Clinically Intelligent is written by Juwon Akinyande in a personal capacity and is not affiliated with, endorsed by, or representative of Barts Health NHS Trust or any other NHS organisation. Content is for general educational purposes only. It does not constitute clinical, legal, or information governance advice. Before applying any guidance to your own practice, consult your Trust information governance lead, your Caldicott Guardian, your line manager, and your professional body.
Good to be back. This one took time to build properly. But here we are. Let's get into it.
You have read the issues. You have the Safety Check. You have seen the tools. The next thing is the part nobody actually teaches. How to talk to these tools so they give you something useful back.
That is what this issue is.
The reason you are getting poor results.
If you have tried AI and the output was rubbish, the tool probably was not the problem. The prompt was.
AI tools cannot guess what you want. They can only deliver what you ask for. The more specific you are, the better the output. The more context you give, the better the tool understands what you actually need. The less you give it, the more it fills in the gaps itself. And when it fills in the gaps itself you end up with something generic, something inaccurate, or something that sounds right but does not actually work for your situation.
I always think about it like this. Imagine you are briefing a junior member of staff on a task they have never done before. You would not just say write me a letter and walk away. You would tell them who it is for, what it needs to cover, what tone to use, what to leave out, how long it should be, and what to do if they are not sure about something. You would give them enough to actually do the job properly.
AI is the same. Except you probably need to be even more specific than you would with a junior colleague. At least the junior colleague has some context from working in the same building as you. The AI has none. It only has what you type.
That is the whole argument of this guide.
What the guide covers.
Three stages. Nine checks total.
Stage one is what every prompt needs. Three checks you run before you hit send. Have you named the role and the clinical context. Have you given enough context for the output to be useful. Have you specified what the output must include and exclude. Answer yes to all three before you proceed.
Stage two is what good prompts do. Three moves that separate prompts that produce usable output from ones that do not. Are you asking the AI to accelerate your thinking or replace it. Are you requiring evidence, sources, or reasoning. Have you told the AI what to do if it does not know something. The clinical judgement stays with you every time.
Stage three is when to stop and try again. Three signals that tell you the prompt is not working. The output is too generic. You cannot verify the claims. Something feels clinically off. A poor output is feedback about the prompt. Not permission to lower your standards.
Each stage has three checks. If a prompt is not producing usable output one of the nine checks is the reason. Find it. Fix it. Run it again.
The guide also has three worked examples showing the same task done with a poor prompt and a strong one. Clinical documentation. Evidence retrieval. Patient facing material. The difference in output quality is the whole argument made visible.
And a one page reference you can keep open while you work. Six elements. Role, context, task, output, constraints, sources. Use it every time.
Download Here: The AI Prompting Guide
How to use it.
Read the three stages once. That is the framework. You do not need to memorise anything.
Come back only when a prompt is not working. One of the nine checks will tell you why. Fix that one thing. Run it again.
The first time you use it pick the task you do most often. A discharge summary. An evidence search. A piece of patient education. Apply the framework. Compare what comes back to what a generic prompt would have produced. That single comparison will tell you more than reading the whole guide twice.
One thing worth saying.
There is no single correct way to prompt. There is a clearer way and a vaguer way. The clearer way works better more often. That is all this is.
The framework in this guide is what I arrived at through practice across twelve issues of testing AI tools in clinical work. Not the only approach. One approach that has earned its place through use. Adapt it, break it, improve it. The discipline is in the doing, not the document.
That is all for Issue 12. Next week, not a tool. Something I keep hearing from colleagues since coming back.
Clinically Intelligent drops every Wednesday. If you are not yet subscribed you can join free at clinicallyintelligent.com.
The AI tools discussed in Clinically Intelligent are consumer products. They have not been independently assessed by the author against DCB0129 or DCB0160 clinical risk management standards, and they may not be approved for clinical use by your employer. Before using any tool described in this newsletter in connection with your clinical practice, you must satisfy yourself that its use is permitted under your Trust information governance policy, your DSP Toolkit obligations, your professional registration requirements, and any applicable contractual terms with your employer. The author accepts no liability for use of any tool or workflow described in this publication. Patient identifiable information must not be entered into any consumer AI tool under any circumstances, irrespective of any guidance contained in this newsletter.