Travel planning guides

AI planning · 7 min

We tested an AI trip planner with a three-day Tokyo brief

The first draft omitted day three and did not account for travel between stops. Here is the review process we used before treating it as an itinerary.

TripNotePlan 편집팀

TripNotePlan 개발·운영팀

Published · Updated

How this guide was made

  • We sent a controlled, fictional three-day Tokyo brief to Gemini 3.1 Flash-Lite and preserved the first response before editing it.
  • We compared the requested constraints with the returned days, place count, route notes, and items that still required human verification.

On July 16, 2026, we ran a controlled test with fictional data: one adult, three days in Tokyo, food and culture as priorities, and no more than four core stops per day. The response created only day one and day two. It also listed time at places without adding the travel time between them. The useful part of the test was not the draft itself, but the checklist of issues a traveler still had to resolve.

The brief was designed to be easy to audit

We did not include a real name, hotel booking, or user itinerary. The input was limited to facts that could be checked against the output: traveler count, destination, number of days, interests, and a daily stop limit.

A testable prompt is more useful than an open request for “the best trip.” If the prompt names three days, the output should contain three days. If it requests meal and buffer time, those blocks should be visible rather than implied.

Controlled input used on 2026-07-16
FieldInputWhat it let us check
Traveler1 adultNo family or group assumptions
DestinationTokyoSingle-city structure
Duration2 nights / 3 daysWhether all three days appeared
PrioritiesFood and cultureWhether suggestions stayed focused
Daily limitUp to 4 core stopsWhether the draft became overloaded

What the first draft missed

The most visible failure was the missing third day. The draft also treated estimated time at a place as if it were enough to build the day, but moving between stations and entrances was not included.

The response warned that opening hours, reservations, and live availability still needed verification. That warning matters: an AI-generated place name is not evidence that the place will be open or bookable on your travel date.

Request versus first response
CheckRequestedReturnedHuman action
DaysDays 1–3Days 1–2Add departure-day structure
Travel timeMeal and buffer timeTime at placesInsert route and transfer blocks
Operating detailsRealistic scheduleRecheck warningUse official venue sources
ReservationsActionable planNot confirmedMark fixed bookings separately

How we would turn the draft into a usable plan

First, add the missing departure day with check-out, luggage, and airport or onward travel. Second, calculate travel between every pair of stops using the transport mode you will actually use. Third, lock timed bookings and place meals and rest around them.

Optional places should stay outside the core schedule. In TripNotePlan, the draft can be assigned to trip days and reordered after the map review. The AI response is the starting material; the day count, route, and booking checks are the editorial work.

  • Match the number of requested and returned days.
  • Separate time at a place from time traveling to it.
  • Verify opening hours and tickets on official sites.
  • Keep at least one optional stop removable without breaking the day.

A prompt structure you can reuse

A concise reusable brief is: “Plan three days in Tokyo for one adult, focused on food and culture, with no more than four core stops per day. Show travel blocks, meals, rest, and which opening hours or bookings require verification.”

After generation, run a five-part review: day count, first and last stop, travel time, fixed reservations, and optional stops. Do not call the itinerary finished until those five checks have evidence outside the AI response.

Action checklist

  • Confirm that every requested travel day exists.
  • Add transfers and station navigation between stops.
  • Verify hours, tickets, and reservations with current official sources.
  • Mark optional stops before following the plan.

Sources checked

  1. Google AI for Developers · Gemini API models (checked )