AI straps a huge rocket to iteration

AI has changed how we build businesses, products and software. Implementation is increasingly cheap, but the quality bar hasn’t moved. If anything, there are fewer excuses to ship low-quality products.

The real leverage is iteration. You can go from idea to reality at incredible speed. The mistake is treating the first pass as the finish line instead of running the loop again and again until the result shines.

I’m building Plural Matter this way. Small repos, narrow jobs, baked-in dogfooding.

As the amount of generated code grows, I’ve stopped looking at the code itself. The product is the outcome, the code is just a means of getting there. AI writes code that works. I care much more about the architecture, the boundaries between systems, and the product itself.

The goal is still customer value and a polished product.

Your product should be clean and easy to iterate on. AI might have changed building but it hasn’t changed what good looks like. Ideas are fast to prototype, but speed alone doesn’t get them to the quality bar. Great products require decisiveness, but AI (and humans) can overbuild or overreach without direction.

AI gives you an enormous solution space before you blink. Judgement is about deciding which direction is worth another pass and which path will lead you off a cliff.

The bottleneck moves from making changes to choosing which changes are worth making.

Here’s what that looks like for me:

I start with an idea then pressure-test it and shape it further through conversations with friends and AI.

I use AI design tools to turn it into initial mockups, from a brand system generated in a similar way. These mockups are surprisingly good out the gate, but they almost always need judgement and taste to remove and refine rough edges: content that adds nothing, pointless eyebrows, unnecessary detail, or copy that’s hard to read. Chip away at the mockup until you have the beautiful essence of it.

Then I task a powerful agent to make a build plan, based on as much context as I can provide + the refined mockups. It’s usually solid from the start.

Recently for the actual build I’ve been using a “dumber” AI as a coordinator which outsources work to “smarter” AIs when needed, following the “idiot boss theory”. So far the results have been promising: the quality seems roughly equal to using a state-of-the-art model for everything, but it burns far fewer tokens.

From there, I manually test, iterate, and ship to production. Shipping constantly beats letting one long-running branch accumulate complexity. It forces you to chunk things out into useful milestones. A series of cupcakes instead of a wedding cake.

By the end of this first loop you have something good enough to put in front of people. That’s where the next loop starts. The care you put in keeps slop at bay.

The second you put something in front of other people, the cracks start to show. The ones that seem so annoyingly obvious with hindsight but were invisible beforehand. People hold the product differently to you. Their vision of how it should work isn’t always the same as your own.

That’s why it’s good to get things into people’s hands as fast as possible. Spinning up landing pages, billing systems and early betas quickly. The faster you meet reality, the faster you can get out of your own bubble and find out what really needs work.

Getting comfortable sharing early is a skill worth building.

The best builders won’t use AI to stop after a single pass. They’ll use it to make far more passes, in far more arenas.

AI straps a huge rocket to iteration. It doesn’t lower the quality bar. That was always on you.