Prompts for IME work, and the limits that matter
Short version: the safe uses of a general assistant in IME work are the ones that need no patient data: your boilerplate, your reasoning structure, literature, and checklists. Copy the 8 prompts below. For anything built on the file, check for the 6 failure modes and run the 10-minute check before you sign.
Limits: We read each product's help pages, terms and pricing on 2026-11-12 and did not test how any product handles PHI. Plans, prices and laws change, and nothing here is legal advice for your state.
8 prompts that need no patient data
Each prompt is written so you never paste PHI. Fill the brackets with general facts, not file facts. They work in any assistant; on a personal plan, turn model training off first. Use the copy button or select the text.
1. Tighten your own boilerplate
Here is the methods paragraph I use in independent medical examination reports. Rewrite it for clarity at a 10th grade reading level. Keep every factual statement. Do not add claims, citations or qualifications I did not write. Mark any sentence you think a cross-examining attorney could attack, and say why in 1 line.
[paste your boilerplate, with no examinee details]2. Stress-test a causation argument in the abstract
I am writing an IME opinion on causation. Without any patient details: a [age range] worker with [general condition] reports [general mechanism]. List the 5 strongest arguments for and the 5 strongest against a work-related cause. For each, name the type of record evidence that would support or undercut it. Do not reach a conclusion.3. Literature with sources you can check
Find peer-reviewed studies or society guidelines on [topic, for example natural history of lumbar disc herniation without surgery]. For each, give the full citation, the year, the study design, the sample size and the 1 finding relevant to an IME causation or prognosis section. If you are not sure a source exists, say so instead of citing it.4. Records to request for an injury type
For an IME on [injury type, for example rotator cuff tear after a fall at work], list the record types I should expect in a complete file and the ones most often missing. Group them by source: emergency, primary care, specialist, imaging, therapy, pharmacy, prior claims. Keep it to a checklist.5. A report outline for a referral question
Draft an outline for an IME report that answers these referral questions: [paste the questions, with names, dates and claim numbers removed]. Use these section headings in this order: [your headings]. Under each, list what evidence the section needs. Do not write findings or opinions.6. Plain-language explanation of a term
Explain [term, for example "Waddell signs" or "maximum medical improvement"] in 3 sentences for a claims adjuster with no medical training. Then give the 1 most common misunderstanding of the term and correct it.7. Check a de-identified draft for internal contradictions
Below is a report draft with all names, dates, places and numbers replaced by placeholders. List every place where 2 statements conflict, where a conclusion is not supported by a finding stated earlier, or where the same fact is given 2 different ways. Quote both passages. Do not rewrite the report.
[paste the placeholder draft]8. Audit cited AI output from a record tool
Below is a chronology exported from my record review software, with identifiers replaced by placeholders. List every entry that has no page citation, every entry whose date is outside [date range], and every pair of entries that look like duplicates of the same visit. Output a table: entry, problem, page.
[paste the de-identified chronology]Prompts 7 and 8 assume you replaced every identifier. If the text still carries real dates or claim numbers, use them only in a product under a signed BAA. The "can I paste this?" checklist helps you decide.
6 failure modes to check for
These are the 6 patterns we check every AI output on records for. The first 3 matter most in IME work, because nothing on the screen looks wrong.
| Failure | What it looks like | How to catch it |
|---|---|---|
| Omission | A prior injury, a gap in care or a failed treatment never appears | Ask "what happened before the date of injury?" as its own question; scan the file index for providers that never show up |
| Lost negative | "No radiating pain" becomes "radiating pain" | Open the page behind every symptom your opinion rests on |
| Wrong date | A fax date or print date replaces the visit date | Spot-check entries where the date falls on a weekend or long after the visit |
| Wrong or invented page | A cite points to a page that says something else, or to no page | Open 10 cites at random; 1 bad cite means check them all |
| Merged duplicates | The same MRI counted as 3 studies, or 3 visits merged into 1 | Compare the count of imaging studies against the imaging index |
| Invented source | A study or guideline that does not exist, in a literature answer | Look up each citation yourself before it goes in the report |
The 10-minute check before you sign
- List the 5 to 10 facts your opinion depends on. Date of injury, mechanism, prior history, imaging findings, treatment response, current status.
- Open the cited page for each one. Read the line, not the summary of it.
- Ask 1 question 2 ways. "When was back pain first documented?" and "List every visit before 2023 that mentions the back." Different answers mean an omission.
- Open 10 random cites from parts of the file you did not rely on.
- Check the records list against what the referral says was sent. In New York, the report must list every record received and reviewed.
If a tool cannot give you a page for a fact, you cannot do step 2, and the fact should not be in your report until you find the page yourself.
Limits by tool type
- Consumer chatbots: no BAA, so no file content at all. Training may be on by default.
- BAA-covered general assistants: legal for PHI on the covered features, but no page citations to your upload in the help pages we read, and large files hit published caps. Fine for a finished report, weak for a 2,000-page packet.
- Clinician literature tools (OpenEvidence, ChatGPT for Clinicians, Doximity Ask): they cite journals. They do not know your file.
- IME record tools: they cite pages, but a citation proves where a fact came from, not that every relevant fact was found. Step 3 above exists for that.
No tool in any category forms the opinion, and in California workers' comp only the signing physician may review and summarize records or draft conclusions under Labor Code 4628. The rules page has the details.