Your agent is only as smart as what it reads. On Day 2 you are going to give it a small dataset, and it will answer questions using that data. So before you build anything, you decide where the data comes from. There are three legit sources, and you only need to pick ONE for now.
1) AI-generated synthetic data. This is the easiest one, and it is the default we use together on Day 2. You ask the AI to make up a realistic table for you. Example: 'Make a table of 15 fake clinics with name, ZIP code, services offered, whether they take Medicaid, and walk-in hours.' Nobody real is in it, you control exactly what columns you get, and it is ready in seconds. If you have never touched AI before, choose this one. You will not break anything.
2) Public or open datasets. These are real numbers the government already publishes for free, and you are allowed to use them. They are AGGREGATE, meaning they count groups of people (a county, an age band, a state), never a single named person. Real working examples for healthcare: - U.S. Census, data.census.gov, for health-insurance coverage rates by county or state (how many people are uninsured where). - CDC, data.cdc.gov, for public-health numbers like chronic disease rates, vaccination rates, and the PLACES local health data. - HealthData.gov, healthdata.gov, the U.S. Health and Human Services hub with thousands of downloadable health datasets. - Data.gov, catalog.data.gov, the giant catalog (over 500,000 datasets) where you can search a keyword like 'diabetes' or 'clinic' and filter to health. All four are real, free, and live right now. If your problem is about access or affordability, these give your agent real grounding.
3) Your own made-up examples. You type out a handful of realistic rows yourself in a note or a doc. Good when you know your topic cold (you run a clinic, you volunteer somewhere) and you just want to hand-write 10 example cases. Totally fine, just keep them invented.
Now the one rule you cannot break. Never use real personal data from a real, identifiable person. No real patient names, no real medical records, no real phone numbers, no real address tied to a real human, no screenshot of an actual person's chart, no your-cousin's-actual-diagnosis. Public aggregate stats (a county uninsured rate) are almost always fine because they are about a group, not a person. One healthcare-specific caution: glance at any row built on a very tiny population, because a single ZIP or census tract with a rare condition and only a handful of people can quietly point back to one real person even though it looks like a group number. If a public row covers thousands of people you are fine; if it covers three, skip it or zoom out to a bigger area. Data about one real human is never fine. Two reasons: it is the right thing to do in healthcare, and it protects you, because these free AI tools may train on whatever you paste in, so you never want a real person's private info going in there. When in doubt, make it up or use a public aggregate that covers a large group.
Now write it down. Add a new line to your one-page Agent Design Brief that says 'DATA SOURCE:' and name which of the three you picked plus what the data covers. Examples: 'DATA SOURCE: AI-generated, 15 fake low-cost clinics in my city with hours and Medicaid status.' Or 'DATA SOURCE: CDC PLACES public data, diabetes rates by county in Georgia.' One line. That is the whole step.
If you have done this before: pick a real public set today and skim its actual columns so Day 2 is just trimming, not hunting. Heads up, public files are often huge (tens of thousands of rows), and you will only paste a small slice on Day 2, so note now which 2 or 3 columns and which filter (one state, one county, one year) you actually need. Synthetic is still the faster path if your problem is narrow and you want clean, controllable columns.