How we taught software to run a boutique design process

For the makers and the curious: this is how you turn a hands-on, judgment-heavy, sit-with-the-client design service into software that runs itself, and the parts that were genuinely hard. Honest about scope, because the version shipping first is a real beginning, not a finished cathedral.

The starting point was the method from earlier in this series. Eight steps we'd run by hand across a year of clients. Audit, study, keep-rewrite-cut, find the one true thing, show real directions, write true copy, build clean, leave a record. The whole bet was that this method was reliable enough to automate without collapsing into the generic average we keep warning about. That bet is the entire engineering story.

First lesson: don't build one giant prompt. The naive version is "AI, redesign this website." You get mush, because you asked one pass to audit, strategize, write, and build all at once, and it does all four shallowly. The method exists because those are different jobs. So the engine mirrors the eight steps as distinct stages, each with one job, each handing a clean result to the next. The audit stage only audits. The copy stage only writes. Structure beats a wish.

Second lesson: different jobs want different models. This surprised people, but it matches how we worked by hand. Some stages are taste and voice, getting the copy to sound like a real, specific owner. Some are precise construction, building clean markup that actually works. We don't use one model for everything. Stronger reasoning where the work is voice and judgment. Fast, exact models where the work is build. The same instinct a good shop has about who handles which part.

Third lesson, the hard one: the seeing doesn't fully automate, so don't fake that it does. The most valuable step is finding the one true thing, and it's the step most resistant to automation, because it's judgment about a specific business. The engine got good at surfacing strong candidates: it reads the site and the business and proposes the facts and stories worth leading with. What it does not do is pretend to have the final say. That's why the whole product is recommend-and-approve. The engine does the repeatable bulk of the work and drafts the last mile. A human confirms the spine before anything goes live. We could have hidden that seam and demoed "fully autonomous." The honest architecture is also the better one.

Quality gates, because speed without checks is just fast garbage. A generated site can be quick and quietly broken: a claim that doesn't match the business, a section that contradicts another, a layout that falls apart on a phone. So there are checks between stages, including a coherence pass that asks whether the finished site hangs together and sounds like one business, not a quilt of plausible parts. The goal isn't "a website appeared." It's "a website appeared that a careful person would sign their name to."

What was genuinely hard: getting the engine to be specific instead of safe. Left alone, models drift to the category average, the exact "could be anyone" copy that's the original sin of bad websites. Most of the real work was pushing the system, over and over, to find and lead with what's only true about this one business. Specificity is the brand. Specificity is the thing software fights you on. That tension was most of the build.

Honest about scope. The first version does the core loop: paste a URL, get a real, specific, owned redesign you review and approve. It's a beginning we're proud of, and it's not everything we want. Billing, one-click live relaunch, the deeper custom work, all roadmap, not shipped. We'd rather tell you that than imply a finished product and let you find the edges yourself. The method is real, the engine is real, and the honest framing is part of the craft.

If you build AI products, here's the transferable lesson. The magic isn't a clever prompt. It's decomposing a real expert process into honest stages, using the right model for each, gating for quality, and refusing to hide the seam where human judgment still belongs. Do that and you get a tool people trust. Skip it and you get a demo that falls apart the second a real business uses it.

We're launching soon, Product Hunt included. If you want to watch the seams and tell us where they show, that's exactly the feedback we want.