Finance & bookkeeping · Claude

Check a hospital bill line by line before you negotiate

Idea + starter prompt, not run by us yet

For: Anyone facing a large itemised medical bill they think is padded

A patient obtained the coded billing form from the hospital, then asked Claude for a spreadsheet of Medicare payment rates per code, flagging anything that needed more research. Claude found unbundled charges and a couple of other oddities. He had ChatGPT check it, spot-checked it himself, wrote the letter and negotiated. The final settlement was a small fraction of the bill.

Steps

1. Ask the hospital for the itemised, coded bill. It lists a procedure code for each charge.
2. Tell Claude your insurance type, location and year, then ask it to build a table of reference prices per code.
3. Ask follow-up questions on blank or zero rows. In the original run, bundled codes came back as zero.
4. Verify the analysis yourself and with a second tool before you rely on it.
5. Write the dispute letter yourself and negotiate directly. Claude only did the analysis here. It is not a lawyer.

Starter prompt

Trial first: use a made-up bill of [5] invented lines with real-looking codes, or a copy of my bill with my name and ID removed. Save the trial table in [trial-table]. Do not use the full original yet.

This job only reads and drafts. It changes nothing outside [output folder] and contacts no one. After my yes, here is my itemised hospital bill with billing codes. I am in [location], my insurance is [type or none] and the service date is [year].

Build a table with each code, what was charged, a reference price from public sources, and the gap. Name the source for every price. Flag anything that looks unbundled, duplicated or unusual and say what I should ask the hospital. Where a row is blank or zero, explain why. Do not state conclusions as fact when unsure. I will write and send any letter myself. I will not paste ID numbers or card details.

Original post: aol.com · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.