I had more health data than ever, spread across six places: Apple Watch, a glucose sensor, 15 DEXA scans, a lifting app, two years of a paper bullet journal, and photos of my meals. Each app answered its own question. None answered mine: what should I do today?

Data I couldn't act on

In 2025 my body fat went from 36.2% to 25.6%, but my lean mass barely moved. Looking back, the reasons were in my data the whole time:

• I lifted weights only 4 times between January and September 2026. My classes felt hard, but they never added weight week by week.
• My sleep averaged 6.3 hours across 161 logged nights this year, short of my 7-hour goal. Muscle repairs during sleep.
• Two of my DEXA scans were taken right after a festival and a long flight, so they likely understated my muscle.

No single app could connect these. A person could, if they read everything every morning. That's the job I gave the agent.

The agent lives in a Claude project. Two documents are its memory: a baseline with my goal, my dietitian's plan and my lifting program, and a data log it updates itself. Each morning an iPhone Shortcut reads Apple Health and asks Claude for my brief; everything else I send as files or photos.

1. The morning brief: one decision a day

An iPhone Shortcut runs every morning. It reads last night's sleep and HRV and turns them into a training call I can follow without thinking:

It ends with one food reminder, such as spreading protein across meals, then adds yesterday's numbers to a running health log. The answer lands on my phone as a notification, which my watch also shows.

How I built it on my iPhone

I built the brief in Apple Shortcuts, so it runs on my own phone with no server. Shortcuts pulls the data, Claude does the thinking.

1. An automation runs it every day

2. It pulls sleep, HRV, heart rate, steps, glucose and weight from Apple Health

3. My rules and the data go to Claude; the reply is logged and shown as a notification

The prompt is where most of the design lives: who Claude is, my targets, a fixed answer format, and the sleep-based training rules. It also handles messy data. My watch and phone both record sleep, so the stages can add up to 14 hours. I added a rule: if the stages add up to more than 12 hours, two sources overlap, so halve them. The brief now catches it and explains what it did.

A real brief. It spotted the double-counted sleep and corrected it.

Debugging with real data

Getting real Health data into Claude took a few rounds. Shortcuts blocked the first run because it tried to share 472 Health items at once; I allowed large data sharing in Shortcuts settings. A statistics step then failed because it read numbers as text, so I removed it and let Claude summarize the raw samples instead.

Blocked: too many Health items in one run

Fix: allow sharing large amounts of data

A math step failed on text, so Claude does the summary instead

2. The health report: day, week, month and year

The agent built a report I can open anytime, in English and Chinese. It starts with the view I care about most: my bullet journal stickers, rebuilt as a calendar from photos of the paper pages.

The week view puts training, steps, sleep, HRV and glucose side by side, so patterns jump out without me hunting across apps.

3. The lesson of the day: learning from my own data

Every day the agent explains one nutrition or training idea, using something that actually happened to me. On October 11 my glucose hit 275 after a syrup coffee, a croissant and fried rice. Instead of a warning, I got a lesson on why it happened and three easy fixes for next time.

Glucose on October 11, from my Stelo sensor

The lesson the agent wrote about that day

Design the rules, not just the screens

Most of the design work was writing: the agent's role, what it may and may not say, its tone, and the decision rules above. For an agent, the instructions are the interface.

Every number shows its source

Each card carries a small label such as Apple Watch, Stelo or DEXA. When data is missing or unreliable, like a workout I forgot to start on my watch, the report says so instead of guessing. Trust comes from being honest about gaps.

Meet the data where it already is

I didn't change my habits to feed the agent. It reads photos of my paper journal and my meals, files from my lifting app, and Apple Health directly through a Shortcut. The easier the input, the more likely I keep it going.

Explain, don't diagnose

Health advice needs clear limits. The agent acts as a nutritionist and coach. For anything medical, it explains what a number means, how urgent it is, and what to ask my doctor.