Light Rail Living went from idea to public beta in fourteen calendar days. The product is small: three questions, one itinerary, no sign-up. But the build pattern is worth writing down, because it's how we ship now, on every project.
The first day was a framing day. We didn't write any code. We sat with the problem (Phoenix's transit is genuinely good but reads as confusing to anyone who hasn't memorized the line) and used an AI assistant to pressure-test five different framings of the product. The one that survived was the most boring: a planner, not a chatbot, not a social network. A planner that uses AI to make a curated index feel personal.
Days two through six were code. Claude Code wrote most of the first pass: the schema, the API surface, the AI integration with our own venue index, the homepage, the responsive layout. Our role was structural: editing the prompts, naming the components properly, deciding what to keep generic versus what to brand.
The unglamorous middle came in days seven through ten. This was content work: writing the venue index, sourcing real photographs, fixing stops where the model got walking distances wrong, and instrumenting analytics. None of this is work AI replaces wholesale. It was a useful drafting partner, but every venue entry was reviewed by a person who has actually walked the block.
Days eleven through fourteen were tightening. We added the weather context, the time-of-day shaping, and the small details that make the product feel like it was made for one specific city by people who live there. We also wrote the copy that admits, in the app, that the recommendations come from an AI model. People respect that more than vague AI marketing.
What the AI did: drafted code, wrote first-pass copy, generated venue context, and acted as a research partner for the framing work. What we still did ourselves: the call on what to ship, the curation of the venue index, the live walks of the rail, the brand decisions, and the dozen calls to local venues to verify hours.
The mistake we see other teams make is assuming AI replaces taste. It doesn't. It removes the friction between taste and shipped software. The best operators in 2026 are going to be people with strong opinions and strong AI workflows, and that's the only way we know how to build.
The AI wrote most of the lines. The shipped product is still ours.