Plans is the AI-powered itinerary tool I've been building for travel agents. It started as a weekend project sixteen months ago and is now used by a small group of travel agents to put together client itineraries. This is what I've learned shipping it.
The original problem was small and concrete. A travel agent I knew was spending Sunday afternoons assembling day-by-day itineraries from scratch — flight times, hotel descriptions, transfers, restaurant notes — for clients she'd already had a discovery call with. The information existed. The shape didn't. AI was unusually good at the shape.
The first version was a single prompt over a Google Doc. Paste a brief, get an itinerary back. It worked badly enough to ship and well enough that two of her colleagues started using it within a month. That's the bar I keep watching for: bad enough to embarrass me, useful enough to spread.
The hard work since then has been almost none of it about the model. The model picks the right hotel most of the time, the right restaurant some of the time, the right tone almost always. The rest gets edited by an agent who knows the destination better than the model ever will. That edit step is where most of the design work has gone.
The biggest single lesson: the agent isn't a reviewer of AI output. The agent is the author, using AI as their first draft. Every interface decision in Plans flows from that. The diff view, the inline edits, the saved variants, the agent-tone settings — all of it treats the human as the writer and the model as the typist.
The biggest single mistake: I built a multi-step "agent" early on that was supposed to chain together flight lookup → hotel match → daily plan. It worked in demos and broke in real use. Every step compounded the previous step's errors and the explanations afterward took longer than the original task. I rewrote it as three independent single-shot calls and the system became reliable overnight. That lesson now lives in every AI build I take on.
The numbers I track each week: agents using Plans daily, itineraries generated, edits per itinerary, and time-from-brief-to-sent. The last is the one that tells me whether the product is actually useful, because that's the variable agents care about. It's the one moving in the right direction.
The reason this matters for the practice: every fixed-price AI build I take on for someone else is informed by what Plans has taught me — what works, what breaks, where users actually live. Plans is the apprenticeship. Clients get the version of me that already made the mistakes on his own product first.
Plans is the long bet. Tee Sheet and Whip are quietly in development behind it. If the next year goes the way the last one did, the writing here will get less abstract and more specific — fewer essays, more numbers.