Card 10 of 13 · after the visit
delegable draft, review required

Referral letters and replies

The letter travels. Whatever goes wrong here does not stay in your office: it lands on a colleague's desk with your signature and becomes the starting point for their assessment. And formal prose lowers the reader's guard, which makes this prime territory for the elegant error.

The case

Twelve letters to write, and the tool drafts a clean one in seconds from four lines of input. It reads better than what you would have written tired at 7 p.m. The question is whether everything in it came from you.

Guiding question: Is there anything in this letter I did not write?

Structuring the letterDelegable
Choosing what goes inYours
Framing the questionYours
Signing itNot delegable
01 · What it does, and what it does not
What it does well

Builds a clear letter in seconds from what you provide: question, relevant summary, workup already done, medications. Adjusts register for the recipient. And drafts the reply back to the referring clinician with the same discipline, in reverse.

What it does not do

Know what you did not tell it, or notice that it filled a gap with something plausible. The skeleton is delegable; every fact inside it is yours.

Clinical design position

The failure here is not bad writing. It is good writing that contains something you never said, and the better the prose, the less likely you are to catch it. So this card puts almost all of its weight on the input and the review, and almost none on the draft itself.

02 · Where it fails here
The invented detail

An exam finding you did not report, a date nobody gave, a history item that does not exist, written in the sober register of a medical document. Surrounded by correct sentences, a fabricated detail reads like a chart entry.

The certainty upgrade

The subtler version of the same failure: your hypothesis goes into the elegant draft and comes out as an established diagnosis, and it travels under your signature.

The lost question

The reason for referral is the soul of the letter, and the model dilutes it into "referring for evaluation." Without a clear question and a stated expectation, the specialist starts from zero and the letter failed at its only job.

And the letter that says too much

The reader here is another clinician, also bound by confidentiality, and the minimum-relevant principle still applies: the letter carries what serves this assessment, not the entire chart. A sensitive detail that does not change the specialist's assessment does not go in by default; it goes in when it is relevant, by your decision. Keep the ruler in mind, though: when the reader is outside confidentiality, an employer, a school, a court, the same defect changes category, and the principle tightens from minimum relevant to minimum necessary. That is Card 12.

03 · The safer workflow
01The question before the draft.What exactly you want from your colleague, and how urgently. One line, yours. The letter exists to carry that line; everything else is context.
02Minimum relevant input.You select what goes in; you do not paste the chart. The draft is born anonymous, and identifying details are added afterward, in the final document, outside the chat.
03Gaps become brackets, not prose."Do not add findings, dates, values, or history I did not provide; mark anything missing as [confirm]."
04Fixed structure.Question and urgency at the top; relevant summary; worked up and treated; current medications; your hypothesis, as a hypothesis; what you expect; your contact.
05Directed review before signing, and the same mold in reverse for the reply.Every date, dose, and finding against the source; the hunt for additions; the relevance pass; the question still clear at the top; certainty at the right level. The reply back to a referring colleague follows the same rules: what you found, what you did, what you are handing back, what they should monitor.
04 · The bracket technique
Why a gap must look like a gap

Left to itself, a language model resolves missing information the way prose does: smoothly. The dose it was not given becomes a plausible dose, the missing date becomes a plausible date, and nothing in the sentence signals that anything was supplied.

A bracket in the middle of a letter is an honest gap you close in seconds. A fluent sentence in the same spot is a fabrication you will not see.

What it costs, and what it buys

It costs a slightly uglier draft, and one pass of filling in blanks before the letter goes out.

It buys the only reliable way to tell, at a glance, which facts in this letter came from you and which the tool would have had to invent.

Console · exercise

Find what you did not write

Synthetic case. What you gave the tool: 52-year-old woman, three months of altered bowel habits and iron deficiency anemia (Hb 10.2, low ferritin); no known prior colonoscopy; taking losartan; your hypothesis: lower GI blood loss, to be investigated; your question: prioritization for colonoscopy.

The letter that came back
I am referring a 52-year-old woman with confirmed lower gastrointestinal bleeding of three months' duration and iron deficiency anemia (Hb 10.2; low ferritin). On examination, the abdomen was soft, non-tender, with no palpable masses. Colonoscopy five years ago was unremarkable. History of a depressive episode in 2019, in remission. Currently taking losartan 50 mg daily. Requesting evaluation for colonoscopy on an expedited basis.

Well structured, correctly formatted, and ready to send. Five of its clauses did not come from you.

The reveal, in order of danger
worst"Colonoscopy five years ago was unremarkable." You said there was no known prior study. An invented recent normal exam is exactly the kind of detail that leads a colleague to deprioritize the patient who most needed the workup.
certainty"Confirmed lower gastrointestinal bleeding." Your hypothesis left the draft as an established diagnosis, and it travels under your signature.
exam"On examination, the abdomen was soft..." An exam you did not provide. In formal register, it reads as performed.
precision"Losartan 50 mg daily." You did not give a dose; the letter picked one. A small fabrication wearing the look of precision, and exactly what [confirm] exists for.
relevance"History of a depressive episode in 2019." True in the chart, and it does not change a colonoscopy prioritization. It comes out, not because it is false, but because it is not relevant here.
Four fabrications and one relevance call

The first four are errors: the letter states things that are not true or not established. The fifth is not an error at all, which is what makes it worth practicing.

Deciding what to leave out is a clinical judgment about the reader, and it is the one part of this review the tool can never make for you.

What survived well

The complaint, the labs, the clear question with urgency at the top.

The skeleton was good. The contents needed you, which is the whole reason this card is a delegable draft and not a delegable letter.

05 · Before you sign

0 of 6

The second box is the one that catches the elegant error, and it is best done by reading the letter against your input rather than against your memory of the visit.

The prompt
Draft a referral letter to [specialty].
Question for my colleague, which should open the letter:
[opinion / take over care / procedure] with urgency
[routine / expedited / urgent].
Relevant case: [summary aimed at the question].
Already worked up: [tests with dates and results].
Current medications: [list with doses].
My hypothesis, which must remain labeled as a hypothesis:
[hypothesis].
Rules: use only what I provided; do not add findings, dates,
values, or history; mark anything missing as [confirm]; do not
include patient identifiers; do not upgrade a hypothesis into
a diagnosis.
06 · Evidence

Generated clinical letters have been rated highly on formal quality since 2023, and that is precisely the trap of this card: the fluency that impresses a reviewer is the same fluency that hides the invented detail from the person signing.

  1. [article] Ali SR, Dobbs TD, Hutchings HA, Whitaker IS. Using ChatGPT to write patient clinic letters. Lancet Digital Health, 2023;5:e179-e181.
  2. [article] BRIDGE: benchmarking large language models for understanding real-world clinical practice texts. Nature Biomedical Engineering, 2026. The gap between form and clinical fidelity. doi.org/10.1038/s41551-026-01719-2
Educational content; synthetic case and letter. Not a substitute for clinical judgment or for the rules that apply where you practice.