A kitchen tablet that rates dinner

My grocery agent learns from what I change in the Picnic cart. Swap jarred pesto for fresh pesto, and next week it gets it right. (How that works: How my grocery agent works.)

But there was one thing it never learned from: the food itself.

Whether a dinner was a hit or something nobody wanted to see again, that verdict only lived in our heads. Maybe someone mentioned it at the table. A week later, nobody remembered. And the agent happily suggested the same mediocre dish again three weeks later.

So I put a tablet in the kitchen.

What's on the screen

The dashboard runs in Home Assistant, which I already use at home. It's in Dutch, so I've translated the labels below. It has two parts.

This week, big and easy to read while cooking:

• the dinners for this week, each with a photo and the cooking time
• a Preparation button that opens a separate view with a large photo, the ingredients and every step, so nobody has to dig through an email with a pan on the stove
• an Open in Picnic button for the original recipe
• after dinner: a rating from 1 to 5, an optional note, and Eaten, which hides the dish from the list

Next week, compact:

• six candidate dinners
• we pick about four of them with I want this

That second part turned out to matter more than I expected. Choosing four out of six is a lot easier than coming up with four ideas. Nobody has to think of what to eat, but everybody still gets a say.

From stars to rules

When I ask the agent for next week's groceries, the first thing it does is pull everything from Home Assistant: the ratings, the notes and the picks for next week.

Then a few simple rules kick in:

• 5 stars: the recipe becomes a favourite
• 3-4 stars: saved in the history, nothing else
• 1-2 stars: the recipe never comes back

The picks for next week go straight into the menu. And as soon as an order is ready, the agent immediately puts six new candidates on the dashboard, so the shortlist is never empty.

That closes the loop. The agent no longer only learns from what's in the cart, but from what we actually thought of the food.

The screen is dumb, the files are smart

The part I like most about how it's built: Home Assistant doesn't own any of the data.

Everything still lives in the same plain files the agent already uses. Home Assistant is just a screen in the kitchen. Two small scripts move data back and forth:

files ──push──→ Home Assistant   menu, recipes, next week's candidates
files ←─pull─── Home Assistant   ratings, notes, picks

Even the dashboard itself isn't built by hand. A script creates the helpers Home Assistant needs for each dish (the name, the link, the rating, the note) and generates the complete dashboard through Home Assistant's API. If I want to change the layout, I change the script and run it again.

That keeps the whole thing replaceable. If I ever swap the tablet for something else, the agent doesn't need to know.

The small things took the longest

Generating a dashboard was the easy part. Making it pleasant to use on a kitchen tablet took a lot more fiddling.

Recipe links didn't open. A normal link to a Picnic recipe opens inside the Home Assistant app, and Picnic doesn't allow its pages to be shown there. So the Open in Picnic button uses a link that hands the recipe to the Picnic app itself when it's installed. On Android, a long press forces it to open in the Picnic app.

Sliders are terrible with wet hands. The first version used a slider for the rating. On a tablet, in the middle of clearing the table, that's surprisingly fiddly. Now it's five big buttons, plus a reset.

Long recipes cluttered everything. Putting every step on the main screen made it endless to scroll. Now the overview stays compact, and the full preparation lives on its own page with a clear way back.

"Eaten" needed to mean something. Without it, the list kept showing dishes we'd already had, and it was hard to see what was still planned. Now an eaten dish disappears until the next week is pushed.

None of this is technically impressive. But it's the difference between a dashboard that's used every day and one that's ignored after a week.

What I took away from it

• Put feedback where the moment is. Nobody opens a laptop to rate dinner. A tablet on the kitchen counter, right after eating, works.
• Let the screen be dumb. Keeping all data in plain files makes the dashboard easy to rebuild, and easy to replace.
• Choosing is easier than inventing. A shortlist of six turns "what do you want to eat?" into a two-minute decision.
• Usability beats features. Big buttons and a working recipe link did more for adoption than any extra functionality.

What's next

I want the kitchen tablet to become the main place for feedback: not just ratings, but also quick things like "we're out of milk" or "not home on Wednesday". And I'd like the planning to use more of what Home Assistant already knows about the house.

But for now, the most important thing already works: the food itself has become part of the feedback loop.