In search of the holy kale: meal planning with Claude

And now for something completely different. Most of what I write about AI is work: agents running a business, assessments, workflow design. This post is about dinner.

It's here because I think it makes an argument the work posts can't: some of the most interesting AI systems are the ones you build for exactly one household, to fix exactly one recurring annoyance. Ours was the weeknight scramble of "what can we actually make with what's in the kitchen?"

A borrowed idea that didn't survive contact

The idea wasn't mine. I read a blog post about wiring up a Hermes agent to run weekly meal planning, thought it was clever, and tried it. Hermes choked. The context window couldn't hold a pantry inventory, a recipe library, a weekly plan, and a shopping list all at once without losing the thread somewhere in the middle.

Which led to the obvious thought: I already live in Claude all day. Why not build it there?

So I spun up a Claude Project and built the thing with Claude itself over a few sessions. The goal was deliberately narrow: based on what's actually in our pantry and fridge, propose 3–4 weeknight dinners I can get on the table in under 30 minutes.

How it works now

Each week, Claude proposes around 15 dinner options and I pick 3–4. The proposals arrive as a little tappable picker interface Claude builds on the spot, and the mix is deliberate: a couple of double-batch candidates, one or two freezer meals, one recipe from our to-try queue, the rest fresh. Once I pick, the most perishable ingredients get scheduled earliest in the week, which is the kind of small intelligence that makes the plan actually survive to Thursday.

The options come back already shaped to us, because the Project knows our preferences: the vegetables we lean on, our spice tolerance, the cuisines we gravitate toward, and the fact that I like cooking in a wok and will pick the wok recipe nine times out of ten. Preference files need tending, though. At one point we had to add a written rule that liking gai lan does not mean wanting gai lan in every single Chinese dish, because it had started showing up in most of them.

The structural pieces:

• Notion is the home for everything, connected through the MCP connector: a hub page Claude rewrites each week, plus four databases it reads and updates: the recipe library, the pantry inventory, the weekly plan, and the shopping list. It has become the central place our household's cooking actually lives.
• The shopping list syncs two-way with Apple Reminders, the shared list my wife and I already use. This was the part I expected to be a project in itself, and it turned out Claude supports Reminders natively. When I pick the week's recipes, the missing ingredients get added (pantry staples excluded, so no phantom "buy soy sauce" entries). Claude tags every item it creates so it can tell its own additions from ours, and if we swap a meal mid-week it removes the dropped dish's ingredients without ever touching anything my wife added herself.
• The pantry data came from the least glamorous data-collection exercise of my life: 20 genuinely ridiculous minutes walking around the kitchen transcribing everything out loud, down to every jar of condiments and bag of rice. I mention this because every useful system I've built had a moment like this. The magic runs on somebody doing the tedious inventory once.

A while back I also had Fable 5 take a maintenance pass over the whole thing: tightening the workflow, rewriting the instructions, and cleaning out the Notion databases. It deleted about half the recipe properties we'd been dutifully filling in, on the grounds that it could infer things like cooking method and spice level from the recipe text itself. It was right. The system got noticeably sharper, which is its own small lesson: these setups are worth revisiting as the models improve.

The test that matters

There's only one evaluation metric for a meal planner. The first recipe out of the system was dry fried green beans with beef, in the wok. Cooked it that night. Delish.

Why bother writing about dinner

This use case is not shaping Nvidia's stock price.

There's a version of AI discourse where everything has to be transformation-scale to count. But the skills this little system taught me, structuring household knowledge so a model can act on it, wiring connectors between tools, knowing when a context window will choke and what to do about it, are exactly the skills that transfer to the systems that do count. Building for yourself is the lowest-stakes rep you'll ever get, and the reward is that a real problem in your actual life disappears.

If you want the full setup, the complete write-up lives at Kitchen HQ, and the source files (the system doc, the preferences file, the whole thing, with private IDs removed) are on GitHub, so you can adapt it to your own kitchen.