Review a hundred-page case file in minutes, with the citations at hand

Who it is for

Anyone who decides based on case files: insurance claims adjusters, credit analysts, procurement teams evaluating tenders, firms reviewing background records. Any job where reading a hundred pages is the bottleneck.

The problem

Every case comes with a long, disorganized file. According to the source, a medical insurance claims adjuster spends more than 100 minutes per case just reviewing records. The key information is spread across different documents and nobody can read all of it carefully.

Ingredients

• Case files in PDF with selectable text (if they are scans, run them through OCR first)
• The list of the eight to twelve questions you always answer when reviewing a case
• A paid claude.ai or ChatGPT account, or API access for volume
• Two case files already reviewed by a person, for comparison

Step by step

1. 1

   Write down the questions behind your decision

   Do not ask for "a summary". List what you actually look at: key dates, amounts, who signs, what is missing, what is contradictory. Those questions are the backbone of the prompt and of the output.

2. 2

   Require a source citation for every data point

   The prompt must ask for the page number or document name next to every statement. That is what turns a nice-looking summary into one you can trust and verify.

3. 3

   Test with two cases already resolved

   Run the prompt on case files a person has already reviewed and compare: did it find the same things? Did it invent anything? Adjust the questions until it matches on the essentials. Do not skip this step.

4. 4

   Explicitly ask for the gaps

   A mandatory section for "what is not there" and "what is contradictory" is worth as much as the summary: that is where the problem cases hide.

5. 5

   Scale only what works

   Once the prompt gets real cases right, then think about the API and a workflow that processes whole folders. According to the source, EXL built its solution by combining document extraction with specialized language models; you start with the prompt.

Replace what is in brackets with your company details, or use the guided mode above: it asks you and fills it in for you.

# Role
You are the senior analyst for [TYPE OF CASE FILES, FOR EXAMPLE "insurance claims"] at [COMPANY NAME]. You read rigorously, you do not fill in with assumptions what the case file does not say, and you always indicate where you got each data point.

# Context
Attached is the complete case file. The decision I have to make is: [DECISION YOU MAKE WITH THE CASE FILE]. The questions I always need to answer are:
[LIST OF KEY QUESTIONS, ONE PER LINE]

# Task
1. Answer each question with evidence from the case file, citing document and page.
2. List the data missing to decide and the data that conflicts between documents.
3. Point out any date, amount or name that appears with different values in two places.

# Output format
Section A, "Answers": question → short answer → citation (document, page). Section B, "Missing". Section C, "Contradictions". Section D, "Timeline" with the relevant dates in order. If an answer is not in the case file, write "Not on record" instead of inferring it.

# Examples
Question: "Date of loss?". Expected answer: "March 14, 2026 · Claim notice, p. 2. Note: the medical report (p. 11) says March 15; see Contradictions".

What usually goes wrong

• ✕Asking "summarize this case file". You get a general text that does not answer your questions and has no citations: useless for deciding.
• ✕Uploading scans without text. The model misreads or does not read at all; run the PDFs through OCR and check that the text can be selected.
• ✕Trusting without comparing. The first two cases are always checked against the human review; that is where the missing questions show up.