Personal Cook Bot — What Should We Cook Next?

Some of the most interesting product problems are the ones that sound almost too ordinary to be worth solving.

"What should we cook today?"

It is a question that happens in households constantly. It is also surprisingly difficult to answer well.

The obvious solution is a recipe app. Search for a dish, browse some recommendations, pick something, and start cooking. But I don't think recipes are really the problem.

The problem is the decision that happens before the recipe. What should we cook today given everything else that is happening in the household?

Maybe there are vegetables that need to be used. Maybe yesterday's dinner was heavy, so today should be lighter. Maybe someone doesn't like a particular ingredient. Maybe the cook only has thirty minutes. Maybe there are guests coming. Maybe we've eaten the same thing three times this week and nobody wants to eat it again.

The answer isn't sitting inside a recipe database. It sits inside context.

Rather than creating another app where someone searches for recipes, I built an assistant that understands the household and tells the cook what to prepare next.

There's no app to open. Every morning at 5 AM, a message arrives in our family Telegram group — the place where household coordination already happens. No new tool, no new habit.

Instead of returning twenty recipes, the assistant says something like: "Make vegetable pulao and raita today. You have rice and vegetables available, you haven't had it recently, and it can be prepared within the time you have."

That recommendation is useful because the system knows something about the people asking. It knows what is available. It remembers what was cooked recently. It understands preferences. It knows what the cook is comfortable preparing.

The whole thing runs on three pieces, deliberately kept boring:

Google Sheets is the household memory — pantry inventory, what was cooked on which day, per-person likes and dislikes, and dishes the cook is confident making. A spreadsheet because anyone in the family can edit it without being taught anything.

n8n is the orchestrator. A 5 AM cron trigger pulls the current state from the sheet, assembles it into context, calls the model, and formats the result into a short, readable message.

Telegram is the interface. The recommendation is pushed into the group the household already uses, and replies in that same thread become feedback that writes back to the sheet.

None of these are impressive technologies. That's the point — the interesting engineering is in what goes into the prompt, not what runs it.

The system observes household context, makes a recommendation, gets feedback, and uses that feedback to improve the next recommendation.

If the cook says, "Everyone liked this, make it again next week," that's useful information. If they say, "Don't suggest this anymore," that's even more useful. If they say, "We have too much spinach. Use it tomorrow," the system now has another piece of context to incorporate.

Over time, the product becomes less like a recipe database and more like a household memory.

A generic AI model already knows thousands of recipes. That isn't the difficult part. The difficult part is understanding the messy, constantly changing context of a real household.

The difference isn't the intelligence of the model. It's the quality of the context surrounding the model.

It would be easy to add a chatbot to a meal-planning app and call it an AI cooking assistant. But the chatbot itself isn't the interesting part. The interesting part is removing a recurring decision from someone's day.

This project forced a question I keep coming back to in my day job: where does AI actually create a better product experience? Usually not by generating more options — by removing the need to choose between them.