A week of vegan dinners, planned by agents.
Four AI agents that turn "what should we eat?" into a week of dinners, a shopping list and a cookbook of favourites.
- Role
- Product, design, build, test
- Built with
- Claude agents
The problem
"What should we eat?" My partner and I could never agree.
And deciding was only the start. I wanted the shopping list written for us, with a rough price for everything at each store. I wanted the stores near us close at hand. I wanted a way to rate each dinner, and a cookbook of the ones we loved.
It began as a course exercise. In my Product Management with AI course we were asked to build one agent. I had built agents before, so I set myself a harder brief: four agents that work together on a problem from my own life.
What was in the way
- A plan is only useful if you can trust it. Prices change and shops differ, so Vegan Chef had to be honest about what it knows and what it estimates.
- Two people, one plan. Both of us choose, rate and shop, so nothing could get out of step.
- Agents need boundaries. Each one has a single job and stays out of the others' work.
What I built
What should we eat?
Every night, the same question: what should we eat?
Every night, the same question: what should we eat?
Sample data only
I wrote a product brief first: the problem, the users, the job to be done, goals with a way to measure each one, what is in and out of scope, and acceptance criteria.
Then I built four AI agents, each with one job, that hand the week from one to the next.
- 01
Find dinners
- 02
Pick together
- 03
Plan and shop
- 04
Rate and keep
Here is what a week looks like from our side:
- New dinner ideas arrive without anyone searching.
- We each pick what we fancy.
- A plan for the week arrives with one shopping list. It shows the approximate price of each item across our stores, so we can see where the week is cheapest.
- The stores near us are a tap away.
- After we eat, we rate the dinners. The ones we love go into our own cookbook.
How the agents do all this stays in my kitchen.
What changed
I turned the acceptance criteria into 25 tests and tracked each one in Linear. Testing caught things I would have missed, like a gluten-free week that suggested an ingredient containing wheat. I fixed each one and tested again.
My partner used it too, and liked it. His feedback is shaping the next version.
What I'd do next
- Make the changes from my partner's feedback and start fresh in December.
- Finish the remaining tests, including the ones that need both of us signed in.
Next project →