The lemon that bought a Schweppes (and other AI dinners)

The shopping list said "lemon". The cart ended up with a Schweppes. Technically it isn't wrong: the word lemon is right there on the label. It's the kind of mistake that gets a laugh at the dinner table and a frown in production, because it shows exactly how a machine reasons when nobody has explained to it what a lemon is.

The project is called Settavola (roughly "seven at the table"): seven days, seven dinners, everyone at the table. It does one thing, and it does it for one family, mine. The AI proposes the week's menu, drawing on the recipes of our kitchen robot; we approve it, or pick between the alternatives, or send it back with a comment. Once approved, the menu becomes a shopping list and the list becomes a ready cart at the online supermarket. Then a notification arrives, and we're the ones who pay. The system never pays: it's written in the privacy impact assessment, and the project doesn't contain a single line of code that knows how to pay.

First commit on 9 August, in production since 26 August, 236 commits in just over a month. The first real menu is from 27 August. AI cost for that week: about 16 US cents. The first real cart, in early September: 33 products out of 37. Put like that it sounds like a success story. The good part, though, is in the four that are missing, and in the ones that were there but shouldn't have been.

The lemon, the sugar and the pears

The heart of the system is a translator: from "sugar, 200 g" to a real product on a real shelf. Sounds easy, until you watch it work.

• "Lemon" → a lemon Schweppes.
• "Sugar" → it suggested a Pepsi, which does contain plenty of it.
• "Pears" (pere) → risotto rice, hot dog buns, pizza mozzarella. In Italian the match had fired on the preposition per ("for"): riso per risotti, pane per hot dog, mozzarella per pizza.
• Spring onions: 700 grams bought for 60 grams needed.

The fix wasn't "a smarter prompt". It was a two-step chain: first a broad search that brings home every candidate; then a small, cheap model that asks one question only, like a grandmother at the market: is this really the ingredient, or does it just look like it?

The constraint that lived only in the prompt

The most instructive mistake isn't about shopping. It's about rules.

The menu had to respect some constraints: dishes ready in 30 minutes at most, dinners that are actually dinners, variety between meat and fish. I had written them in the instructions to the AI, carefully. The AI read them, understood them and, every now and then, ignored them. One evening the proposal was a 90-minute dish. Another time Monday's dinner was a milkshake. Swordfish wasn't recognised as fish, so a week with five fish dinners out of seven took no penalty at all.

The project's issue log has a sentence in bold: a constraint written only in the prompt is not a constraint. And right after it, in capitals, the way you write when you're cross with yourself: FOUR times the same bug. Protein, cooking time, banned ingredients, variety: one mistake, four disguises. The solution was to move the rules from the text into the code: the prompt asks, the program checks. If a recipe takes more than thirty minutes, it isn't the AI discarding it out of good manners: it's a check that has no opinions.

The test that lied

The quietest failure is my favourite, because it isn't funny at all. For nine days the system said the session had expired and that I had to log in again. The session was fine. The check was broken: it tested for a condition that only existed in the test, and that never showed up that way in reality. The test passed, the code was wrong, and the error message blamed someone else.

The lesson is one to pin above the desk: a test written against an imaginary world only checks that the imaginary world works.

The name, in passing

The project was born as "Assistente Chef" and became Settavola on 28 August. Three agents screened eleven candidates. One, Settimenu, fell because someone else had registered its domain two days earlier. The runner-up, Impiatta ("plate it up"), lost on meaning: plating is the very last gesture, and the project does everything that comes before it. The repository, the packages and the deploy project still carry the old name: renaming them touches addresses and system identifiers, and for now it isn't worth the effort. Naming debt is real debt, it just doesn't make any noise.

What goes in the cart

• The AI proposes, the family approves, the system doesn't pay. Three verbs, three different subjects, and nobody does anyone else's job.
• Finding and judging are two jobs. Whoever does both in one go buys the Schweppes.
• The rules that matter belong in the code. In the prompt they're just good wishes.

The AI's bill for that first week stayed below the price of a coffee. And the lemon is now a lemon.

The yellow kind.