Your AI level and your wine level are two separate things

The piece about the wine profile in a Claude project got me two replies that contradicted each other. The first came from someone who has been buying wine for thirty years and found the ten-question interview a waste of time: “I already know all of this about myself.” The second came from a developer who got the idea immediately and then stalled on question eight, the one about oak. He had no idea whether he likes it.

They were both right. They just weren’t standing in the same place.

Two axes, not one ladder

I have been writing about AI and wine together for months, and I caught myself writing for a single reader: someone who knows a bit of both. That reader barely exists. What you can do with AI and what you know about wine are completely independent.

Four boxes, four different starting points. Below is where you start in each one, a prompt that belongs there, and the mistake I see most often there.

Top left: one question, one bottle

You have barely used AI and you buy wine by the label or the price. Don’t start with a profile. Start with tonight.

I want one bottle for tonight. I'm eating [dish], I want to spend at most
[amount], and I buy at [shop] in [country]. Give me one bottle, not a list.
Tell me what temperature to serve it at and why that suits it.
If you are not sure, say so.

Those last two lines do all the work. “Not a list” forces a choice, and without it you get five bottles with an enthusiastic paragraph each, after which you pick on price anyway and learn nothing. “If you are not sure, say so” is there because a model sounds every bit as confident about a bottle you cannot check as about one it actually knows.

The trap here is the list. Five options feel like a better answer than one, and it works the other way round.

Top right: the model as note-taker

You know wine and you have hardly used AI. Then don’t ask it for advice, because this is the corner where the model knows less than you do. Have it write things down instead.

I'm going to dictate tasting notes. Don't make anything sound better than I say it.
Put my words into a fixed shape: colour, nose, palate, finish, verdict.
Ask me to fill in any term I use without explaining it.
Never add a tasting descriptor I did not say myself.

I only added that last line after a week. The model had been padding my short notes with descriptors I never tasted, and they sat there so convincingly that on rereading I took them for my own. That is worse than bad advice, because it contaminates your own memory.

So watch out for being polished. Your note reads better and means less.

Bottom left: wine as a dataset

You build things, but your rack is a shelf of bottles you know nothing about. Wine makes a good practice dataset: small, messy, and you have a real stake in the answer.

Here is my cellar as a CSV: [paste].
Write a script that estimates a drinking window per bottle and sorts by
what should be opened first. Have the script state the assumptions it makes
and fail loudly on an empty column instead of inventing something.

The temptation in this quadrant is to build instead of drink. I once spent an evening on a dashboard for a cellar of fourteen bottles. The flat text file did the same job in twelve lines.

You end up building infrastructure for a problem that fits in a note.

Bottom right: your own data, your own test

You know wine and you build things. This is the only quadrant where you can check the model against something you know the answer to, and that is the most valuable position there is.

I'll give you twenty questions from my own field with an answer key.
Answer them without searching. Give a confidence percentage per answer
and flag where you are guessing. Then compare against the key and count
separately how often you were wrong while your confidence was above eighty percent.

That last number is the only one that matters. A model that is wrong and hesitant is usable. A model that is wrong and sounds certain is dangerous, and that is exactly what the hallucination audit measures.

The risk is that you build your test out of what you already know, so it finds precisely the errors you saw coming. Have someone else write five of the twenty questions.

Ethics note

The quadrant is a way to pick a starting point, not a label to stick on yourself. I see people file themselves under beginner and stay there, when they have long since moved along one axis.

What it leaves out matters more. In all four corners the model does not taste. It recognises patterns in language about wine, and language about wine is not wine. People on the right know that and use the model as a library. People starting on the left risk mistaking the description for the flavour, and that is a habit you have to unlearn later.

And watch what you hand over. Your taste profile is harmless, your budget less so, and the moment you share a cellar file you are also disclosing what is standing in your house.

Start here

Work out your quadrant with two questions. Have you ever written a prompt longer than two sentences? And can you name a wine you poured away this year, and why?

Two noes puts you top left, and the prompt above is enough for tonight. Two yeses puts you bottom right, and the most honest test you can run is your own twenty questions, against the answer key you already carry in your head.