How to approach prompting of AI chatbots

Experiment, experiment, experiment

Getting AI-enhanced chatbots to provide the information and format you want takes experimentation and patience. There is no secret strategy. It’s like a web search or a library search in that regard. The best searches go through many iterations and use an evolving strategy. The best prompts for chatbots likewise take repeated attempts.

This document focuses on prompts for extracting information from large language models like ChatGPT, Bing Chat, Bard, or Claude. Other pages on the CTE site address such things as the use of generative AI in teaching and learning (including the meaning of common AI-related terms), academic integrity and AI, and the types of AI tools instructors and students might consider in their work.

Prompting deserves specific attention, though, because generative AI chatbots require guidance in the form of natural language. That’s where prompting comes in. Prompting is really just a matter of providing clear directions for what you want the bot to do. The bot can’t read your mind, and it has no context for the type of information you seek. You have to guide it with clear, specific commands. Some bots have access to the internet and can supplement the information they provide through search. Others draw only from the data they were trained on, which is often a year or two old.

As large language models evolve, prompting may need to change, too. For now, though, most chatbots respond in similar ways, meaning that prompts for one will generally work with others. The training and algorithms of the bots vary, though, so the answers you receive will vary, even with the same prompts.

Prompting basics

There is no single way to prompt chatbots, but following a few simple steps will help. Think of the process as a series of branches. Each branch allows you to provide additional direction, focus, or depth. You can also back up at any time and pursue an idea in another branch. You don’t have to provide all the branches at the beginning. Rather, you build and remake those branches as you evaluate the bot’s output.

Here’s an example prompt. The bulleted list that follows refers to this prompt and explains the rationale behind it.

You are an expert in nursing and nursing education. Create a 500-word case study about the use of artificial intelligence in nursing. Provide discussion questions related the case, and create an instructor guide for using the case in class.

Use natural language

The large language models that allow chatbots to work are trained on natural language text and programmed to interpret natural language prompts. The best prompts use clear, concise directions in a way you might explain to a coworker who will complete a project. (One professor equated the process to providing directions to a new intern.) Include any keywords that offer direction or provide context.

Identify the role you want the bot to take

In the example above, that is the first sentence: You are an expert in nursing and nursing education. You could use the same approach in any discipline. For example: Assume the role of a researcher in cultural history or Act as an expert in theoretical physics. The phrasing doesn’t really matter as long as you provide clear context and give the bot a role or perspective to limit the scope of its replies. Don’t worry about getting that perfect. You can always adapt it or change it later.

Provide specific requests

In the nursing example, we tell the bot what we want (a case study, discussion questions, and an instructor guide), and how extensive the output should be (500 words). All generative AI models have limits on the amount of material they can provide in a single response. It’s often best to start small and ask the bot to provide an outline or questions (although some users prefer to pack an initial prompt with directions and focus). You can ask it to elaborate by using follow-up prompts. Other examples:

• Create a summary paragraph and five bullet points.
• Create three sample problems. Show your work and provide explanations about how you arrived at the answers.
• Create a study guide that highlights the six most important topics and provides study questions for each topic.

Drill down with follow-up prompts

Don’t expect a chatbot to provide a perfect response after your first prompt. Instead, examine the initial response for weaknesses and for areas that need refinement and elaboration. By asking for more specifics in follow-up prompts, you can add depth to the results. In some cases, you might want to ask for a specific number of examples. In others, you might want to ask for specific sources or for expansion or elaboration on a particular element. In the nursing example, ChatGPT created a one-paragraph scenario for the case study. A follow-up prompt asked it to elaborate:

Expand the case scenario to 300 words. Provide specific information and scenarios nurses might face.

That helped flesh out some details of the case, with ChatGPT creating a character called Mary, who is a registered nurse in a busy hospital. (Note that ChatGPT defaulted to an example of a white woman as a nurse. It is just one example of the biases that often show up in material that generative AI produces.) We can continue with follow-up questions to fill in details of the scenario:

• What are five challenges Mary is likely to face in working with AI as a nurse?
• What type of training should Mary seek out?
• What are the most common AI tools nurses use in hospitals today? How can Mary become more adept at using those before she leaves nursing school?
• Now provide five specific examples of how Mary can use AI to improve her workflow. Include tools she can use and guides she can consult.

Provide sources to draw on

By having a chatbot read a particular source or document, you can direct it to follow a specific style or draw on specific information. Some bots allow you to upload documents. Nearly all allow you to provide a URL of a webpage or PDF and ask the bot to read it. You can also paste in information into the prompting box and ask the bot to read it. The first example below asks the bot to process the information, pause, and prepare for additional questions. The second asks the bot to process the information and create a summary:

• Read the following document and tell me when you are ready to answer questions about it.
• Read the page at this URL, summarize the information, and provide a bulleted list of the five most important aspects.

Provide information about audience and tone

Chatbots will usually make assumptions about audience based on the type of prompts you provide and the sources you ask them to draw on. It is helpful, though, if your prompts include information about your audience, and directions about the tone or style of language you want. You can also direct the bot to simplify complex information or to take a particular tone. For example:

• Answer in language appropriate for a general audience and avoid jargon.
• Respond in language appropriate for researchers in social psychology.
• Explain the concept as you would to a 10-year-old.
• Provide your explanation in the form of an employee memo for a law firm.
• Take a lighthearted tone in your response, but avoid jokes or cheeky phrasing.

Prompting as critical thinking

Prompting chatbots draws on such skills as researching, setting goals, understanding audience, understanding disciplinary language, interviewing, and evaluating information for strengths and weaknesses. It helps students develop skills in critical thinking and information literacy.  

Prompting helps students develop metacognitive skills, which are crucial to learning. Prompting forces them to focus their thoughts and to consider broader aspects of topics and outcomes of projects. It is certainly possible to engage in mindless discussion with a chatbot, but extracting solid information and helpful direction requires well-considered input.

The advice here offers is simply intended to help you get started. It is based on many hours of experimentation and on advice offered by other users of generative AI. As you and your students work with chatbots, you will find approaches that work best in your discipline. Experimenting and sharing will help your students develop better prompting skills.