# Why GPT Still Needs Work for Trip Planning Recently, I had the chance to take a discounted train ride to Paris, accompanied on the train by a [friend and labmate](https://abhinandshibu.com) who used the train to ultimately go to the CDG airport to catch a flight back home. I was not planning to stay overnight, so I had to catch the train back to Lausanne on time. At the same time, I didn't want to depart from Lausanne very early, so as not to harm my already fragile sleep schedule :) As a result, my final purchased tickets allowed me to stay in Paris from ~ 11:10 to 17:50. I had long wanted to visit Paris and most of its main sights again, since my last trip several years ago. However, with a total of 6:40 hours of available time, including the time for lunch, snacks, and drinks (especially in that week with an ongoing heat wave), that seemed impossible. Having just watched the livestream of OpenAI's GPT-5 announcement a few days earlier, including the demo of GPT-5 providing a planned schedule after accessing the user's calendar, it left me wondering: can I use GPT to help me visit the key places in Paris in such a short time, while still keeping buffer times for unexpected plan changes and the time for lunch/snacks? Moreover, as my last trip to Paris was a long time ago, could GPT help me have a quick-to-read summary of the important points to know about each place along the way, so that I can quickly learn about the sightseeing places without eating into my already-packed schedule? This blog post explains my experience in two parts: first, using GPT for planning, and second, for learning information about each place. I will then conclude with my thoughts on how the models and the user experience can be improved for the common task of trip planning. ## GPT for planning times I initially started with a zero-shot prompt by simply giving GPT the start and end times. It had a thinking time of about 3 minutes, and then provided me with a schedule that left out several famous landmarks in Paris (e.g., Eiffel Tower, Avenue des Champs-Élysées, etc.). This forced me to adjust the prompt by gradually adding more information, one piece at a time, through editing the initial prompt (it should be noted that merely asking a follow-up question to adjust the plan led to direct schedule edits, failing to reschedule the whole plan from the ground up to satisfy the new change properly): - **Modes of transportation:** To best enjoy the atmosphere of Paris, in my prompt, I specified that the model should mostly provide walking directions (and only use metro A) if there is really no other way that fits in the schedule, B) if walking wastes a lot of time that can be spent on seeing places, or C) the walk is more than 30 minutes so that it gets boring). I specifically had to add the last point to prevent the model from suggesting that I walk entire stretches of notable streets (e.g., Champs-Élysées), given both the packed schedule and the hot weather on the day of the trip. - **Lunch place:** While the model was able to give suggestions for a lunch spot, it suggested places that either required prior reservations (which wasn’t ideal, as my lunch time could have been easily shifted due to something as simple as a train delay), had low review scores on Google Maps, or were closed (either permanently or for the summer holidays). I expected the model, especially the thinking mode, to use the web search functionality before replying back with the information of the suggested restaurants. - I finally added the following to the prompt: `For any restaurant you recommend, you should check and confirm these online in their website. The requirement: they don't need reservations, they are not a very fancy restaurant (but more casual), and they are open at that time and date (e.g., not on a summer break; this can be verified in the internet).` - This part took the most iterations for GPT to get right, but at least the final outcome was *delicious* :) - **Lunch and coffee break timing:** I wanted to have lunch around 12:30 to 1:00 PM, not earlier or later. However, I didn't expect the model to know this information without my input. What the model should have maybe known, however, was to not put the coffee break directly after lunch in the plan, but after a bit more sightseeing. As a result, I had to add the following to the prompt: `There should be time (and other activities) between the lunch and café, not directly after each other.` - **Places to visit:** The model tended to not include many of the top tourist attractions in Paris, likely because it assumed they wouldn’t fit into the packed schedule. As a result, I had to mention the names of places manually in the prompt (which I found via a simple Google search). Automating the search for places using another web tool call might have been a missed opportunity on ChatGPT's side. Moreover, it didn't include a souvenir shop in the plan, which I had to ask the model myself (not necessarily a model fault, however). - **Low amount of buffer times:** The initial prompt led to no buffer time for any possible delay on the train arriving in Paris, as well as only 15 minutes of buffer time before getting back to the train station (which was risky for a first-timer, given how huge the station is). Moreover, no time was specified for using WC, if necessary. I had to specify manually: `Have gap times throughout the schedule, for things such as going to WC etc, and also if something unexpected happens.` - Regarding WC: I hadn't known (or remembered) that Paris has a high number of public toilets throughout the city. When I asked about including WC, it would have been nice if ChatGPT had mentioned this (but it didn't). - **Possibly inaccurate ETAs:** I saw in the reasoning / thinking summary of ChatGPT that the model noted it did not have access to a mapping tool call. Adding this can possibly have a significant effect on the accuracy of the ETAs and the usefulness of GPT as a travel planner. Finally, given the fact that I had set the model chooser to Auto (to choose automatically between the base and thinking models), on several runs, it chose the normal model automatically, and in others, the reasoning model. Whenever the normal model was chosen, I changed it manually to use the thinking model, to force a more thoughtful and planned answer. Due to all of the issues above, in the end, after receiving the final output, I opened a maps application myself, checked all the suggested locations and distances manually, and adjusted the planning. Not fun, yes, but waiting for GPT to improve at trip planning would have possibly taken longer than just adapting the schedule manually myself... ## GPT for learning about Paris While generative AI had its own set of issues when creating the schedule, it generally worked very well for the task of learning facts about the different landmarks in Paris. I initially prompted the model to provide for each place a short summary, key points as bullet points, and less-known interesting facts. With that prompt, the answers were pretty long, so I couldn't fit each landmark into a maximum of 2 minutes (the time I intended to spend on each description while reading during the trip, to minimize reading time and get straight to the key points). When I asked GPT-5 `this is too long! I want to read it max 1-2 mins per place`, the summaries became too short to be useful. When I asked it to increase the summary again, although it followed the request, I found the response to use difficult literary words, which didn't feel suitable for a quick on-the-go read. Moreover, the summaries were initially not factual enough, rather being written with a poetic and literary style, like story books. Also, the key points were rather short, e.g., for the Notre-Dame Cathedral, it mentioned "2019 fire" without any further explanation on what was the cause of the fire. I asked ChatGPT to correct all of these issues, which eventually led to a final, proper prompt. In the end, I asked ChatGPT to give each response in a Markdown format, to make editing easier. I finally imported them all into the Pages app on my Mac and exported it as a PDF file for easy use on the go. ## Where to go from here? I believe that the experience of trip planning with generative AI can be improved in two main ways: - From the **model** aspect, the models can have access to mapping API tool calls, enabling them to obtain information on A) distances between places, and B) details of each place (opening hours, user reviews, etc.). With that said, the models will still likely have to rely on web search or the model’s internally encoded knowledge for *map exploration*, as it should be difficult to come up with agents that can freely explore a map (each movement on the map, along with seeing all details on the map, and then generating the next step, should be a computationally expensive task). With that said, future models might enable agents with higher abilities to navigate maps and explore environments within a reasonable cost and time window. Moreover, it would have been great if the models asked a set of follow-up questions from the user, before jumping to provide the plan, e.g., asking about what type of traveler they are, whether they prefer cultural or historical landmarks, whether they would want to buy souvenirs, etc. - From the **user interface and user experience** aspect, adjusting prompts and then having to check all the details manually can feel daunting for many people, making them either avoid GPT altogether for this task, use a previously-prepared travel material (with the cost of not being personalized to the exact time constraints or interests of the users), or follow a less-than-optimal plan. A user-centric app for travel planning can verify all of the details of the trip before showing the final itinerary to the user, and can also be integrated into the map applications by automatically suggesting the route to the next destination on the list whenever arrived at a landmark, delivering the information of the landmark as a short auto-generated podcast in your headphones, and re-routing in case of any unexpected deviations from the plan. (I have not used any AI-assisted app for travel planning; if any reader has an opinion, I’d love to hear about it!) Moreover, in my personal experience, I found getting the phone out of the pocket every time for checking the next landmark was a bit tedious. While the Maps app on the Apple Watch proved very useful for navigating around Paris, I had to enter the landmarks manually from my plan into the app. For anyone developing an app for travel planning, having a companion smartwatch app (or even better, in the future, an AR app on smart glasses, showing the next destinations to go using arrows in the air or on the ground) can all be viable solutions to this problem. If any of you have had the experience of planning your travels using generative AI, and have supporting or contradicting opinions to this post, it would be great to hear from you :)