It was a good day for the grill. Spring, warm, the kind of afternoon that calls for something light and seasonal rather than heavy. We wanted to stay outside and keep it simple.
Working through the menu with Claude, we landed on grilled swordfish, with tuna as the backup depending on what the store had. Either way, the plan was a caper and tomato relish built with red wine vinegar. Briny, acidic, a little bright. That preparation would work with either fish.
With the dish taking shape and two possible fish still in play, we could turn to the wine. We have a large cellar, and something in there would work. But I needed to know what to pull and get chilling before I left for the store. The relish would shape the pairing as much as the fish itself, and I wanted to be ready before I knew which one was coming home.
I started with Claude's recommendations. The relish was the deciding factor. Capers, red wine vinegar, a preparation that needed a wine with enough salinity to hold its own:
↳ Claude:
The relish is the deciding factor here. The brine from the capers and the acidity from the red wine vinegar call for a white with enough salinity to hold its own. Vermentino from Sardinia is the natural fit. A white Burgundy works if you want more richness. A dry Provencal rose covers either fish.
None of those matched what I thought I had. We track our cellar in an app called VinoCell, and I was fairly sure we didn't have any of those. I shared a screenshot of the Sauvignon Blancs I had on hand, figuring something there would work well enough.
Claude offered options from what I'd shown. But I was giving it a limited view into what was actually possible. I was doing the filtering, not the pairing. It occurred to me that if I handed over the full inventory, I could let the recommendation follow from the actual data rather than from my assumptions about what might work.
The app has an export function. I pulled a full inventory file, dropped it into the conversation, and asked Claude to look again.
↳ The Aided Chef:
Here's my full cellar inventory as a data file. Does anything in here change your recommendation?
↳ Claude:
Yes. You already have the Castellani Vermentino 2024. That's the bottle.
We had it. Not something I remembered buying, so not something I had thought to search for. The recommendation hadn't changed. But now it was pointing at something already in the cellar.
I got it chilling and headed to the store.
They had swordfish. Beautiful steaks.
The relish came together while the grill heated. Cherry tomatoes halved, capers drained, parsley rough chopped. The red onion went into the red wine vinegar first. My husband doesn't love raw onion, and a quick pickle takes the bite out without losing what the onion contributes. By the time everything else was ready, it had done its job.
The swordfish went on a hot, clean grill. The afternoon got away from us. It was beautiful out, the kind of day that makes you linger, and we let it go a little long. The fish came off slightly overcooked, firm where it should have been just tender. Still good. Swordfish has enough density to hold up through some error, and the relish was lively enough that the dish worked even with the fish not quite where we wanted it.
We rounded out the meal with grilled asparagus that cooked alongside the fish and some crusty bread to catch the juices from the relish.
The data found the bottle. The dish proved it.
The quality of a recommendation depends on the quality of the input.
Claude had workable options from the screenshot I shared. I would have pulled a Sauvignon Blanc and called it close enough. Not a bad choice. Just a choice made from a limited view of what was possible. Handing over the full inventory changed what came back. The search was no longer running through my assumptions. The data did that work instead. There's a difference between asking for a recommendation and giving the actual material to work from. The first produces something reasonable. The second can surface something specific.
A good pairing doesn't sit beside a dish. It works inside it.
At the table, the Vermentino earned its place. My husband noticed it first. The fish tasted more savory, more meaty, less like fish. The relish came alive alongside it. The parsley in particular seemed to amplify rather than just sit there. It wasn't a wine that stood politely to the side. It pulled the whole plate together.
If I cook this again, I'll pull the fish a minute earlier than feels right and trust the carry-over. That's the one thing I'd change.
The Vermentino was in the cellar the whole time. I wouldn't have found it without the data. I didn't remember buying it, and I hadn't thought to search for it. Handing over the complete inventory instead of my own filtered version of it changed what came back. Most of us develop instincts about what we have and what might work. Those instincts shape what we ask for.
Have you ever handed AI more than you thought it needed and gotten back something you wouldn't have found on your own? I'm curious what that opened up.
The full recipe from this entry is available for subscribers. If you cook this, I'd love to know what you changed and what prompts you used to get there.
