Part 2 | Putting it to work
Chapter 10. Ten prompt techniques
Key points
- Knowing the main techniques makes it easier to adapt
- Mapping them onto therapy work is what makes them practical
- The feel for deepening a question matters more than the "correct" operation
Prompting is the design of a question
How you ask changes the quality of what comes back. The instruction you give an AI is called a prompt. This chapter covers ten representative techniques that adapt well to therapy work.
1. Zero-shot prompting
The most basic use. Ask directly, with no examples.
"Explain low back pain." "Summarise this paper." "Suggest a home exercise programme."
Fine to begin with. The weakness is that it tends towards generalities.
2. Few-shot prompting
Show a few examples first, then ask.
"Write the case report in the following form: e.g. background → assessment → intervention → outcome"
Useful when you want a consistent voice or format: standardising SOAP notes, templating case reports.
3. Chain of thought
Instruct it to
"think step by step, in stages"
Like clinical reasoning, have it work through
- gathering information
- forming hypotheses
- testing
- planning the intervention
in order, and you see not just the conclusion but the reasoning process.
4. Generated knowledge prompting
Instruct it to
"first lay out the relevant knowledge, and then reach a conclusion on that basis"
For example:
"Lay out the psychosocial factors relevant to chronic low back pain. Then, on that basis, propose an intervention."
This separates assembling the premises from reasoning. Structurally, it is almost identical to clinical reasoning.
5. Proactive prompting
Let the AI ask you things.
"If information is missing, ask me for it."
Say that first, and it will come back with "How old?" "Activity level?" "How long has the pain lasted?" Useful for practising history-taking, for improving how you work, and for career conversations.
6. Score anchoring
Give it numerical footing:
- pain NRS
- FIM
- MMSE / MoCA
- Barthel Index
- gait speed
and organise around "what does this score imply?" For example:
"Consider support for an older man living alone, MoCA 20, with reduced attention."
Anchoring to assessment measures keeps proposals from floating free.
7. Virtual parameters
Set hypothetical conditions.
- "What if he attended half as often?"
- "What if she lived with family?"
- "What if they were rushing back to competition?"
Varying the conditions widens clinical reasoning and case study work. Easy to adapt for teaching and training.
8. Persona prompting
The same family as the role setting in Chapter 7.
"You are ◯◯."
placed first, to shift the standpoint. A warm senior PT and a strict supervisor give quite different feedback.
9. Self-critique
Instruct it:
"Now criticise your own answer." "List three weaknesses or things to improve."
The AI reviews itself and produces something sharper. Useful for polishing abstracts and study-group material.
10. Structured output
Specify the shape of the output:
"Produce the output under these headings:
- overview
- key points
- cautions
- sources
- likely questions"
Save it as a template and you get the same shape back every time.
Therapists are a profession that makes questions
This may look complicated, but at bottom it is about
how you frame a question
A therapist assesses, hypothesises, organises information and interprets. That is
the work of making questions
which is exactly why prompt design sits so well with a therapist's thinking.
A prompt to try (combined)
You are an experienced attending physician in rehabilitation medicine.
[Request] Work through the case below in the following order.
- First, lay out the three main clinical problems.
- For each, give up to three possible causal hypotheses.
- For each hypothesis, propose assessments that would test it.
- Finally, narrow to a single intervention to prioritise now, and say why.
- When you are done, list three weaknesses in your own answer.
[Case] (paste the anonymised case here)
Chapter 10 summary
- Ten techniques will carry you a long way
- Combine role, examples, reasoning process and self-critique
- Think about designing the question, not about the "correct" operation