Fieldnote

THE PIVOT

My first prototype was a cleaner version of what already exists, somewhere between Noom and a tidy logging app. I built it in Figma Make far enough to see the problem: better designed, not more useful. Things got logged, data appeared as a chart, and the burden of deciding what to do with it stayed exactly where it had always been.

So I made the decision that reframed everything after it: stop designing a place to put data and start giving people a reason to collect it. The home screen stopped being a glucose graph and started asking a question — what do you want to figure out about your body? Everything downstream of that question got easier to decide.

Finding the variance before asking for any effort

Onboarding connects Dexcom and Omnipod, reads thirty days of history the person already has, and reports back where their variance actually lives: glucose running about 40 mg/dL higher on pod days three and four, afternoons spiking around 35 above other meals. No empty log, no setup labor. The product proves it has something to say before it asks for a single minute of work.

Showing the sample behind every claim

An LLM is reading someone’s health data here, so every finding carries an explicit confidence label tied to the sample it came from: twenty-four of twenty-eight pod cycles, nineteen of thirty days. A strong signal and a weak one should not look alike. The analysis screen also states that processing runs on-device, placed at the exact moment someone is deciding whether to hand over three months of their own medical history.

Committing to a question before collecting anything

Every experiment states its hypothesis as a sentence the user agrees to test, and shows what they will see at the end, before day one. Required and optional fields carry filled and open dots, so the minimum bar for a valid day is visible rather than implied. For questions no protocol covers, a plain description — I keep waking up around 250 and I don’t know why — returns a designed ten-day protocol with the user’s own sentence quoted back, so they can check the system understood them before committing three weeks to it.

Making the daily loop small enough to finish

The check-in is one question per screen across five steps, with a segmented progress bar and a skip on every optional one. The active experiment screen is a single card: which day you are on, what tonight’s condition is, one button. The protocol remembers where you are in the cycle so the user doesn’t have to. My principle throughout: an experiment nobody finishes teaches nothing, so every design choice went to the twenty-first day rather than the first.

Keeping the AI quiet most of the time

The obvious thing to build with an LLM is a daily feed of insights, refreshed every morning. But someone managing Type 1 already has a CGM talking to them all day, and a second voice with opinions is not help, it is more anxiety. Fieldnote’s intelligence appears three times: at onboarding, when an experiment is proposed, and when one ends. I also cut summary dashboards, general logging, duplicated manufacturer values, bolus suggestions, and open-ended chat, each of which was a different way of handing the burden back.

Moving fast enough to throw work away

Generating first concepts in Claude Design and Figma Make put a full working direction in front of me in hours instead of days — high enough fidelity to evaluate faster than low-fi sketching would have allowed. It also made the pivot possible: throwing away a logging app I’d spent a week on was a decision I could make; throwing away a month of it would not have been. Drafting experiment content at speed let me evaluate ten protocols instead of designing two.

Steering against the default of more

AI reaches for more by default: more fields, more insights, more screens. Nearly every design decision on this project turned out to be a subtraction. It also reaches for certainty, stating findings as conclusions. The confidence labels, the sample sizes, and the patient-generated framing on the clinical report are all constraints I imposed on it, because in a health product the distance between a finding and a suggestion is the entire product.