Nobody has ever felt a calorie. You feel a day.

Because none of them use a unit anybody can feel. A calorie is a unit of heat. A step is a unit of walking. Neither one tells you what it cost you, so the number goes up and nothing changes.

Life expectancy research already prices these behaviours — in minutes and hours. That unit was sitting there unused. Not Today is what happens when you make it the interface.

Steps, sleep and workouts come straight from Apple Health, so most days are already counted before the app is opened. Scoring runs entirely on device, and entries sync to your own iCloud so the app survives a new phone. The exception is a shared challenge: name, progress and weight go to a server, because the other person has to be able to see them.

Screen time is deliberately not taken from the iOS API — no permission, no privacy question, and the app is not owed that access to do its job.

Not one mockup. The brief was text: I described what should happen on a screen and gave no references at all. For the first months I did not touch the look of it — it stayed default while the functionality got built.

The rule was the point of the experiment, not a shortcut. If I had started drawing, I would have been answering a different question.

When there were more tasks than fit in my head, I connected Notion to the model. Cards get written during the discussion itself, the screen flow updates on its own, and a separate page holds a shot of every screen — shared memory, so the product never has to be explained from scratch again.

Each card carries a Start checkbox. I go through the backlog and tick the ones that do not need me. The model looks at Notion a couple of times an hour, picks those up and works through them — overnight, while I sleep. I read the results in the morning.

What gets done is my call. When and how is the model’s. Alongside it sits a set of agents with roles — product, research, UX — that argue an idea before anything gets written.

The raw API model counts calories badly — confidently and wrongly, while the chat product on the same family does it well. There was no clever fix: I worked through models until one counted acceptably, then kept it.

Once I turned to design the app got much better to look at. The UX barely moved.

My guess is that there are so many habit trackers in the world that the models have more than enough references to arrive at a sensible flow without me. The taste was worth adding. The structure was already there.

A week or more went to Apple certificates and provisioning: done in App Store Connect, not showing up in Xcode, round and round. That is the one part of the project where the model helped with nothing — the problem was never in the code.

The rest of the tail came with it: the subscription tier and its paywall, the store listing, the age rating, the privacy answers.

And I did not learn to. That was part of the deal: the moment I go into the code myself, the experiment is over and the answer is worthless.

I can find a corner radius or a type size. Everything else is brief, review, and deciding what is good enough to keep.