The machine that makes the machine

Everyone wants to build the product. Fewer people want to build the thing that builds the product. But that second thing is where the real leverage lives.

When I started shipping AI-generated content at scale, my first instinct was to make each piece great by hand. It worked — and it didn't scale. The breakthrough wasn't a better prompt. It was a pipeline: research, draft, judge, publish, measure, improve. A flywheel that gets smarter every turn.

Build the loop, not the artifact

An artifact is a single output. A loop is a system that produces artifacts and learns from how they perform. The artifact decays the moment you ship it. The loop compounds.

• ◆Measure something real (citations, conversions — not vanity polish).
• ◆Decide what to change based on that measurement.
• ◆Improve the input, not just the output.
• ◆Let it run. Then watch one example end-to-end before trusting the aggregate.

If you can't watch one real example travel the whole loop, you don't have a loop. You have a demo.

That last line is a rule I live by now. It's easy to declare a system "done" because the code merged. It's done when a real input goes in one end and a real, good result comes out the other — observed, not assumed.