Two Weeks of Work. Two Hours. Here’s How.

Last week, a CEO of a PE firm sent me a long list of feature improvements for SearcherOS. I replied “done” two hours later. He picked up the phone immediately.

“I was an engineer. That was at least two weeks of work. How?”

I'm a guy with zero Python skills. So how does that work?

The Shift

About six months ago, there was a massive shift in AI models with the introduction of Opus 4.5. That's when developers at the major shops stopped writing their own code and started orchestrating teams of developer agents.

That was when I went all-in.

The first time I tried building with AI, I asked it to build me a full CRM. It gave me a beautiful mess. Nothing connected. Nothing worked. I almost walked away.

Then I learned the real skill isn't prompting. It's decomposition.

Decomposition

If you tell Opus 4.6 to go rebuild a major CRM today, it will fail every time. That's not how you do it.

You ask it to break the problem down into phases. You write a spec for the first phase. Then a spec for each feature within that phase. Then atomic-level tasks that a junior engineer could execute to actually build that feature.

If you have the patience to break the problem down into smaller and smaller pieces, there is nothing you can't build.

Agents Grading Agents

But decomposition is only half of it.

I write code in English. My architect, Bill, translates that into specifications. My engineer writes tasks. My supervisor tells an agent what to write and then reviews the work. My QA engineer validates the output. Bill then tells me how we did and what needs to happen next.

Agents supervising and grading each other. That's how a two-week feature list gets done in two hours.

Where to Start

People always ask how they can learn to do this. I've tried teaching it, but that doesn't work nearly as well as you'd think. You need a pain point in your life that you want solved.

Download your favorite voice-to-text app — 3x faster than typing — and just talk to Opus 4.6 for five minutes about your problem and how you might solve it.

One problem. Five minutes. That's how every system I've built started.