How I used Budge to plan a 3-week trip (founder story) — Budge

I started planning a trip to Japan in November. The trip was three weeks in late March — cherry blossom season, which I knew was complicated timing — covering Tokyo, Kyoto, Hiroshima, Osaka, and possibly a night or two somewhere more rural. I'd been to Japan once before, briefly, and wanted to do it properly this time.

The planning started the way it always starts: sixteen browser tabs in a week. Two Reddit threads about cherry blossom timing. Four blog posts about Tokyo neighbourhoods that contradicted each other on whether to stay in Shinjuku or Shibuya. A Kyoto accommodation thread where everyone had a strong opinion about ryokan versus business hotel and nobody agreed. A JR Pass calculator that required me to know my exact itinerary before I could calculate whether the JR Pass was worth it, which required me to already have my itinerary. The familiar planning spiral.

The broader case for using AI in travel planning covers why this kind of research spiral happens and what AI does to address it. What I want to describe here is what it actually felt like to use a purpose-built tool — Budge, which is the thing I built specifically because I kept running into this problem — rather than the generic version.

What I tried first

Before using Budge, I spent two sessions with ChatGPT and one with Claude (the general-purpose assistant, not Budge) trying to build the Japan itinerary.

The ChatGPT experience was the same as most people describe it: useful when I gave it a detailed brief, not useful when I treated it as a search engine. The first message I sent — "help me plan a 3-week Japan trip in late March" — produced the standard Tokyo-Kyoto-Osaka circuit with sensible suggestions and absolutely no relationship to my specific situation. When I gave it more context — the previous Japan trip, what I'd found dull (too many temples), what I wanted more of (food culture, architecture, less famous places) — the quality improved substantially. But every new session started from scratch. The third conversation picked up nothing from the first two. I re-explained myself each time.

The specific limitation that drove the Budge decision: I wanted to have an evolving conversation about the trip over several weeks, where each new question built on what I'd already decided. ChatGPT's conversation memory within a session is excellent; between sessions it's zero unless you're using the paid memory feature, and even then it's inexact. I wanted something that knew my trip as a persistent object — the dates, the preferences, the decisions already made — and could add to it incrementally.

I've described how different AI tools compare on this dimension — the persistent context problem is the core reason purpose-built tools differ from general AI.

What Budge did differently

The first session I used Budge for the Japan trip was about cherry blossom timing. I described the trip — late March, three weeks, travelling with my partner who had never been to Japan, specific interest in food and architecture over temple circuits — and asked whether the late March timing made sense or whether I should consider a different period.

The response did two things differently from what I'd gotten from generic AI. First, it asked a follow-up question before answering: had I already booked anything non-refundable? The answer changed the advice. If I'd already bought flights for late March, the question was how to plan around the cherry blossom uncertainty. If I hadn't, the question was whether late March was actually the right choice given what I'd said I wanted from the trip.

I hadn't booked anything yet. The advice shifted: given that we wanted food culture and architecture over famous scenery, autumn was probably a better choice. The autumn foliage is as visually spectacular as cherry blossom and the crowds at the famous sites are substantially smaller. More specifically: the type of Japan experience I'd described — eating at small restaurants, exploring neighbourhoods, avoiding tourist-conveyor-belt sightseeing — doesn't depend on any particular season, and late October to mid-November would make the sites I specifically wanted to visit less crowded and the booking logistics less compressed.

This was the right answer. It wasn't the answer I'd been working toward — I'd mentally committed to cherry blossom season — but it was correct given what I'd actually said I wanted. A general AI would have helped me plan the late March trip I'd asked about. Budge pushed back on whether late March was the right premise.

Building the itinerary across multiple sessions

Over three weeks of planning, I had eight or nine conversations with Budge about the Japan trip. Each one picked up where the previous one had left off. By the fourth session, I wasn't re-explaining that I wanted food and architecture over temples, that my partner hadn't been to Japan before, that I wanted to avoid the tourist-restaurant trap, that we were travelling in late October. The tool held that context and filtered everything through it.

The specific things that worked well:

Logistics reasoning with accumulated context. In session six, I had a rough itinerary and asked whether the routing made sense. The response took my existing decisions (5 nights Tokyo, 4 nights Kyoto, 2 nights Hiroshima, 2 nights Osaka) and pointed out a specific inefficiency: I had a day trip to Nikko scheduled from Tokyo and a day trip to Nara scheduled from Kyoto, but the day I'd put Nikko on was the same day as an arrival from a bullet train from Kyoto that I'd moved earlier in the session. The logistics contradiction was caught because the tool held the full plan.

Follow-up questions that narrowed to specifics. When I asked about Kyoto accommodation, the first response was a question: were we prioritising being near the main sights, staying in a traditional ryokan experience, or minimising the accommodation premium that central Kyoto carries? The answer changed the recommendation. We ended up in a business hotel in Fushimi — close to the Inari shrine we'd decided to prioritise, cheaper than the Gion area options, still in central Kyoto for evening walks.

Proactive flags I hadn't thought to ask about. In session seven, when I pasted in a draft day schedule for Kyoto, the response flagged that one of the experiences I'd listed — a specific tea ceremony reservation — typically requires booking 4–6 weeks in advance and was likely to be unavailable if I tried to book it the week before the trip. I hadn't thought to check this. The trip was 10 days away. I checked, confirmed it was unavailable, and found an alternative. That flag alone justified the tool.

Where Budge got things wrong

The honest section: two things didn't work as well as I'd hoped.

Local restaurant recommendations remained unreliable. I asked specifically for ramen shop recommendations in the area near my Kyoto accommodation and got a list of names with confident descriptions. I checked two of them on Google Maps. One had permanently closed; the other had reviews from recent visitors describing a significant decline in quality from the period the description implied. This is the known limitation of AI restaurant recommendations — training data doesn't track current status — and Budge didn't solve it differently from any other AI tool. For specific local recommendations, I ended up using Google Maps sorted by most recent reviews, which remains the only reliable source.

It occasionally optimised for coverage over pace. When I asked for a Kyoto day plan, the initial response included more activities than a comfortable day allows — the kind of itinerary that's physically achievable if nothing takes longer than expected and nothing catches your interest enough to linger. When I pushed back ("that looks like too much for one day"), it revised immediately and the revised version was better. But I had to push back, and a tool with a stronger model of travel fatigue would have started with a more realistic plan.

Both of these are known limitations of the current category. The restaurant problem is structural; the pacing problem is a prompt calibration issue that I worked around. Neither broke the experience.

How the trip went

We went in late October. The timing was right — the Kyoto maple colour was at the beginning of its peak when we arrived, the famous temples were busy but not impassable, and the food recommendations from the planning sessions (specifically the Fushimi neighbourhood izakaya strip that Budge had flagged in session four) were consistently excellent.

The planning felt less stressful than previous trips of similar complexity. Not because the logistics were simpler — Japan in any season has complexity — but because I'd worked through most of the decisions before the trip rather than during it. The questions I might have spent hours on in-trip (is this routing efficient? is this accommodation area right for what we want?) had been settled. That freed up the trip itself for the things that can't be planned: the ramen shop we found by smell in Yanaka, the morning we spent longer than expected at a tool museum in Osaka because it turned out to be extraordinary, the afternoon in Hiroshima that changed the rest of the trip's register.

The honest truth about what the tool changes

Budge is free to try — you don't need to plan a 3-week trip to find it useful. Start with one destination you're curious about and ask it something specific.

The tool doesn't replace the judgment call of being there. It doesn't know what the ramen shop smells like or how a street feels at 7am in the fog. It doesn't account for the way travel fatigue accumulates differently for different people, or the spontaneous decision to extend a day in one place because you're not ready to leave.

What it changes: the quality of the skeleton you arrive with. The decisions that would have cost you a day to sort out on arrival are pre-sorted. The flag that your tea ceremony needs advance booking gets caught in a planning session, not on the morning you wanted to go. The itinerary has been challenged — "is this actually what you want?" — rather than built from search engine results that reflect other people's priorities.

The trip is still yours. The planning is just less of a chore.