How I Plan Every Trip with AI in 2026: The Killer Prompt, the Workflow Around It, and What It Means for Hotels

Two years of trial and error with my own family. Four stages. One copy-paste prompt. And a closing word for everyone on the supplier side of the booking screen.

Seven or eight months out from a trip, I don’t book anything. I open Claude (or ChatGPT, or Gemini, depending on the day) and I dream out loud.

That’s the part most people skip.

The dominant mental model for AI travel planning right now goes something like this: type “plan me a 7-day trip to Italy with kids” into a chatbot, scan the day-by-day output, screenshot it, move on. If it’s wrong, blame the model. If it’s right, marvel at the magic. Either way, treat the AI like a vending machine. One prompt in. One answer out.

Honestly, that’s how I used these tools at first too.

But after a couple of years of actually using AI to plan, book, and survive real trips with a family in tow, I’ve come to a simple conclusion: the people getting genuinely great results aren’t writing better prompts. They’re running better workflows.

A workflow respects how trips actually get planned. It moves in stages. Dream. Plan. Travel. Reflect. Each stage uses AI for a different job. Each stage has a model that’s best suited for it. And each stage quietly builds on what came before.

Let me show you what mine actually looks like, including the one prompt I copy-paste at the start of every trip to figure out where to go in the first place.

Why workflow beats prompt

If you’ve ever asked an AI to “plan me a trip” and gotten back a beautifully formatted itinerary that mentions four restaurants that don’t exist and a museum that closed in 2018, you already know the failure mode.

The issue isn’t the model. It’s the framing.

A single prompt forces the AI to make every decision at once: destination, duration, daily structure, budget, vibe, transport, food, kid logistics, weather, your random allergy to humid air. With no follow-up loop. No iteration. No memory. Recent benchmarks show that something like 90% of AI-generated itineraries contain at least one factual error, and state-of-the-art models produce a fully viable end-to-end trip plan in roughly 4% of attempts when you stress-test them on complex multi-constraint trips. That isn’t an indictment of AI. It’s an indictment of treating AI as a vending machine.

A workflow splits the problem the way you’d split it with a human travel agent who actually knew what they were doing. First we figure out where you should go. Then how long, and roughly what shape. Then day by day. Then we adapt in real time. Then we save what we learned for next time.

You don’t get there by typing harder. You get there by setting the table differently.

Stage 1: Dream. The killer destination prompt.

The first stage is the one most people get wrong, because they ask the AI a thinly disguised version of: “Where should I go?”

The AI’s response to that is, broadly, “Bali.”

You can do much better. The trick is to give the AI enough constraints, context, and friction points that it can’t reach for defaults. The single highest-leverage field, in my experience, is “places I’m tired of hearing about.” It singlehandedly rules out the entire pile of generic suggestions and forces the model to actually think.

Here’s the prompt I copy-paste at the start of every trip. Drop it into Claude (Opus with extended thinking is my pick), ChatGPT (a thinking model like GPT-5.5), or Gemini Plus, fill in the placeholders, and watch what happens.

Quick note before you copy: everything in [square brackets] is mine. Swap it for yours before pasting. Don’t be vague. Vague inputs get vague destinations. The “real reason we’re going” and “places I’m tired of hearing about” lines are the two that decide whether the output is genuinely useful or just well-formatted.

You are an elite travel architect. Treat destination selection as a
constrained optimisation problem. Use your best knowledge of weather
patterns, flight routing, visa rules, and local conditions, but flag
anywhere I should verify independently (visa rules and direct flight
routes change; seasonal weather has variance). Do not suggest
over-touristed defaults unless they genuinely fit. Be direct and
opinionated; if my constraints don't fit together, say so.
MY SITUATION
- Who's traveling: [e.g. me + partner + 2 kids aged 6 and 9]
- Trip length: [e.g. 7 nights, flexible by ±2 days]
- Travel window: [e.g. last 2 weeks of July 2026]
- Departing from: [e.g. your home airport, with code]
- Max flight time: [e.g. 6h direct, or 9h with one stop]
- Total budget (all-in): [e.g. USD 6,000 incl. flights, hotel, food]
- Passport(s) and residency: [e.g. your passport(s) and country of
  residence, so the model can think clearly about visas]
- Weather needs: [e.g. something cool with no rain]
- The real reason we're going: [e.g. we're exhausted and need to
  do nothing / kids' first wow trip / anniversary / mid-life reset]
- Things I love when I travel: [e.g. food markets, swimming in the
  sea, walkable old towns, no driving]
- Things I want to avoid: [e.g. crowds, long transfers, resort
  bubbles, dressing up, time-consuming or expensive visa
  requirements]
- Hard constraints: [e.g. one vegetarian + nut allergy, no malaria
  zones]
- Places I'm tired of hearing about: [e.g. Bali, Maldives,
  Santorini, Dubai]
RESPOND USING THIS EXACT STRUCTURE
## Constraints I'm optimising for
One short paragraph summarising the core limitations and the
trade-offs they create.
## Destination 1: [City/Region, Country]
**Fit:** Strong fit / Good fit with caveats / Stretch, and one
sentence on why that tier.
**The logic:** Why this works geographically and climatologically
given the origin and time of year. Be specific about flight time,
season, and what the weather actually does there in that window.
**The local flavour:** One genuinely specific activity (not
"explore the old town") and one dining insight tailored to the
exact dietary needs.
**The reality check:** One honest deal-breaker or friction point:
specific local cost, transport pain, crowd level, scam risk, or
whatever the guidebooks downplay.
**Verify before booking:** Visa requirement for the passport(s),
and any other point that needs current confirmation.
## Destination 2: [...]
[Same structure]
## Destination 3: [...]
[Same structure]
## My pick if I were you
One destination, one paragraph, why.
## If the constraints don't actually work
Only include this section if my parameters are geographically or
seasonally incompatible. State the conflict directly and offer the
closest workable alternative, including which constraint I'd need
to relax.
Skip obvious answers unless they're genuinely the best fit. No
filler, no hedging like "could be a great option". Commit to a view.

A few small things make this prompt punch above its weight.

It tells the model to flag what to verify rather than confidently hallucinate the visa rule. That single line cuts a meaningful chunk of “AI travel” disasters before they happen.

It treats destination selection as a constrained optimisation problem, which is the actual shape of the decision. You don’t want the “best” beach. You want the best beach given budget, flight time, season, passport, and the fact that one kid won’t eat anything green.

It gives the model explicit permission to say no. A startling number of prompts ask AI to please everyone, then are surprised when the answer is mush. “If my constraints don’t fit together, say so” lets the model push back. It’s the single biggest quality unlock I’ve found.

It demands a reality check for every destination. Most AI travel content reads like it was written by tourism boards. Asking for the specific friction point (the scam, the price spike, the crowd hour, the bus that isn’t running anymore) gets you the kind of insight a friend who’s been there would actually share.

Run that prompt with honest inputs and you’ll get three options that are genuinely tailored to your life, not three pages of “consider visiting the historic old town.”

(One more move: when you get the output, ask the same model to “now stress-test these three picks. What would make me regret each one?” The negative case sharpens the choice in a way the positive case never does.)

Stage 2: Plan. Let the AI interview you, not the other way around.

Once you’ve picked a destination, the workflow shifts.

This is where most people revert to vending-machine mode and type “day by day itinerary for 7 days in Lisbon with kids please.” Don’t.

The trick at this stage is to flip the conversation. Have the AI interview you.

What you get back is a small interview. Some questions you’ll have strong answers to. Some will surface things you hadn’t thought about (Are you OK with kid-noise restaurants or do you actually want a date night somewhere quiet? Is the partner more “see one thing properly” or “see five things lightly”? Are you willing to take a train for a half-day side trip or is the base city the whole point?). Answer them in plain prose. Then ask for a draft.

The day-by-day itinerary you get from a 10-minute interview is materially better than the one you get from a single prompt. Not because the model got smarter. Because it has the right inputs.

Two more upgrades I use here:

Force the AI to think in constraints, not activities. Before asking for a day plan, paste in your hard rules: “No more than 2-3 hours of driving in a day. One sit-down meal per day, the rest grab-and-go. Vegetarian options need to be real, not a sad side salad. One slow morning every three days, because we have kids. Build the days around those constraints.” Thinking models (the ones with extended reasoning) handle this kind of multi-constraint plan dramatically better than vanilla chat. This is exactly the kind of work that hard-mode reasoning is actually good for.

Make it verify itself. End with: “Now check the draft against my constraints. List anywhere it might break them, and propose a fix.” Half the time the model catches its own scheduling conflicts before I do.

Stage 3: Projects. The unfair advantage most people miss.

Here’s the move that quietly upgrades everything else.

Use Projects (or the equivalent in whichever tool you’re in: Claude’s Projects, ChatGPT’s Projects, Gemini’s Gems, Claude Cowork’s session folders). Create one per trip.

Drop in:

• Your draft itinerary (as a file, not just a chat message)
• Booking confirmations (hotels, flights, transfers, reservations)
• Visa or insurance docs if relevant
• Custom instructions specific to the trip (something like: “You are helping me run a 7-day Portugal trip for a family of 4. Kids are 6 and 9. One vegetarian, one nut allergy. We prefer walking and trains over rental cars. Always check what I’m asking against the attached itinerary before answering.”)

Now every conversation inside that Project starts from a position of context. When you’re standing at the wrong train platform on Day 4 and ask “is there a faster way to get to Sintra from here?”, the AI already knows where you’re starting, where you’re trying to go, what time your dinner reservation is, and the fact that you don’t have a rental car. You skip ten minutes of re-explaining the situation and get straight to the answer.

This is the single largest quality-of-life improvement in my travel workflow. After a long day, the last thing you want is to brief an AI from scratch on what trip you’re even on. Projects mean you don’t have to.

And when the plan inevitably breaks (the museum is closed, the train strike is real, the kid is melting down in 38°C heat), Projects let you replan against everything you’ve already committed to without losing the thread.

Stage 4: Travel. AI as the in-trip co-pilot.

This is where AI quietly stops being a planning tool and starts being a sense organ.

A few things I now use AI for in real time, often without thinking about it:

Live menu translation, with context. Point your phone’s camera at a menu in a Lisbon tasca. Get an instant translation plus flag-anything-with-nuts, plus “which of these is the local specialty?”, plus “what should the vegetarian actually order here?” Camera-based translation overlay (Google Lens, ChatGPT vision, Gemini multimodal) has gotten genuinely good. The unlock isn’t just translation. It’s annotated translation.

Signs, customs, etiquette, on the fly. That confusing parking sign with three numbers and a symbol that might be a horse? Snap a photo, ask what it actually means. The dinner invitation from the local family that you absolutely want to accept but aren’t sure what to bring? Ask the model. The hand gesture that just earned you a strange look? Ask.

The crowd-avoidance hack. One traveler I read about asked ChatGPT to suggest the optimal time to visit Château de Chenonceau and was told to wait until 9am the next morning instead of arriving at opening. They walked in as the fifth person through the gate and had a 16th-century chateau effectively to themselves for the first hour. This is the kind of advice that used to require knowing somebody. Now you just have to know to ask.

Real-time pivots when things break. Flight delayed, missed your transfer, kid running a fever, restaurant closed for a holiday you didn’t know about. The single most underrated AI travel use case is “the plan just broke, help me build the next 4 hours.” Because the AI knows your Project context, you skip the “wait, where are you again?” cycle and get straight to options.

Conversational translation when it matters. Two-way voice translation has improved enough that you can have an actual short conversation with someone who doesn’t share your language. Not “ask the way to the train station.” More like “explain to the doctor that the nut allergy is severe and we need to know what was in the dessert.” That’s the kind of moment where you’ll remember how new this all is.

Reading the room you’re in. Take a photo of the inside of the restaurant before sitting down. Ask the model what it tells you about how formal the place is, what the dress code likely is, what local people probably order here. Sounds slightly ridiculous until you try it.

(Honorable mention: shopping. Snap a label in a foreign supermarket and get an instant breakdown of what’s actually in it. Particularly useful when you’re trying to feed a kid with allergies and the packaging is in Greek.)

Two practical notes here. First: pick your model with intent in the moment. Gemini’s tight Google Maps integration is genuinely useful for routing, drive times, and “what’s open near me now.” ChatGPT’s voice mode and broad app ecosystem is great for the conversational in-trip stuff. Claude is sharper on detail and budget-consciousness when you want a thoughtful second opinion before you commit. Second: download an offline language pack before you fly. Connectivity is the assumption that breaks first.

Picking the model: a quick honest take

A working theory I’d offer, based on running this workflow across all three for a few years now and what current benchmarks are showing:

ChatGPT is the strongest for creative brainstorming, narrative drafting, and the messy exploratory front of trip planning. It also has the deepest app ecosystem and the most polished voice and image features for in-trip use. The trade-off: of the three, it’s the most willing to confidently hallucinate. Verify the specifics.

Claude is my pick when I want a careful, opinionated, detail-conscious second opinion. It’s the most budget-aware planner I’ve used and the most willing to push back when your constraints don’t fit. Cowork mode is excellent for the dreaming and itinerary-building stage where you want a desktop-class workspace and longer reasoning. The trade-off: it’s desktop-first for now, so the on-the-go layer of the workflow lives elsewhere.

Gemini has the structural advantage of plugging directly into Google’s Maps, Search, and travel data. For routing, drive times, current weather, and “is this thing actually open today” questions, it’s hard to beat. The trade-off: the live-data confidence sometimes papers over outdated underlying sources, so cross-check.

You don’t have to pick one. You can use all three for different jobs. The workflow is the same. The model is the instrument.

(For the record: I do almost all my “Dream” stage work in ChatGPT or Claude with extended thinking on. Most of my “Plan” and “Project” work also lives in Claude or ChatGPT depending on the trip. Most of my in-trip work happens on ChatGPT and Gemini because they’re on my phone, where the action is. This will change. It’s already changing.)

Stage 5: Reflect. Closing the loop.

The stage everyone skips.

At the end of a trip, spend 20 minutes in the same Project asking the AI to help you write a personal trip review. Not for the blog. For you.

What you get back is a small artifact that pays compounding interest. Next trip’s Dream prompt? You can paste in last trip’s review as context. The model will remember that you actually hate driving more than 90 minutes at a stretch, that you regretted the rooftop pool hotel, that the kids surprised you by loving train stations. The workflow keeps learning about you, trip after trip.

That’s the part most travel AI conversations miss. The interesting question isn’t “can the AI plan one trip well?” It’s “can the AI get better at planning your trips, the more trips you take?” Answer: yes, but only if you set up the loop.

So what does this mean if you run a hotel?

If you’ve made it this far, you’re either a traveler nodding along or a hotelier wondering when I get to the point. Here it is.

Everything I’ve described is already happening, at scale, without you. Your future guests aren’t waiting for your concierge AI to roll out. They’re already shortlisting their next trip using a prompt that has explicit instructions to avoid over-touristed defaults and demand reality checks. They’re already in a Project that holds their booking confirmation, their dietary constraints, and their preferences from three trips ago.

So the question isn’t “should we adopt AI?” It’s “where in this workflow do we actually add value, and where are we just adding friction?”

I’d map it to the same four stages.

Dream stage: be findable to the prompt, not just to Google

When a traveler runs the killer prompt I shared above, the model has to make a judgement call about which destinations and which properties to surface. That judgement is being shaped right now by whatever the model can read about you on the open web, in OTAs, in reviews, in the structured data on your site, and (increasingly) inside answer engines and AI-native search.

If your destination story online is generic, you’ll get matched generically. If it’s specific (the actual neighbourhood feel, the actual signature experience, the actual reason somebody would pick you over a near-identical resort), you have a shot at being surfaced when somebody describes a real human reason for travelling.

This isn’t SEO with a thin coat of AI varnish. It’s a structural shift in how you’re discovered, and it’s already underway. Anyone selling you “AEO” or “GEO” as a buzzword-friendly upgrade to your existing SEO retainer is selling you a bandaid on a broken leg. The real work is content that’s actually distinctive enough to be retrievable.

Plan stage: meet guests inside their workflow

The traveler is now deep into a Project, building day-by-day plans, asking constraint-heavy questions, talking to an AI that knows their kids’ ages and dietary needs.

Most hotels are still meeting that guest with a generic chatbot bolted onto a brochureware site, a “personalised offers” email built on six fields from the loyalty programme, and a booking funnel that has barely changed in five years.

The opportunity isn’t to build “your own chatbot” (please, no more of those). The opportunity is to be structured, queryable, and genuinely useful inside the workflows guests are already running. That means clean, accurate, machine-readable information about your property, your local recommendations, your accessibility setup, your dietary capabilities, your unique experiences. It means writing your concierge content not for a human scanning a PDF, but for a model summarising it for a guest who’s already three layers deep into planning.

It also means picking integrations carefully. I wrote earlier this year about testing every ChatGPT travel app so hoteliers don’t have to, and the lesson stands: most “integrations” added effort without adding distinctiveness. If you’re plugging into someone else’s surface, the question isn’t “can we be there?” It’s “are we showing up with something a guest can’t get from the property next door?” Loyalty and personalisation are the two areas where most groups have the data to actually differentiate, and the two areas where almost nobody is using it well yet. That gap is the opportunity.

When the AI assistant asks “what’s the kid-friendly restaurant scene like near this hotel?”, the answer is being constructed somewhere. Make sure it’s being constructed from material you actually wrote, not from a 2019 TripAdvisor review.

Travel stage: reduce friction or get out of the way

This is the stage where most “AI hotel” experiments add friction rather than removing it.

I love a good AI concierge in principle. In practice, the bar most of them clear is “marginally better than the printed in-room directory” which is a low bar that nobody will return for. The deployments I’ve genuinely been impressed by have something in common: they reduce a specific, painful step in the guest journey. Late check-in friction, language barriers at the front desk, dietary requests that used to get lost in translation between the booking notes and the kitchen, real-time room temperature and lighting requests that used to require a phone call. Small, sharp, frictionless wins.

The deployments I’ve been unimpressed by have something else in common: they’re a chatbot bolted onto a website that wasn’t great to begin with, and they’re optimised for the marketing team’s “we have AI now” announcement rather than for a guest’s actual day.

A test I’d offer: if your AI deployment doesn’t pass the “would I, on Day 4 of my own trip, voluntarily open this instead of calling the front desk” test, it’s not ready. (And I’d argue most aren’t yet.)

The bigger shift coming is agentic AI, where a guest’s assistant doesn’t just chat with your concierge but actually books your spa, orders the in-room service, modifies the dining reservation, and arranges the car, on their own. The lab-side capability is months away, not years. Google quietly announced Universal Cart at I/O 2026 just a few days ago, and the others are right behind.

The version that actually plugs into a hotel company, though, is a longer story. Tapping into this properly will require overhauling legacy tech stacks, cleaning up the underlying data, and building the kind of internal processes and capabilities that have been the industry’s blind spot for two decades. The labs are sprinting. Most of us are still tying our shoelaces. That gap matters, because while it’s open, your real customer at the moment of decision will increasingly be an algorithm optimising for a utility function: price, rating, dietary fit, accessibility match, schedule constraints. Not your story. The agent.

It’s worth sitting with that.

Reflect stage: the loyalty conversation is changing

Here’s the part the industry has barely started talking about properly.

When guests run a Reflect stage like the one I described, their AI is quietly building a richer, more honest model of what they actually liked about your hotel than any post-stay survey you’ll ever run. “The room was lovely but the air-con clicked all night” is the kind of feedback that lives in their personal trip notes, not in your 1–10 star rating. Next time their Project assistant is helping them pick a hotel in a new city, that note is in context. Yours isn’t.

The question for hoteliers: how do you become the kind of property that earns a place in someone’s personal post-trip memory bank, not just a star rating in a global database? The honest answer is the same answer that’s always worked. Be specific, be human, give people money-can’t-buy moments, and don’t pretend service is loyalty. (It isn’t, by the way. Service is the price of entry. Loyalty beyond reason is rare and almost always emotional.)

The mechanism is changing. The substrate underneath, what makes a guest actually love a place enough to remember it three trips later, hasn’t.

(I wrote a longer piece earlier this year on what I learned testing every ChatGPT-integrated travel app, and the takeaway was similar: most of the integrations didn’t add enough value to clear the “would I open this twice” bar. The ones that did, did one specific thing exceptionally well. If you’re rolling out AI to your guests, pick that one thing.)

Where this lands

Travel planning with AI isn’t a prompt. It’s a workflow. And the workflow has been quietly maturing while most of the conversation has been stuck on “look how scary this is” and “look how magical this is.”

The genuinely useful version is more boring than either. You dream with the model. You plan with the model interviewing you. You travel with the model in your pocket. You reflect with the model so the next trip is better. The model isn’t the protagonist. You are.

For hoteliers, the same workflow is a quiet rearrangement of how guests discover you, evaluate you, experience you, and remember you. Some of the rearrangement is uncomfortable. Some of it is a real opening for properties willing to be specific, distinctive, and human. The biggest mistake would be to assume guests will keep finding you the way they used to, while you take another two years to figure out your “AI strategy.” They won’t. The water is getting colder. The bridge is in line of sight.

The good news, in the Stoic optimist’s reading: this is one of those decade-long shifts where the winners aren’t the ones with the biggest tech budget. They’re the ones with the clearest sense of what their guests actually want, the discipline to build that into AI-readable form, and the humility to keep iterating.

Seven or eight months out from a trip, I open a chatbot and dream out loud. By the time I land, I’ve co-built the whole trip with an AI that knows me, my family, and what I’m trying to feel by the end of the week.

For travelers, that’s the new normal already.

For hotels, it’s an invitation. The kind that rewards whoever opens it first.

The workflow is here. The geek in me remains ever hopeful we use it well.

If you want a clean copy-paste version of the destination prompt above, scroll back up, change the placeholders in [brackets] to your own situation, paste it into your favourite model with extended thinking turned on, and see what happens. Be honest in the answers. Especially the “real reason we’re going” line. That’s the one that decides whether the trip is good.