The record so far
Computed at build time from the committed weekly metrics — not typed in by hand. Covers August 2025 through the week of 21 September 2026; weeks with no commits are omitted, so the count above is weeks of activity within a 57-week span, not an unbroken run.
The contract
I publish the parts where I was wrong.
Most writing about building with AI is written from memory, shortly after something impressive happened. This is written from the record. Every factual claim traces to a commit, a file diff, or a dated artifact — and when the data contradicts something I’ve published here, that gets said in the same piece rather than the next one.
One of the early pieces is about how I accumulated 19,000 lines of test coverage in two days, felt very good about it, then quietly let the suites covering my three highest-traffic screens sit disabled for three months — finding out by accident, four days before the thing reached a single user.
Anyone can write the version where the tooling is remarkable and the author is prescient. That version is worthless to you, because you can’t tell it from a version that’s made up.
The corresponding ask: if I get something wrong, tell me. I’d rather be corrected than agreed with.
What this isn’t
- It isn’t build-in-public. No revenue graph, no subscriber count, no running commentary on my feelings about the week. Build-in-public optimizes for narrative momentum. I’d rather optimize for being right.
- It isn’t a demo reel. There’s a whole genre that shows you a thing being built in four minutes and stops before the part where you have to maintain it. The four-minute part is real. It’s also the least useful part, and it’s thoroughly covered by other people.
- It isn’t thought leadership about where AI is going. I have no idea. What I have is a growing body of decisions with recorded outcomes, which is a different and much narrower thing.
Who this is for
People making these decisions rather than reading about them. Product leaders and founders working out what changes operationally, not just what becomes possible. If you’re responsible for what a team can deliver next quarter and you’re trying to separate real leverage from the appearance of it, we’re looking at the same problem.
Posts
New here? Start with The Parts That Don't Demo Well — it sets out the contract the rest of these run on.
Models Got Better. Rework Got Worse.
The spike I assumed was users turned out to be the first time anyone looked.
My Agents Started Talking Without Me
I spent months trying to get out of the middle. The thing that worked was one I'd ruled out without testing.
The Test Rig
How I went from killing my own car battery to supporting 1,000s of vehicles.
I Traded 10,000 Lines of Tests for 380
I booked a cost reduction as profit. It was a liability.
Verbatim Spine
The constraint stopped being how much I could build. It became how much I could check.
Two Agents, One Checkout
I'd been treating a systems problem as a management problem.
72× the Output. Half the Rework. Here's What That Number Hides.
A year of throughput data, including the four reasons not to trust it.
The Parts That Don't Demo Well
A year of operating four products where most of the work wasn't mine.
Who’s writing
Twenty years across product and engineering, most of it spent responsible for outcomes rather than output. Until July 2026 I was Chief Product Officer at SignUpGenius, where I built the product function from the ground up and helped drive 14x revenue growth across SaaS, advertising and payments for a platform serving 70M+ users.
I got there by way of engineering, and I’ve been starting companies the whole time — including an aftermarket auto parts business I grew past 40,000 SKUs and sold in 2011, which is most of the explanation for why one of the products here is an automotive app.
Why any of that matters: I know what normal looks like. I’ve run the planning cycles, owned the consequences of being wrong about capacity, and sat through enough technology cycles to be suspicious of my own enthusiasm. When I tell you something changed, I have a baseline to measure it against.
- The studio: rbzilla.com
- The products: kaijumechanic.com and codextitan.com
Subscribe
Free, and it stays free. Unsubscribe in one click.