How to prompt AI for medical record review

What record review AI does, and what it does not

It does four things well.

It reads everything. A ten-thousand-page record set gets the same attention on page 9,000 as on page 3. Human review degrades across a long file. Machine review does not degrade in that particular way.

It finds the same thing everywhere it appears. Ask for every mention of a body region and you get every mention, including the one buried in a nursing note from an unrelated admission.

It holds many documents in view at once. Contradiction detection across twelve providers is a task humans do badly, not because it is hard to judge but because it requires holding twelve documents in working memory at the same time.

It restructures. The same facts as a timeline, as a table by provider, as a summary by body region, in seconds each.

It does four things badly, and knowing which is the whole game.

It cannot reliably tell you what is absent. A model asked "are there gaps in treatment" will answer from what it sees. Absence is not a pattern in the text; it is a hole in the text. Prompts must construct absence questions carefully, which is what Section 3 is about.

It conflates who said what. Medical records are layered speech: the patient tells the intake nurse, the nurse writes it, the physician repeats it in a history, and three years later it reads like a clinical finding. Models flatten this. So do tired humans, but a model does it confidently.

It degrades on bad scans. Handwritten notes, faxed records, stamped dates, and rotated pages produce OCR errors that surface as confident wrong dates and misread values. The output looks identical whether the input was clean or garbage.

It fills gaps with plausibility. Asked a question the record cannot answer, an unconstrained model will produce a reasonable-sounding answer. This is the failure that ends careers, and Section 5 covers detection.

None of these are arguments against using it. They are the reasons the prompts in this library are shaped the way they are.

Two doctrines

Everything in Part II and Part III follows from two rules. If you remember nothing else, remember these, because they let you write your own prompts.

Doctrine one: three-state answers

Every factual claim about a record set is in one of three states.

Record-supported. The record says it, and you can cite the page. Record-silent. The record does not address it. This is a finding, not a failure. Inferred. Someone reasoned to it from what the record says. Possibly correctly, but it is not the record talking.

Most bad AI record output collapses all three into one confident voice. Most bad human summaries do too. The fix is to demand the distinction in the prompt itself, every time.

In practice this means writing "cite the source for each element, and note any element the records do not establish" rather than "summarize the incident." It means "separate provider findings from patient-reported symptoms." It means "report the alignment as an observation with dates rather than a conclusion."

Every prompt in this library enforces at least one of those three separations. When you write your own, enforce one too.

The three-state rule matters differently by lane. For a plaintiff attorney it prevents a demand letter built on an inference the record will not support at deposition. For an IME physician it is the difference between a defensible report and one that gets excluded. For an expert witness it is the line between an opinion and a guess, and opposing counsel will find that line whether or not you did.

Doctrine two: absence is a finding

The highest-value findings in medical record review are usually absences.

The gap between February and August. The referral to neurosurgery with no neurosurgery record. The recommended fusion that was never scheduled. The MMI determination no provider ever made. The imaging the note references and the file does not contain. The functional capacity evaluation nobody ordered.

Extraction-shaped prompts miss all of these, because absence produces no text to extract. You have to build the absence question with a scaffold: establish the expected set, then check the record against it.

Three scaffolds recur throughout this library.

Referral-to-encounter. List all referrals. List all encounters. Report referrals with no matching encounter. Reference-to-document. List every imaging study, report, or record referenced inside a note. Check each against the produced file set. Report references with no document. Billing-to-record. List all billed services. List all clinical records. Report billed services with no supporting record, and date ranges with billing but no notes.

Each scaffold turns "what is missing" into two extraction questions and a comparison, which is work a model does reliably.

When a prompt in this library asks the model to "report if none exists" or "state explicitly where the record establishes no baseline," that clause is not filler. Remove it and the prompt will quietly return only what is there.

The anatomy of a working prompt

Six elements. Not every prompt needs all six, but every prompt needs the first three.

1. The scope. What subset of the record, and what date boundary. "After <incident date>." "Across the full record set." "Limited to records postdating the referral." Unscoped prompts return unscoped answers.

2. The output shape. Table, timeline, list, one row per finding. Name the columns. A prompt that does not specify shape returns prose, and prose cannot be checked against a record efficiently.

3. The citation requirement. Per row, per finding, per quotation. Say which. An output without citations cannot be verified, and an unverifiable output cannot be used.

4. The separation. Which of the three states must be kept apart here. Objective from subjective, provider finding from patient report, observation from conclusion, record-supported from inferred.

5. The absence clause. What the model must report if it finds nothing, or what it must flag as not established. Without this the model reports only presence.

6. The verbatim requirement. For anything that will be quoted, argued, or challenged, demand the exact language rather than a paraphrase. Paraphrase is where meaning drifts, and drift is what gets impeached.

Compare. Weak: *summarize the treatment timeline.* Working: *build a chronological timeline of every medical encounter. For each row: date, provider, encounter type, presenting complaint, objective findings, assessment, and any change to the treatment plan. Cite the source document per row. Note any date range with billing activity but no corresponding note.*

Same task. The second one produces something you can check.

Six failure modes and how to catch each one

These are ordered by how much damage they do.

5.1 The fabricated citation

The model attributes a real-sounding finding to a page that does not contain it, or to a document that does not exist. This is the failure that has produced sanctions in reported cases and it is the one to design against first.

Detection. Spot-check citations against the source before any output leaves your desk. For anything that will be filed, deposed on, or relied on in an opinion, check every citation, not a sample. Check the page, not just the document.

Reduction. Demand verbatim quotation alongside the citation. A model that must reproduce the language is much less able to invent the finding. Quotations are checkable in seconds; paraphrases require reading the page.

5.2 The confident partial

The record set is incomplete and the summary does not say so. The output reads as a complete account of the treatment because nothing in it signals otherwise, and the missing orthopedic file simply does not appear.

Detection. Run the completeness prompts before the substantive ones. S2 provider roster and S4 missing records audit exist for this reason and belong at the start of every review, not the end.

Reduction. Always pair a summary prompt with a scope statement: what record set this covers, which providers are represented, and what is known to be outstanding.

5.3 Speaker collapse

The patient's account becomes a clinical finding. "Patient reports constant burning pain radiating to the left hand" becomes "documented radicular symptoms." The distinction survives in the original note and dies in the summary.

This one is quietly the most consequential in causation disputes, because the entire argument often turns on whether a finding is objective or reported.

Detection. For any finding that matters, open the source and read who is speaking.

Reduction. The separation clause. "Distinguish provider findings from patient-reported symptoms." "Separate objective findings from subjective reports, in two columns."

5.4 Date drift

Scanned and faxed records produce OCR date errors. A 2019 becomes 2016. A stamped receipt date gets read as an encounter date. The timeline reorders itself around the error and the resulting narrative is wrong in a way that reads perfectly smoothly.

Detection. Sort the timeline and look at the boundaries: the earliest entry, the latest entry, and anything out of sequence relative to its neighbors. Errors cluster at the edges and in handwritten material.

Reduction. Ask for the date as it appears in the document alongside the normalized date. Discrepancies become visible instead of silent.

5.5 The unasked absence

Covered in Section 3. The model reports what it found and stays quiet about what it did not, and quiet reads like clean.

Detection. If an output contains no absences, no gaps, and no missing records, that is a signal to re-run with an explicit absence scaffold, not a sign of a clean file. Genuinely complete record sets are rare.

5.6 Anchoring to the framing

Ask "identify the weaknesses in this case" and the model finds weaknesses, including some that are not there. Ask "identify the strengths" of the same file and it finds those. The framing shapes the finding.

This one matters most for IME and QME physicians and for expert witnesses, whose value rests on not having done this.

Detection. Run the mirror. Ask for contrary evidence with the same specificity you asked for supporting evidence. If the two lists are wildly asymmetric, the asymmetry may be in the prompt rather than the record.

Reduction. Neutral phrasing. "Report the alignment as an observation with dates rather than a conclusion." Prompts I6 and E2 exist entirely to force the mirror.

Verification, tiered by consequence

Verify in proportion to what an error would cost. Three tiers.

Tier 1. Check everything. Anything filed with a court, sworn to, served on an opposing party, or relied on in a written expert opinion or IME report. Every citation opened, every quotation matched, every date confirmed against the source. No exceptions and no sampling. The output is a draft until this is done.

Tier 2. Check what carries weight. Internal case memos, demand letter drafts, reserve evaluations, mediation preparation. Verify every finding that would change a decision or a number: causation statements, permanency and impairment language, restrictions, gaps, contradictions, cost drivers. Spot-check the rest at roughly one in five.

Tier 3. Check the shape. Orientation summaries, provider rosters, first-pass triage. Confirm the output is structurally sound, the date range is right, and the provider count matches expectation. Detailed verification comes when a finding gets used.

Two rules cut across all three tiers.

Anything you will quote gets checked at the source, regardless of tier. A quotation is a representation about a document.

Anything that moves between people gets its tier attached. If a Tier 3 orientation summary is going to a colleague, say that it is a Tier 3 orientation summary. Most verification failures are handoff failures, where work produced as a rough first pass gets used as a finished product by someone who never knew which it was.

Run order and chaining

Prompts are not independent. Run them in this order and each one improves the next.

Stage 1, establish the record. S2 provider roster, then S4 missing records audit. Do this before anything substantive. You are answering "what am I actually looking at" and the answer changes how much every later output is worth. If S4 returns significant gaps, resolve them before building a chronology you will have to rebuild.

Stage 2, build the spine. S5 master treatment timeline. Everything downstream reads better with this in hand, and errors caught here are errors caught once. Then S1 incident summary and S3 body region inventory for orientation.

Stage 3, split time. S6 pre- and post-incident split. This is the hinge of every causation argument in every lane, plaintiff and defense alike, and it should happen before anyone forms a view.

Stage 4, clinical detail. S7 through S13. Imaging, coding, injuries, trajectory, medications, MMI, functional findings. Run what your lane needs.

Stage 5, integrity. S14 through S20. Gaps, missed appointments, contradictions, pre-existing conditions, intervening events, timeline correlation. Run all of these regardless of lane. The side that finds them second is the side that is surprised.

Stage 6, causation. S21 and S22.

Stage 7, your lane. Now run the lane pack, which assumes the spine is built.

Two chaining habits worth forming.

Feed outputs forward explicitly. When a lane prompt depends on a spine output, say so in the prompt: "using the treatment timeline established above." This keeps the model working from the verified artifact rather than re-deriving it, and re-derivation is where inconsistency enters.

Re-run after new records arrive. A supplemental production invalidates every gap analysis and every completeness finding you have. S4, S14, and S15 in particular are only true as of the record set they ran on. Date every output with the production it covers.

What good output looks like

Structure only. Every cell below is a placeholder in angle brackets, and nothing here asserts a clinical fact. A blank template gives shape without making a claim, which is the only kind of example this section can honestly carry. Worked examples on real material belong with the founder-provided sample files.

A timeline row from S5 has seven fields and two dates:

The two date columns exist so OCR drift is visible rather than silent. The plan-change column is what makes the timeline readable as a course of treatment instead of a list.

A gap finding from S14 is not a duration, it is a duration plus a reason plus a status:

That third status value is the one people leave out. A gap that is really an unproduced file is not a gap, and S4 is what tells you which you have.

A contradiction from S16 is always a pair, never a summary:

Both statements quoted, both cited, both dated. If the output gives you a characterization instead of the two quotations, it cannot be checked, and an unverifiable contradiction is not usable.

Three signals that output is working, whatever the prompt:

Citations resolve to a page, not a document. "Smith records" is not a citation. "Smith records, p. 214" is. Absences appear. A result with no gaps, no missing references, and no unestablished items usually means the prompt did not ask. Attribution survives. You can still tell who said each thing: the clinician, the patient, or the analysis.

---