AI / GenAI·7 min·21 October 2026

A tasting night that writes itself up

At a tasting the best remarks land on the second and third bottle, and by the sixth they have evaporated. Whoever takes notes isn’t tasting. Whoever is tasting isn’t taking notes. This has been true so long that nobody treats it as a problem any more.

A phone recording solves it, and brings along the only real issue in this piece: you are recording other people.

What you want

One recording of the evening, and by the next morning a document with, per bottle, what was said about it, who thought what, and which bottle won the conversation. Not a summary of the evening, which is useless. A report per bottle.

The prompt

This is the recording of a tasting with six wines.

Split the transcript into six parts, one per bottle. I say each bottle's
number out loud before we start it, use that as the marker.

Per bottle: write down the tasting impressions as they were actually said.
Use the speakers' own words and add no descriptors of your own.
Record opinions that contradict each other as two separate lines, not as
an average.

Refer to speakers by a letter, no names.

End with the bottles ordered by how much was said about them, not by how
positive it was.

That last line is the most useful of the set. The bottle that gets talked about longest is not the best one, it is the most interesting one.

Taken apart

Mark with your voice, not afterwards. Saying “bottle four” out loud takes half a second and saves half an hour of hunting through a transcript. It makes every recording searchable without any editing.

Forbid averaging. If four people think one thing and two the opposite, the average says nothing. The disagreement is the information, because that is where the gap sits between what the wine does and what people expected.

Letters instead of names. Anonymising up front is easier than redacting afterwards, and it makes the document shareable without any fuss.

Where it goes wrong

Three things are predictable once there are more than three people at the table.

Talking over each other. Transcription separates speakers reasonably well as long as they take turns. At a tasting table they don’t, and the bottle everyone has an opinion about at once is the one that turns into mush. That is usually the interesting bottle.

The marker that drowns. Saying “bottle four” out loud only works if you wait for a gap. Say it while people are still talking and it never comes back as a marker, which shifts the entire report by one bottle.

Words belonging to nobody. Without the rule forbidding added descriptors, the model rounds short reactions up into tidy tasting notes. They sound like something someone could have said, which is exactly the problem: afterwards you cannot tell who thought what.

Ethics note

On this subject the caveat comes before the technique.

You are recording people who are guests in your house. That requires consent up front, out loud, at the table, before the first bottle opens, and not a remark afterwards that “something was running”. Someone saying no means the recording doesn’t happen, not that you edit them out.

Say what happens to it afterwards too: where the file lives, whether it goes to a service for transcription, and when you delete it. A reasonable term is the day after the report. A recording of an evening with friends is exactly the kind of file you forget exists three years later.

And more gets said at a tasting than wine talk. Something personal comes up, someone talks about work, there is gossip. The report should not be about any of that, and the transcript contains all of it. That is why the file has a short life and the report only covers the bottles.

Take it

This is the recording of [a session with numbered parts].

Split the transcript per part. I say each part's number out loud, use that
as the marker.

Per part, write down what was actually said. Use the speakers' words, add
nothing, and record contradictory opinions as two separate lines rather
than an average.

Refer to speakers by a letter, no names.
End with an ordering by amount of discussion, not by approval.

A small exercise for the next time you are at a table with more than three people: ask out loud whether you may record. The answer is more interesting than the report.