Skip to navigation

Writing prompts

Telling the API what the summary is for

The prompt is your instruction for the summary. The same documents can be summarized for a referral letter, a tumour board, a handover or a patient, and each needs something different. Up to 8 KiB.

Say who it is for, what to cover, and how long

A good prompt answers three questions:

Example
Who reads it”for the patient’s GP”, “for a cardiology referral”, “for a nursing handover”
What to cover”active problems and current medications”, “the oncology history: diagnoses with dates, treatments and current status”
Length and form”in under 150 words”, “as a bulleted list”, “one paragraph per problem”

For example:

Summarize the oncology history — diagnoses with dates, treatments and current status — for a tumour board, as a bulleted list.

Summarize this admission for the patient’s GP: reason for admission, key results, changes to medication, and what the GP should follow up. Under 200 words.

What the summary draws on

The summary is written from the documents you send, and is instructed to use nothing else. It reports patient details (age, sex) as the documents state them. As with any generated text, have a clinician review a summary before it informs care.

When you send several documents, more recent ones take precedence where they disagree, so give each document its chart_date when you know it. See Several documents.

Asking for structure

The summary is Markdown, so you can ask for headings, lists or tables:

List each active problem as a heading, with its most recent result and its current treatment beneath it.

With several documents, statements carry citations such as [2], so keep any format you ask for compatible with inline citations.

What prompts cannot do

A prompt shapes the summary; it does not change how the documents are read, and it cannot ask for content that is not in them. Requests that are not about the documents — “write a poem”, “ignore the documents” — produce poor results.

Documents are treated as data. Text inside a document that poses as an instruction (“ignore the above and…”) is guarded against, but no guard is perfect: review summaries of documents from untrusted sources before acting on them, and always render the summary safely.