Visiting a wine region takes the kind of preparation nobody enjoys: who receives visitors, by appointment or walk-in, what time the door shuts, and how far it is to the next estate. Twenty of those little files costs an evening, and it is exactly the sort of work you delegate.
I delegated it and then spent an hour improving it, and that hour is what this piece is about.
The job
One block per estate: address, appointment or walk-in, opening hours, and the source those hours came from. That last field is the difference between a usable dossier and a list that merely looks good.
My actual prompt
I'm going to [region] for four days. Build a dossier of twenty estates
that receive visitors.
Per estate: name, village, appointment or walk-in, the opening hours, and
the URL you took those hours from.
No source means the estate goes at the bottom under "unverified", with the
reason. Do not invent hours and do not invent phone numbers.
For each estate, give the date shown on the source page, so I can see how
old the information is.
Group by driving distance, not by fame.The line about the source date is there because of what happens without it, which is the interesting part.
Taken apart
Ask for the source per fact, not per answer. A list of sources at the bottom is decoration. A URL next to the opening hours is checkable, and you only notice the difference when you open it.
Ask for the source’s date. This is the field I had forgotten myself. Tourist pages are often years old and look exactly like a page from last week.
Make the unverified stand out. Anything the model could not substantiate belongs at the bottom in its own list, with a reason. Hidden among the rest, it gets treated exactly like the checked rows.
Group by distance. A dossier ordered by fame is a list you cannot drive.
What broke
This is not a fault in the model, which makes it more instructive.
I used the Jura as a test. What came back on opening hours matched the sources, and the sources themselves are a mess. One estate’s hours are published by a tourist board with a validity range running to 31 December 2050. Another runs to 2049. Those are not opening hours, they are fields somebody filled in to get past a form.
The same information also sits in three languages on three subdomains of one tourist board, with different hours per language, because one version was updated later than the others. And a third estate says “visit and tasting, by appointment only”, right next to a line saying open 10 to 17, which is the kind of pair a summary flattens into the wrong half.
So the model summarised correctly. The problem is that correctly summarising messy sources produces a tidy dossier in which you can no longer see that it was messy. Hence the source URL and the date: not because the model lies, but because the source material does.
The fix is not technical. For anything I am actually going to drive to, I phone. Five calls for four days, and the dossier’s job is to tell me who to call.
Ethics note
Small estates are not businesses with a front desk. They are often two people who also work the vines, and a dossier that happens to send twenty visitors at once means something different to them than to you. Call, arrive at the agreed time, and buy something.
Watch what you take from them too. The text on an estate’s site is usually written by the maker and it is their story. A dossier for yourself is fine; republishing that story as your own travel writing is not.
And the energy side, which belongs here once you have a model read twenty pages: this is one of the few tasks where that is defensible. Summarising twenty pages costs less than the evening you would otherwise spend with twenty tabs open, and far less than the detour you drive when you find out halfway that the estate is closed.
Take it
I'm going to [place] for [duration]. Build a dossier of [number] [what you
are looking for].
Per item: name, location, the practical details I need, and the URL you
took them from.
No source means the bottom of the list under "unverified", with the reason.
Do not invent hours or phone numbers.
For each item, give the date shown on the source page.
Group by driving distance, not by fame.A small exercise: open the source for the first item and check whether the date matches what is in your dossier. If the source is from 2019, you know how to read the rest.