Reading time: about 11 minutes
This case is not about writing better: it is about what does not leave your pharmacy. And it has a very concrete shape, because "anonymise" sounds like a lot and in practice is five deletions. What you learn here you will use in every case that follows.
Step by step
-
Copy the text into a notepad, not into the chat
Open your notepad and paste the text there. This intermediate step looks trivial and is what makes the whole thing work: it separates "I have the document" from "I have sent it", and thinking fits between the two.
Pasting straight into the chat is irreversible. Deleting the message afterwards does not undo the send: it has already gone.
-
The five deletions, in this order
You do not need a complicated rule. Find and replace these five things, always with a label, never with a plain deletion — leave a gap and the model invents what goes in it:
What you remove Why What you put instead Full name Identifies on its own [PATIENT]ID number, record number, card number Identifies on its own [ID]Phone, email, address Identifies on its own [CONTACT]Date of birth With two more details, identifies [74 years old]← the age you do needYour pharmacy’s name and the doctor’s Identifies the other party [PHARMACY],[DOCTOR]Look at the fourth row. The date of birth goes and the AGE stays, because age is what changes the clinical answer and identifies nobody. That is the whole rule: remove what points at a person, keep what the question needs. -
Read it once more looking for what is not a field
The five deletions catch whatever looks like a field. What does not look like a field stays, and in a small town that is what really identifies: "the woman from the corner shop", "the nurse’s son", "the one who comes on Tuesdays with the dog".
No tool and no rule catches that: you catch it by reading. Replace it with
[PATIENT]too. If removing it makes the question stop making sense, that detail was part of the case — then keep it general: "an older person living alone". -
Replace, do not delete
A small detail that changes the whole result. If you delete the data and leave the gap, the sentence goes lame and the model fills the gap with whatever seems to fit — and that guess then reads as though it were your data.
Leave a label in its place:
[PATIENT],[AGE],[DRUG],[DATE]. It keeps the sentence intact, reads just as well, and makes clear what was removed.And use the same label for the same person throughout. Put
[PATIENT]in one place and[THE LADY]in another and the model may take them for two different people and answer about two cases. -
And now yes: paste it with the task in front
Put what you want first, then the text. The other way round the model reads four paragraphs without knowing why and starts summarising on its own.
Act as a community pharmacist. I am giving you a report from which I have already removed personal data; you will see labels like [PATIENT] or [ID]. Do not try to work out who it is. Tell me, as a short list: 1) which medicines appear and what each is for, 2) what you would ask this person at the counter, 3) what is missing from the report to answer them properly. If you are missing data, say so instead of assuming it. --- paste the cleaned text here -
Save your cleaning template
You are going to do this a hundred times. Save the five-deletion table and the prompt above in a note on your phone. The first time costs five minutes; from the third it is twenty seconds, and that is the difference between doing it always and doing it when you remember.
One detail you will thank yourself for in six months: you can number the labels.
[PATIENT 1]and[PATIENT 2]let you discuss two people in the same conversation without them merging.
A full example, and what is wrong with it
A real discharge summary — altered, but shaped like the ones that arrive — and its cleaned version alongside. The interesting part is not what gets removed: it is what cannot be removed without wrecking the question.
What arrives: "Patient Mary L. G., 78, record no. 4471982, of 14 High Street. Admitted 3 Feb for decompensated heart failure. History: AF on warfarin, hypertension, T2DM. On discharge: anticoagulant per haematology (Dr Ferrer, County Hospital), furosemide 40 mg/24 h, enalapril 20 mg/12 h, metformin 850/12 h. Review in clinic 28 Feb."
What you paste: "78-year-old woman. Admitted for decompensated heart failure. History: anticoagulated atrial fibrillation, hypertension, type 2 diabetes. Discharge medication: anticoagulant per protocol, furosemide 40 mg/24 h, enalapril 20 mg/12 h, metformin 850 mg/12 h."
And now, what is wrong with that reply:
- The AGE stays and the name goes. Seventy-eight changes the answer — renal function, falls risk, interactions — and identifies nobody. The name changes no answer and does identify. That is this whole case in one line.
- The admission DATE goes, the "on discharge" stays. 3 February, crossed with "78-year-old woman" and your neighbourhood, points at a person. That it is discharge medication does matter for the question. If you need the timeline, write it relative: "two weeks ago".
- The DOCTOR and the HOSPITAL go, harmless as they look. They are not patient data, but they narrow the universe: a 78-year-old at County Hospital under Dr Ferrer is a very short list. And to know whether the anticoagulant interacts with something, who prescribed it is completely irrelevant.
- Brand → active ingredient, and it is not cosmetic. The brand identifies nobody, so it could stay. It is swapped for the active ingredient for a different reason: the model answers better, because that is how nearly everything it has read is written. It is the only substitution here made for quality rather than privacy.
The cleaned text is 40% shorter and answers exactly the same question. That is the proof that what was removed was not needed: if the question still makes sense without it, it was surplus.
And the other way round, which is the error on the other side: if removing something stops the question making sense, do not delete it — generalise it. "14 High Street" is not deleted, it becomes "lives alone" if that was what mattered.