I built an app on parental leave (mostly from my phone)

I’ve wanted to build this app since COVID. Back then I made a version called What’s Next?, a Ruby on Rails web app. I’d taught myself Rails after years as a product designer, with whatever coding I had left. My friends used it, the ones I’d convinced and invited. People were literally bored.

To grow, it needed to be a native app, and that’s where it stopped. It was hard to find tutorials that really helped you wrap a web app in a native container and get it out there. And people, even today, have a hard time installing a web app on their home screen unless they’re on Android.

So it stopped at the App Store door. Keep that in mind for later.

This year David, my best man and my friend from university, told me about his struggles with cooking: going through his recipe books, trying to plan a week. The idea had always been Firuze’s and mine. I’d started building products with AI by then. “Why don’t we build this together?”

In May David was in Berlin, so we did the kickoff at home. Over a Sunday and a Monday we worked with Claude and came out with a folder full of files: research on what people complain about in recipe apps, research on recipe copyright, a spec, a click-through prototype. Firuze, a Captain Cook power user, was first on the list to interview.

Nine days later David was back in Cologne and we had a remote session. I wrote the first prompt for Claude Code at 21:07, and the first commit landed at 21:29. By 23:18 we both had a very, very early draft.

Then we both went back to work. I spent my time on a different product I needed sooner, and the draft sat there.

Our son arrived and I was home for two months. I cooked a lot. The concept was still sitting there.

I had a Mac mini at home that I’d set up as a dedicated machine for that other product. So the plan was simple. Grab my phone, use Remote Control, and every time I had some time left, build it on the phone while I was taking care of the little one.

A lot of it happened on long walks. Outside alone, Nuri asleep in the carrier. I used to listen to podcasts on those walks. Instead of listening to what other people are doing, I was doing my own thing.

I couldn’t have built this without that. Otherwise I’d have spent hours at the computer, which would have taken far more time away from the family, or it would just have been super slow. Months instead of weeks.

On 21 September I picked it up again. From then on there was at least one commit every day for 32 days.

I’m not reading code. I want to say that early, because people assume otherwise.

I describe an idea to Claude, usually by talking, and ask it to sketch the concept. It comes back as a Claude Artifact, and those got so much better than before. They’re not just HTML any more. They’re very interactive, an artboard I can click through on my phone, annotate and comment on.

This was the moment it clicked for me. It reminds me of the design critiques we have as designers, where someone shows artboards and workflows and we all leave comments. Now it’s exactly the same process, with the agent.

When we’re both happy, it writes a development concept. I read that. Then it builds, I open it on my phone over the home network, and I try it in my hands.

What I bring is a general understanding of how applications work. And taste, which only works after the fact. You can only know it when you see it. That’s what the speed is for. I don’t need many mockups to feel the real thing, because I have it in my hands.

Every time something is wrong, it becomes a rule. The agent keeps a memory for this project with 70 notes, and 28 of them are corrections from me. Never say “dinner” or “tonight”. No icons in big buttons. A recipe opens at its own servings. Each one is something I only knew was wrong once I saw it.

Our plan from May was clear: no feed, no community in the first version. Five days after I picked it up again, community shipped anyway, in three parts on one day.

That was 100% speed. And it was Firuze’s mother.

Havva was staying with us during the leave. She cooked us a pizza she’d found on YouTube Shorts, showed me the short and sent it to me. She kept watching it over and over and over again. And I was like, “Man, I need to make this show when she saves it, and I can also see it, and we can like it, and we can share it.”

The short was completely in Turkish. So the next day the app could translate a recipe into German or English, and turn tablespoons and American or British measures into German ones.

Then the bigger thought. Why don’t I invite Havva, Firuze’s aunt, my neighbours and my friends, and they can cook as a household together? That was the whole piece I always wanted.

By October I had a running web app that people had to install on their home screen. The same place What’s Next? had stopped, five years earlier.

So I asked: “Hey can you try to turn this into a native app?”

I remember it as ten minutes until I had a first running iOS version. The session’s record says I asked at 13:41 on 8 October, and at 13:53 I was already asking it to try a native tab bar. Twelve minutes. Xcode and the simulator were on the Mac mini, which is another reason that machine matters.

By the evening I told it to “do all things necessary for App Store submission”. The public TestFlight went to Apple for review, and then the App Store review too, because why not. At 21:30 I installed it from TestFlight and found the first bug myself.

To be fair about what it is: a native shell around the live website. The tab bar, sheets, menus, the widget, Live Activities and notifications are iOS. The screens inside are the website. That’s why it took hours instead of months. It’s still waiting in App Store review.

WhatWeCookPlan the week togetherIn review

Henning Witzel-Acikgoz·Closed beta on TestFlight

When I was first asked what the agents are bad at, I said they’re not really bad at anything. That isn’t true.

On 14 October two TestFlight builds went out without the file that tells the app where the server is. They opened on “WhatWeCook needs a connection”. The agent had built and uploaded them from a fresh worktree without checking that the file was there. It was only me, and I caught it right away because I install every build.

Two days later it told me, from its own notes, that a feature wasn’t built yet. It had shipped two days before. In the same conversation it said opening a recipe fires about 15 AI calls. The real number is at most two.

It gets worse as a session gets long. Compacting helps. Sometimes you simply have to restart, and that needs me at the Mac mini. I can’t do it from the phone. So sometimes the trust is really low, and that’s why I do a lot of manual testing.

And nobody reviewed the code. For the sake of speed, I didn’t review pull requests. There’s no CI either. The only reviews were agents reviewing agents after the fact, and on 22 October they found real things: no limit on how many AI calls one person could trigger, and outside fetches that could be pointed at internal addresses. Fixed the same day. I could put a review step into every pull request, or have a GPT model review the code. That’s something I will do. I haven’t yet.

One more thing I didn’t expect. Working with Claude feels like a slot machine. Not adrenaline. Positive emotions: you made something, and you want to keep going. 50 of the commits landed between midnight and five in the morning. You need to understand where to put it down, because you can do way more than you normally could.

On 14 October my agents wrote David and me a status. The part about what’s built was long. The part about what people do was short.

Outside the three of us there were 11 people. Six had never opened the app. Nobody had planned a meal and then cooked it. Nobody had used the shopping list. The advice was blunt: stop adding features until outside households use the loop.

The same day, the welcome was rebuilt to end in a proposed first week. Ten days later there are 23 people outside the three of us. Seven have saved recipes of their own. Two have cooked a meal they planned. Three households have two of them in it.

Small numbers, going the right way. The agents made building fast. They didn’t make anyone use it.

If you’re in a company that’s being asked to use AI, or you’re a founder, or a designer thinking of building something yourself: this is how I do it.

You don’t need to read code. You need to understand how applications are put together, well enough to ask for the right thing and test it in your hands. Write a rule every time something’s wrong. If you skip the review the way I did, that risk is yours.

Someone asked me what part of this a design leader wouldn’t like. I don’t know. To me it’s literally amazing to have an engine at hand and not have to be at your computer to use it. Every time you have an idea, you can dictate it into the app.

If you have questions, ask me.

The timer I asked for at 10:21 took six minutes. Getting people to cook with the app is taking longer.