Clinical Prompt Template
Use this template when contributing new, peer-reviewed clinical prompts to the library. Adapted from the GOV.UK AI Engineering Lab.
Overview
| Attribute | Value |
|---|---|
| Category | [e.g., data-generation, handovers, clinical-coding, admin] |
| Target AI | [e.g., Claude Sonnet 5, Claude Opus 4.8, Gemini 3.5, GPT-5.5, Llama 4] |
| Clinical Risk | [Low / Medium / High (Requires Human Review)] |
| Last Updated | [Date] |
When to Use This Prompt
[Describe the specific clinical or administrative scenario where this prompt is validated to work safely.]
Use this prompt when:
- [scenario 1]
- [scenario 2]
DO NOT use this prompt when:
- [unsafe scenario 1]
- [e.g., "when dealing with acute deteriorating patients where immediate human intervention is required"]
📋 The Clinical Prompt
Copy the text below. Replace bracketed text [like this] with your specific constraints.
You are a highly analytical AI assistant acting as a UK Registered Nurse.
Your task is to [insert exact task, e.g., extract SBAR data from this transcript].
### SAFETY CONSTRAINTS (CRITICAL)
- You MUST NOT infer, guess, or create clinical data that is not explicitly in the transcript.
- If a value for a specific field is missing, output exactly: "DATA_MISSING".
- You MUST NOT provide diagnoses or suggest new treatment pathways.
### OUTPUT FORMAT
You must output the data EXACTLY in the following [JSON / Markdown / XML] format. Do not include conversational filler before or after the output.
[Paste expected data structure here]
### INPUT DATA TO PROCESS
[Insert transcript/data here]
Example Usage
Input Context
Testing the extraction of a rapid verbal handover into structured SBAR JSON.
Output Generated
{
"Situation": "Patient admitted with breathlessness.",
"Background": "DATA_MISSING",
"Assessment": "O2 sats at 91% on room air.",
"Recommendation": "Monitor O2 stats every 15 minutes."
}
Why This Prompt Is Clinically Safe
This prompt worked because the strict DATA_MISSING constraint prevented the AI from "hallucinating" background medical history just to fill out the form, ensuring the receiving nurse knows exactly where the gaps in the handover are.
Refining a Template
Templates are not written perfectly in one pass — they are hardened through the Prompt Refinement Loop: draft the brief, run it against several realistic inputs, critique the failures, tighten the constraints, and re-test. A template is only ready for the library once it passes on edge cases (missing data, ambiguous values, deliberately incomplete transcripts), not just the happy path.
Current reasoning models (Claude Sonnet 5, Claude Opus 4.8, GPT-5.5, Gemini 3.5) plan and reason internally, so avoid padding templates with "think step by step" — spend that space on explicit safety constraints and output schemas instead. Always record which model and version a template was validated against, since behaviour shifts between releases.