza gets better. Not because Mickey worked harder, but because the right specialists worked in parallel, each doing what they do best, and then the results came together.
That’s the analogy I keep coming back to when I think about agentic AI. You’re building systems where different agents, each with their own skill sets, work simultaneously and then merge their outputs. The result is something none of them could have produced alone.
For expert advice like this straight to your inbox every other Friday, sign up for Pro+ membership.
You’ll also get access to 300+ hours of exclusive video content, a complimentary Summit ticket, and so much more.
So, what are you waiting for?
The problem I was actually trying to solve
In March 2025, Andrej Karpathy, who most people in this space will know, released a GitHub repo called auto-research. The idea behind it is elegant: one agent runs roughly 100 experiments overnight, iterating and improving without any human in the loop. The repo has around 39,000 stars at this point, which tells you something about how many people saw what I saw when I read through it.
I thought, what if I applied this to my own Claude Code skills?
For context, I work in Claude Code and I’ve built a set of skills, essentially structured prompts combined with architectural instructions, that tell the system how to approach different tasks. Things like a Patent Checker, a Prompt Enhancer, and an App Auditor. These skills shape how the AI behaves when I ask it to do something specific.
The challenge is that when you’re working alone, as a solo founder or independent builder, you don’t have 100 people giving you feedback on whether your skills are actually good. You’re guessing. You ship something, it kind of works, and you move on. You never really know how much you’re leaving on the table.
That’s the gap I wanted to close.
