It started, as these things do, with hardware. I had a TV with a streaming stick and an old phone lying around. The obvious move: mount the phone on the fridge, put a kiosk dashboard on the TV, and finally run the household from glowing screens. A weekend project. I could already picture the screenshots.
I stopped before writing any code and asked the question I’d ask any PM pitching me this idea: if and how is this actually needed? Then I did something I’ve never regretted — I interviewed myself.
Five questions, before any code
The interview was the same one I’d run on a raw product idea: whose problem are we solving first? What actually goes wrong today without it? Who authors and signs off on the thing? How does the world respond to it? And four weeks in, what’s different if it worked?
The answers rewrote the project. The failure mode wasn’t a missing display — it was that meals weren’t planned, so they defaulted to whatever was comfortable and carbohydrate-heavy, which is quietly hostile to a fitness goal. Instructions for the person who cooks for our family got lost in verbal relay. My own priorities — work, interview preparation, learning by shipping AI projects, teaching my son to swim — lived in my head and lost to the loudest thing in the room.
And one constraint outranked every feature I could have designed: the person who executes most of the household plan doesn’t read. She understands simple speech — a couple of languages — and pictures. Not text. A text dashboard for her would have been a beautifully engineered thing nobody could use. The right surface for her is audio and big pictorial cards; the right surface for me is a plan I can glance at. Constraints first; features are what’s left over.
The design bets
The plan is the product; the screens are render targets. Dashboards decay because the data behind them goes stale. What changed my week wasn’t a display — it was one artifact: a reviewed weekly plan that a daily brief is rendered from. So I built the plan pipeline first and deliberately postponed the kiosk screens.
Design for rejection. When I built the tool for choosing the week’s meals (a drag-and-drop canvas, here), I put in roughly twice as many options as there were slots. A menu with exactly seven dinners is a form to fill in; a surplus is a set of trade-offs to choose between. Saying no became part of the flow.
Measure the minimum that changes behaviour. The plan that had worked for me in the past was a meal service: the food arrived, the portions were set, and I stopped thinking about food entirely. Recreating that at home meant portioning — but weighing every ingredient is a hobby, not a system. So: two weighed items per meal, the rest eyeballed. The measurement discipline is deliberately thinner than a nutritionist would like, because a plan you actually follow beats a perfect one you abandon by Thursday.
Every loop gets an exit rule. A feedback loop without a decision rule is just a log. The weight trend has a checkpoint that says tighten, hold, or ease off. The swimming progression my son and I are working through logs how comfortable he was each session — two rough sessions and we repeat the week; two joyful ones and we advance early. The rules are written down before the data arrives, so the data can’t talk me into anything.
What shipped
Not a dashboard. A small stack of artifacts: a goal stack with an explicit priority order (one leading goal, everything else parallel), a measured weekly menu with weighed portions, a thirty-day progression for the pool, and a weight log with its decision checkpoints. Plus one real tool — the meal canvas.
I built it with an AI agent, and the most valuable thing the agent did wasn’t write code. It was interview me, hold the constraints in view every time I drifted toward the pretty thing, and keep asking what the loop was before agreeing to build the surface. The code was the easy part.
The automations — daily briefs, weekly planning runs — are configured but switched off on purpose. Run the loop by hand until it’s proven, then turn on the machine. Otherwise you’ve automated a habit you haven’t formed yet.
The transferable bit
Household or product, the discipline is the same: interview yourself before you build; name the job rather than the artifact you pictured; let the hardest constraint pick the surface; make the data the product; and ship the loop before the dashboard.
The screen was never the system.