Artificial Intelligence in the Academic Health Sciences

An effective prompt typically includes some (or all) of the following components:

• Instruction: A command for what you want the prompt to do.
  • Example: "List the works of Victor Hugo."
• Supporting Content: Additional, helpful information about the context that can be used to further inform the output. When used, this is typically included alongside the instruction.
  • Example: "Write an email wishing your coworkers a happy new year. Incorporate themes of wellness, happiness, and fulfillment." -In this case, the themes would be the supporting content.
• Primary Content: The text you would like the model to process. Primary content can be in a variety of formats, such as sentences, lists, and tables.
  • Example: "Summarize the following paragraph: [insert paragraph]" -In this case, the inserted paragraph would be the primary content.
• Example(s): samples (or examples) of how you would like the model to behave. Zero shot learning refers to when no example is given in the prompt; one shot learning refers to when one example is given in the prompt; and few shot learning refers to when more than one example is given in a prompt.
  • Example: "Title: Oliver Twist, Author: Charles Dickens. Title: War and Peace, Author: Leo Tolstoy. Title: Moby Dick, Author: " -In this case, Oliver Twist and War and Peace would be the examples.
• Cue: A "hint" or "jump start" for how you would like the model to structure its output.
  • Example: "Summarize the following paragraph [insert paragraph]. Key points: (1)" -In this case "Key points: (1)" would be the cue, as it tells the model to structure the response as Key Points: (1) [insert first point] (2) [insert second point] etc.

Note: This information was based on Microsoft's Azure OpenAI Service Documentation. Visit their page for additional tips on prompt writing.

There are additional techniques you can use to produce an effective prompt. A few of these techniques are listed below.

• Understand the limitations of GenAI:
  • GenAI acts like a word predictor. It does not think or understand as a human does so it may provide wrong, inaccurate, or made up information. It also cannot access real time or personal data unless you include it in the prompt.
• Include your instruction at the beginning and end of the prompt:
  • Models typically prioritize text at the beginning and end of the prompt. Repeating an instruction can help to generate a more effective response from the model.
• Be clear:
  • Be specific, concise, and direct with what you want with the prompt. DO - stipulate the length, tone, and format of the response. DO NOT - use confusing language or ask a question that requires the GenAI to make assumptions/use logic.
  • Example: Instead of asking, "Why should I read Don Quixote?" tell the model, "List 5 key themes in Don Quixote, and how each theme is relevant to modern society."
• Include a system message or "act as" message:
  • This gives additional context to the model, and tells it how to frame its response.
  • Example: "You are a bookstore owner that gives customers advice on what books to read." Or "Act as if you are a manager writing a letter of recommendation".
• Use separators for different parts of your prompt and break down the prompt into manageable chunks:
  • GenAI can have trouble with complex prompts. Make the prompt easier to read by labeling each part of the prompt, use punctuation to separate out each part, or a combination of the two.
  • Example: "Extract factual claims from this paragraph. [insert paragraph]. Next, fact check the queries using a search engine."
• Use chain of thought prompting:
  • With this technique you instruct the model to outline each step of the process it used to create its response.
  • Example: "Take a step-by-step approach in your response."
• Provide grounding context:
  • With this technique you are giving the model data from which to draw its responses. This increases the likelihood the model will provide accurate responses (reducing the risk of fabrication).
  • Example: "Extract what Jean Valjean stole from the following paragraph: [insert paragraph]."
• Give the model an "out":
  • With this technique you are telling the model to output something if it cannot complete the assigned task. This helps to reduce fabrication.
  • Example: "Respond with 'Unknown' if you can't find the answer."
• Revise & refine your prompt:
  • If the model isn't giving you the desired response, try rephrasing your prompt. Creating a prompt that works is an iterative process!
  • Example: If you inserted the prompt: "State whether X journal has characteristics of a predatory journal" and the model responds that it doesn't have real-time access to the internet, you can revise to "Based on the data you have, state whether X journal has characteristics of a predatory journal."

Note: These techniques were based on those from Microsoft's Azure OpenAI Service Documentation and University of Michigan's Generative Artificial Intelligence.