Ask a Claude agent questions about your WHOOP and blood test data
Idea + starter prompt, not run by us yet
For: Self-trackers with wearable data and lab PDFs who can't query it easily
The author kept a WHOOP export and lab results from emailed PDFs in Databricks tables, then replaced Databricks' Genie agent with his own light Claude agent. The agent drafts SQL, sends it to a read-only tool, looks at what comes back and can ask again. After he fixed two date bugs in his data, Sonnet went from 47% to 90% correct on 30 test questions.
Steps
- Put your data in tables you can query: the WHOOP export and the lab values pulled from your PDFs.
- Give the agent a read-only SQL tool, so it can look but never change anything.
- Write a set of questions with answers you checked by hand, and score the agent against them.
- Check the source data first. Every agent missed the questions affected by two date bugs in the author's table.
- Treat the scores with care. The test was small, and a Claude model graded Claude's answers.
Starter prompt
Test this workflow using synthetic health records in [test database or file]. Do not open real health records or shared files. Use a read-only SQL tool: you may read but never insert, update or delete. First show me the tables and what each column means, and tell me how you plan to answer [my question]. For every answer, show the SQL and the rows it returned. Check the date fields for mistakes before you trust them. This is not medical advice, so say when the data is too thin to tell. Never ask for passwords or account logins.
Original post: GitHub · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.