I heard Greg Brockman on a podcast describing AI at its best as a team of doctors in your pocket — there for you, thinking continuously about how to help. The phrase stuck. I read it less as inspiration and more as a design problem.

Most health apps display data. A doctor pays attention. Those are different products.

I had been tracking myself for years — Whoop, Strava, Apple Health, all collecting data I rarely looked back on. I was curious to use that data as a way to learn what it would actually mean to have a personal doctor in my pocket. I built Signal in conversation with Claude Code. What follows is the result — and, where it helps, the conversations that produced it.

Below the daily Read sits the long-term context layer. Connects to Today — Walking HR has shifted -16.0 bpm this past month — could be improved fitness or a slower walking pace. The agent surfaces ambiguity rather than picking a story. Most days this section is quiet, by design. Long-term observations only surface when they actually inform today's interpretation.

The Watch List below operates by the same principle. Each item has a state (W watching, N notice, F fading) and a day count. They don't fire as alerts; they have lifecycles. Anything competing for the user's attention has to earn its place.

The decision to combine the old two-label treatment into the single "CONNECTS TO TODAY" header was the result of seeing three prototypes the AI built side by side.

The AI built all three competently. The choice — which one was right — was mine alone. That moment clarified the division of labor: the AI does competent execution across parallel possibilities; the designer picks which possibility is right for the product.

Signal isn't chat-first, but chat exists as a depth mechanism. When the briefing surfaces something, the user can tap to ask more. The chat opens with suggested questions drawn from today's data, so the user rarely starts from a blank prompt.

The agent's answers cite specific data points by number. "You've slept 6.2 hours but need 10.7 — a 4.5-hour gap — and over the past 7 days you've accumulated roughly 28 hours of sleep debt." Each cited claim has a numbered footnote linking to the underlying record.

The same flow on mobile — the question, then the answer with its citations — proving the pattern holds in the place most of these checks actually happen.

The cited-sources pattern is one of the more important design moves in the product. An agent that asserts without showing its math is hard to trust. An agent that asserts with a footnote is easier to evaluate. The user can decide whether the evidence supports the conclusion, rather than just trusting the prose.

An agent paying attention across time can't live on a single screen. Signal has four surfaces, each calibrated to a different rhythm of attention. The labels themselves took three passes to get right.

Each label names what the agent is paying attention to at that scale, not just what time period the page covers.

Recent is the weekly synthesis. The strain-and-recovery arc, the weekly Take organized into load/recovery/sleep threads, the resolved watch items, the cross-week swimlanes. Open on a Sunday — or, in this case, a Friday when the weekly arc has become the dominant story.

Long View is the multi-month picture. Resting heart rate over 90 days. Walking heart rate trending down with the ambiguity acknowledged in plain text. VO₂ Max as a sparse 24-month line. Mindfulness as a 12-week calendar heatmap.

Memory is the agent's working transcript. Every check, every observation, every table. The principle: agents users can't inspect are agents users can't trust.

Most AI products either hide all their controls or expose so many they're unusable. Signal resolves this by separating what the agent generates from what gets surfaced at all. Each card has an AI/DATA pill at the top — AI controls govern generation (tone, length, voice continuity); DATA controls govern the deterministic rules that decide what qualifies for the agent's attention.

When you change a setting, you preview the effect against current data before committing. Current / Preview / Diff. The save button reads "Looks good — save" instead of "Save" — framing the commit as confirmation that the preview looked right, not as a default action.

This page is where the labeling discipline of the rest of the product compounds.

AI products default to a system-y register because that's what their training data sounds like. Pulling them toward warmth has to happen explicitly, and it compounds across the product.

Updates on the horizon.

A feedback loop. The agent has no way to know whether its daily reads are landing. A thumbs-up/down stored over time, fed back into the prompt as calibration, would close the most important loop the product is missing.

Persistent life-context. The agent knows my training routine and today's plan, but doesn't know about work stress, travel, or illness. A small notes field for non-physiological context would substantially improve the daily reads on weeks when something unusual is happening.

The whole project was specified in prompts that read more like rigorous design specs than instructions — scoped, with explicit constraints, anticipated failure modes, and "don't do this" sections that were often more important than the positive direction.

The exercise gave me real insight into how design iteration through prompting produces a faster path to a working product. The artifacts are screens; the work is in the dialogue.