Core Engineering Principles For Prompting
-
Be Clear and Direct in Task: Provide Context and Persona:
-
Establish the background, domain expertise, and operational role expected of the LLM.
-
Specify Output Format: Define exact structure (e.g., HTML table, JSON, bulleted list, maximum word count).
-
Define Negative Constraints: Specify what the model must not do (e.g., "Do not use medical jargon," "Do not include Protected Health Information").
-
Use Exemplars (Few-Shot Prompting): Provide input-output pairs to guide the model's reasoning pattern.
Meta prompting: ask ChatGPT or other LLM models to generate your prompt and make edit to that prompt.
Other Tips:
Open the voice-over feature to brainstorm different things and convert my thoughts into written format Grammar and Writing check-As an international student in the US, it is still difficult for me to translate my thougths to a professional writing format. I uses grammarly and ChatGPT. Have Gen-AI role-play to brainstorm example: "I am an OTP. You are an early stage dementia client's wife living in a rural farm. what kind of materials about AI would you want to know about dementia care for your husband?" Use Gen-AI as a learning tool to breakdown hard topics and ask me questions to test my knowledge. NotebookLM for fast learning new materials. Claude has helped me learn SPSS coding within 2 hours after finding my whole day on LinkedIn and IBM learning couldn't solved my issue. A study has shown that different Gen-AI presents with different personalities, which I agree. I tend to tell my Gen-AI to be professional, honest, and straightforward when evaluating my goals or identify if I am mislead.
What you want the AI to do?
Avoid ambiguous language.e.g. brainstorm some ideas, create an presentation outline, generate a google doc, generate an excel guideline, and teach me step by step.
-
Provide the clinical or situational background
-
e.g. diagnosis or condition, functional limitations, precautions, and relevant environmental factors
Example Prompt:
"I'm writing a blog post about tips and tricks for ChatGPT prompt engineering for occupational therapy practitioners in rural healthcare working with dementia care. Provide list a few of tips and tricks with some specific prompt examples. My audience consists of expert OTPs who are not AI experts. Use a friendly and approachable tone."
-
Add any sources, frameworks, standards, or documents
-
e.g. research paper, policies, photo, and video
-
Ask the AI to check its own work against clear criteria
-
e.g. Is the response clinically appropriate and client-centered, is it safe and follow the requested format?
-
Refine the output based on gaps, errors, or new information
-
e.g. what needs to be improved, tone change, clarify ideas, adjust to different literacy level, or different format
Bad Prompt:
"Give me dementia interventions"
Better Prompt:
"Suggest OT interventions for an older adult with moderate Alzheimer's disease working on gardening. (Still missing home environment, environmental human support, and rural context)"
Advanced Prompt (Generated by ChatGPT):
"You are an experienced occupational therapist specializing in dementia care.I am completing a home health visit for a woman in her 80s' with moderate Alzheimer's disease who lives with her husband in a rural community 30 minutes away from Saint Louis. She enjoys gardening but has stopped participating because she forgets the steps and becomes frustrated. Her husband wants to help but feels overwhelmed.Suggest five occupation-based interventions that support participation in gardening while promoting safety, independence, and caregiver confidence.
Please include: intervention description, clinical rationale, grading strategies, caregiver cueing technique, slow-cost materials appropriate for rural home care.
Present the information in a table."
Tips: adding a person's routines, interests, environment, and goals can result in better AI response
However, you might need to start a new chat for each individual with no PHI and PII attached!
Reminder: never use AI generated interventions without your clinical reasoning.