Find every booking confirmation in my mail for travel between [date] and [date], and put them in one note with the date, the supplier and the reference number. List anything where you could not find a reference.
9 prompts · 5 tests · assembly beats invention
Generating a plan is free. Checking it is the work.
Ask any model to plan a trip and you get a beautiful itinerary in four seconds. It is nearly worthless, because every fact in it has an expiry date and a real-world consequence the restaurant that closed, the museum shut on Tuesdays, the train that no longer runs. Travel planning is a verification problem wearing a generation problem’s clothes.
If you only read one bit
Use it to assemble, not to imagine
The short version
Its real travel skill is collecting what you already own
An agent cannot check the world. It can only act on what it can read which, for a trip, means your mail, your calendar and your notes. That turns out to be where nearly all the tedium lives: eleven confirmation emails scattered across six weeks, reference numbers you will need at a desk, and the one leg of the journey nobody actually booked. Pulling that into one place is tedious rather than difficult, which is precisely the rule for agent work.
Assemble confirmations
Find the unbooked gap
Treat suggestions as drafts
Never let it book
Needs the Mail, Calendar and Notes areas granted depending on the prompt each is listed on the cards below. How grants work.
Three jobs, three owners
Suggest, assemble, book
A trip breaks into three kinds of work, and they belong to three different places. Conflating them is why AI travel planning has a reputation for being impressive in demos and useless in practice.
Suggesting what is worth doing, what suits four days, what to skip is a conversation. It benefits from back-and-forth, from you saying “not that, we hate queues”. That is Meta AI, and it is available in most of the world today.
Assembling gathering what is already decided into something usable is an agent job. It is dull, it spans several apps, and it is checkable at a glance. This is where Muse earns its place, and the whole prompt list below sits here.
Booking is yours. It spends money, it creates obligations to real businesses, and it is the archetypal action with no undo. Meta says it confirms before purchases; treat that as a backstop, not as permission to look away.
The demo everyone remembers is an agent booking a trip. The version that actually helps is an agent telling you that day three has no hotel.
There is a fourth thing an agent genuinely cannot do, and it is worth naming: verify the current state of the world. Opening hours change, places close, services are suspended. Anything it tells you about a venue is a starting point for a check, never the end of one which is what the tests below are for.
Nine prompts
Before, during and after
All assembly, all scoped, all asking for something you can check in ten seconds. Square brackets are yours to replace.
Before you go · tasks 1–3
The gap-finding stage, which is the one that saves an actual disaster rather than an hour.
Using that note, list every night between [date] and [date] that has no accommodation booked, and every travel day with no transport booked. Do not book anything.
Create calendar events for each confirmed booking in that note, using the supplier name and reference as the title, and show me the full list before you save anything.
While you are away · tasks 4–6
Short, narrow requests. You are on a phone in a station, not auditing anything.
Find the confirmation for [supplier] in my mail and give me just the reference number, the address and the check-in time. Nothing else.
List everything in my calendar for tomorrow with a time and a location, earliest first. Do not change anything.
Search my mail for any message from [supplier] received in the last 48 hours and tell me if anything changed about my booking. Do not reply to it.
After you get back · tasks 7–9
The strongest use on the whole page, and the one people never think of until the expense deadline.
Find every receipt and payment confirmation in my mail dated between [date] and [date], save the attachments to a folder called [trip], and list each one with the supplier, date and amount.
Go through that list and flag any where the amount is missing or the supplier is unclear, so I know which ones to chase. Do not contact anyone.
Compare the receipts in that folder against the bookings in my trip note and list anything I booked but have no receipt for.
Note that none of the nine asks it to suggest, recommend or evaluate anywhere. Every one is assembly against material you already have, which is why every one is checkable. The four parts of a task.
Practical tests
Five checks to run before you trust it
Each takes under a minute and tells you something a demo never will. Run them on any AI travel output this one, a chatbot’s, or a travel app’s.
The fictional-place test
Invent somewhere. Ask about a café, a museum or a station that does not exist, with a plausible-sounding name, and see what comes back.
- Pass
- It says it cannot find that place, or asks whether you have the name right.
- Fail
- It describes the atmosphere, the speciality and the best time to visit. Everything else it told you now needs checking.
The attribution test
For any fact in an assembled list a time, a reference, an address ask which message it came from.
- Pass
- It names the source and the two agree when you open it.
- Fail
- It restates the fact more confidently, or attributes it somewhere that does not contain it.
The contradiction test
Feed it two things that cannot both be true a hotel booking and a train that leaves before check-out, say and see whether it notices.
- Pass
- It flags the clash and asks which is right, which is the hand-back behaving correctly.
- Fail
- It silently picks one and builds the rest of the plan on it. This is the failure that ruins a morning.
The staleness test
Ask when its information about a venue is from, and whether it could have changed.
- Pass
- It says plainly that it cannot confirm current hours and that you should check with the venue.
- Fail
- It repeats the opening hours as though they were checked this morning. They were not; nothing it has access to could have.
The read-versus-assumed test
Ask it to separate what it actually found in your mail from what it filled in around the gaps.
- Pass
- You get two lists, and the second is shorter than you feared but not empty.
- Fail
- It insists everything came from your mail. Something in a complete-looking itinerary was always inferred.
The one test it cannot pass
None of this lets an agent tell you whether a place is open next Tuesday. Opening hours, closures, strikes and seasonal changes are current facts about the physical world, and the agent’s reach is your Mac. Treat every venue detail as a prompt to check rather than an answer and check the ones where being wrong means standing outside a locked door in another country.
This is not a Muse failing specifically. It applies to every AI travel tool that has ever produced a confident itinerary.
If an output fails tests one or three, do not patch it discard it and go back to assembly. A plan that invents places or hides contradictions is not improved by editing. Why it hands back.
FAQ
Travel planning questions
Can Muse plan a whole trip for me?
It can produce something that looks like a plan in seconds, which is the problem rather than the feature. Generating an itinerary is easy and nearly worthless; the work is checking that every place still exists, is open, and is reachable. Use it to assemble.
What is it actually good at for travel?
Assembly pulling confirmations out of your mail into one place, building a calendar from bookings you already made, cross-checking what is booked against what is not. Tedious rather than difficult, which is the rule for agent work. The nine prompts.
Should I let it book anything?
No. Booking spends money and creates commitments in the real world, and it is the archetypal action with no undo. Meta says it confirms before purchases, and that is a backstop rather than a reason to delegate the decision.
Why does it invent places that do not exist?
Because a plausible-sounding restaurant is easier to produce than a real one, and nothing in the request forced it to check. That is why the fictional-place test is on this page ask about somewhere you made up and see whether it admits it cannot find it. The tests.
How do I test whether an AI travel plan is any good?
Five quick checks: can it attribute each fact to a source, does it date its information, does it flag a contradiction between two bookings, does it admit when something does not exist, and does it separate what it read from what it assumed.
Is Muse better than Meta AI for this?
For different halves. Meta AI is a conversation, which suits deciding what you want to do. Muse acts, which suits assembling logistics once decisions are made. Most trips need both. The comparison.
Can it check whether a place is open?
Treat any opening hours it gives you as a starting point rather than an answer. Hours, closures and seasonal changes are exactly the facts that go stale, and the consequence of being wrong is standing outside a locked door in another country.
What should I ask it after the trip?
Expense assembly is the strongest post-trip use pulling receipts out of mail into one folder or list with amounts and dates. Pure extraction, checkable at a glance, genuinely tedious by hand. Tasks seven to nine.
Does this work outside the United States?
Muse launched in the United States only, which is a particular irony for a travel page. Meta AI has a far broader rollout and covers the conversational half. Where Muse works.
What is the single most useful travel prompt?
The first two working as a pair pull every booking for a set of dates into one note, then list the nights and travel days with nothing booked. The gap it finds is the thing worth knowing before you leave. See them.
Assembly beats invention
The useful version tells you day three has no hotel.
Not the one that invents a charming café. Point it at what you already own, then run the five tests on whatever comes back.
Independent guide not affiliated with Meta. Prompts and tests here are written by us as examples; statements about confirmation prompts are Meta’s own. Nothing on this page should be relied on for travel that matters without checking with the supplier.