What is Prompt Engineering?

One of the most important phrases to keep in mind when interacting with AI is garbage in equals garbage out. Meaning that the quality and quantity of information you put into a generative AI tool will have a big impact on the quality of content you receive back. This is what we call Prompt Engineering. Prompt Engineering is "the art and science of designing and optimizing prompts to effectively interact with language models" to receive more relevant and accurate responses from the AI tool (OpenAI, 2023). This page is designed to introduce you to both beginning and advanced prompt engineering skills that will help improve the content you get.

Best Practices

NOTE: To see the below content in video form, watch OpenAI's video on Prompt Engineering Best Practices.

According to OpenAI, there are three key ingredients to a good prompt: give clear and specific tasks or goals, give a persona and provide additional knowledge, and clarify audience and specify output format. These three ingredients set the foundations for good prompt engineering, and are easy to implement. Image below is from the prompt engineering best practices video.

3 key ingredients to prompt engineering graphic

To understand these three key ingredients better, and to see more tips in creating successful prompts, check out the tips below. Each are crafted to give more in-depth explanations and give sample prompts of what you could do.

GIVE SPECIFIC TASKS OR GOALS

Basic prompts with few details provide more opportunity for inaccurate and generic responses. The more specific you can be with tasks, goals, and output format, the better your response will be. In addition, think through the objective and what information the AI tool needs in order to complete the task. See the examples below:

Write a story: consider including the type of story you want, the plot, the genre, and tone you want, who it is for, and any other details that you want to include.
Send an email: consider including the topic of the email, the tone the email should sound like, any specific personalizations you want it to give you and what the output should look like.
Summarizing Notes: consider including how many key take-aways you want summarized, how many sentences you want each to be, what the next steps should be, etc.
Analyze Data: consider including how you want the data analyzed, any control factors, and in what format (e.g. table, graph, etc.)

The more details you can provide, the better your prompts will be.

GIVE THE AI TOOL A PERSONA

One of the most effective ways to receive more accurate answers is to give the AI model a persona. This can look like many things; for example, asking it to emulate someone or something, or asking it to respond in a specific tone or with a specific perspective.

One of the easiest ways to do this is to ask the AI tool to act as if it is a type of person, process, or objective. Here are some examples:

If you're wanting it to create a course outline, ask it to pretend it is a higher ed instructor.
If you're wanting a recipe for ingredients in your house, ask it to take the role of a chef.
If you're wanting it to analyze a case study, ask it to take the role of that occupation.
If you're wanting to get Excel formula information, ask it to take the role of an Excel spreadsheet.
If you're wanting to see how it does as a tutor, ask it to take the role of a tutor.

By asking it to take on the role of a persona, this helps to train the model and tailor the conversation to respond in the way you want.

STATE WHAT YOU WANT IT TO DO

One way to ensure you're getting the output you want, is to tell the AI tool what you want in your response, and what you do not want in your response. This can improve your result and save you time by receiving a more desired output, earlier. Some examples of how you can use this idea is:

State what format you want the answer in.
If you just want the answer and do not want to know how to do it, let it know.
If you want AI to tell you when it does not know the answer instead of guessing, let it know.
If you want it to evaluate writing using APA but only in-text citations and not reference pages, let it know.

USE EXAMPLES FOR GUIDANCE

If you want your answer to follow a specific format, give AI an example or reference point to follow. This will take it one step further to ensure you're getting more of a response that you want. We don't recommend having it create similar work to famous people, writers, etc. due to copyright infringement, but instead, include an idea of what you want it to say (and make it your own).

Need to create a lot of story problems or scenarios with varying names, locations, etc. Provide a formula for it to follow, and identify what you want changed in each scenario or problem.
Have an outline of an email or letter of recommendation you want to include, share it with the AI tool and ask it to follow the format.

BE SPECIFIC ON TONE AND AUDIENCE

When prompting AI, be specific about the tone and/or style you want in the output and who the audience is. This will provide greater context that improves the output you receive. Examples:

Tone examples: Professional, friendly, witty, firm, happy, sympathetic, sarcastic
Style examples: Formal, informal, persuasive, descriptive, concise, narrative, inspirational, critical,
Audience examples: Conference participants, alumni, donors, higher ed leadership, faculty, staff, students, etc.

CONVERSE WITH THE AI TOOL

It is very possible that you will not receive the correct response or an output you find helpful the first time. If you don't have a conversation with the AI tool to get the answer you want.

Build on the prompts you already have in the conversation
Ask it to change the answer, add or subtract information
Remember the tips that seemed to work and improve the output to work more efficiently in the future.
Ask it to change the tone, style, audience, or reading/writing level
Ask questions to improve.

TRAIN THE AI TOOL

By interacting with the tool, you have the ability to train the AI tool to be more accurate and effective.

If it gets something wrong or gives you output that isn't useful, let it know or regenerate the ouput.
If it provided great feedback and examples, give it positive reinforcement.
Use the AI tool's feedback options, including a thumbs up or thumbs down icon to help the system identify good and bad responses.

This approach helps the system in the long run and helps your interactions in individual chats.

ASK THE AI TOOL FOR HELP

Generative AI tools are designed to help; if you're stuck on how to create a good prompt or get the answer you want, consider doing the following:

Feed your prompt into the tool and ask it what else it needs in order to help solve your problem or get you the response you need.
Ask the AI tool how you can improve your prompt
Give the AI tool what you need and ask it what all it needs to create the desired output.

Although simple, this is a great, effective use of AI to improve your prompts and receive better output.

Going from Good to Best with Prompt Engineering

It can be difficult to know if you're doing a great job with prompt engineering. Below are some examples of good, better, and best prompts with brief descriptions of why. Go through each one and evaluate where your currently level is at. Later in this module, you'll have the ability to practice your prompt engineering and improve your skills.

GOOD

PROMPT: "Analyze The Meeting Notes"

In this prompt, it gives it a task to complete, but it is vague and does not provide enough detail to get the correct answer the first time around. It is most likely that you will need to probe the AI tool further to get the answer you want.

BETTER

PROMPT: "Analyze the Meeting Notes. Include 3 key takeaways."

In this prompt, it includes a task and additional information on what they need to analyze it for. You'll get a better response, and 3 key takeaways from the meeting but it may not be in the format you want, or you still may need to probe further to get the answer you want.

BEST

PROMPT: "Analyze the Meeting Notes. In a bulleted list, include 3 key takeaways, 2-3 sentence summaries for each key takeaways and provide ideas for next steps, and topics that will need to be discussed again at upcoming meetings."

This prompt provides the most detail, and in addition to giving specific directions, provides the formatting of the output it wants. Another adaptation to this prompt would be to ask AI to "act as a meeting note-taker" or "pretend you're an administrative assistant who took these notes from our meeting" (or any other role there may be).

Formal Prompt Engineering Techniques

Once you're more comfortable with general prompt engineering, you can experiment with several formal Prompt Engineering Techniques. You may even find that many of the ideas we've discussed above, align with more formal terms. Select the individual tabs below to learn about different techniques, including the equivalent in human behavior and examples of what it looks like.

Zero-Shot vs. Few-Shot
Chain of Thought

Zero-Shot

Zero shot is a prompting technique where the AI tool is expected to perform a task with little instruction and no additional training. In this type of prompting, no examples are given. Instead, the AI tool is expected to use the general knowledge it acquired in its training to identify and share the correct answer. This is most often used with general knowledge questions or exploring new tasks to give you a starting point. If relating this to human behavior, this is similar to if your boss asked you to complete a task you've never done before, but did not give you any specific instructions or examples. As a human, you'd have to figure out how to complete the task only using the knowledge you've gained previously. AI tools like ChatGPT do the same, they can only use their previous training to try to answer the question.

It is important to note that zero-shot prompting is most successful in advanced AI models. As AI models continue to improve, the ability to use zero-shot prompting will increase. Below are some simple questions or tasks that you could give ChatGPT or a different AI tool that demonstrates zero-shot prompting.

Example Prompts

Classify the following text as positive, negative, or neutral: "I love the sunny day." Response: positive
Translate the following sentence to French.
What's the largest University in Nebraska? Response: University of Nebraska - Lincoln

Few-Shot

Few shot prompting is a technique that expands on zero-shot. Instead of having to rely on previous data, the model is given several examples of the task needing to be performed. If relating this to human behavior using the scenario above, this is similar to your boss giving you a task and sharing 3 examples from other universities to use as guidance. The examples given to the AI tool help to guide the model, provide context, and show an example of the output and structure it wants. Few shot prompting allows for better responses. Below is an example of few-shot prompting where it can identify the correct answer. In this situation, I am asking the AI tool to tell me if a statement is positive, negative or neutral. By training it with an example of a positive, negative and neutral statement, ChatGPT is then able to correct identify the fourth statement as being negative.

Example

Prompt:
Classify the following statements as positive, negative, or netural

"I love a sunny day" = positive

"I hate green beans" = negative

"I am okay" = neutral

"I can't believe you're saying these things to me ="

Output:
I can't believe you're saying these things to me = negative

Chain of thought (CoT) prompting walks the AI tool through the process of human reasoning to receive an answer. It is best to think about chain of thought as a step-by-step process. Often, chain of thought is combined with few-shot prompting to get better results. By giving it examples of how to answer a question, or complete an equation, it provides it with a "how to" to more accurately get the answer. Watch the video below to see an individual walk through solving two problems with chain of thought prompting. If you have time, I recommend watching the whole video. If you do not, I found the example that starts at 7:13 extremely helpful in seeing the chain of thought prompting at work.