In the Beginning: There Was Prompt Engineering: Part 1

An Educator's Overview of Prompt Engineering

After 10 long months of waiting, OpenAI made its first direct communication to educators this past Thursday. After issuing a brief overview of ChatGPT’s biases and “hallucinations,” the company offered up 4 ready-made prompts for instructors to experiment with in their courses. At the same time, educator-oriented social media is awash with advertisements for “the best AI _____ prompts for teachers.” Needless to say, the age of prompt engineering has arrived. In this post, I hope to offer you a little guidance as you begin to experiment with the development of prompts to insert into your LLM of choice in order to generate different kinds of text for your classrooms and other settings.

Let it be known that many others have already forged this path ahead of me. In particular, I want to give a big shout out to Lance Cummings at Cyborgs Writing for steadily making gains in the art/science of prompt engineering since the launch of ChatGPT. My intention here is not to reinvent the wheel, but primarily to bring techniques showcased on resources like Lance’s Substack to my readers to assist them in their initial efforts in making LLMs work for them. It is my sense that the educational community—at least on the educator-side of the equation—is presently approaching something of a sea-change. The new normal will soon be everyday use of LLMs to extend the individual reach of the teacher. As to the student-side of the equation, I still believe the conversation is stagnating somewhat. But more on that later…

---

So what is prompt engineering?

Prompt engineering is the development, design, and implement of particular textual inputs for the purpose of generating specific kinds of outputs from large language models. Prompt engineering dates to the time between humans first could interact with machines using natural language. In the early days, “prompts” were limited to short phrases and simple sentences. With today’s advances in AI, users can now chat with LLMs in extended sentences and multi-step paragraphs in order to create highly complicated products that include essays, novels, websites, etc. Amanda Bickerstaff of Education for AI has accurately described prompt engineering as “coding” LLMs using natural language. Some say that this aspect of LLMs will be the real revolution at the heart of the new AI: a whole world now peopled by competent coders. Just imagine!

---

1. Prompt engineering is a mindset and a skill.

Prompt engineering is an open field of study that everyone can participate in given the right mindset, a time set aside for experimentation, and a little bit of patience. The mindset required is that of an explorer or adventurer. Imagine you are entering an intricate labyrinth, but the walls are only visible from certain angles. The words you enter in and the responses you receive shine light on the walls—the implicit rules and constraints—of the large language model. If the user enters only a small number of prompts, the barest outline of the maze emerges on the screen. But as the user experiments with many different patterns and their iterations, the labyrinth’s design grows more steadily more discernible, allowing for easier navigation and eventual passage. While AI-trainers and even Substacks like this one may promise you the most direct route through the labyrinth, nothing can replace experimental mindset and actual time in the maze. So grab a snack and a flashlight and settle in. Learning curve is sharp, and I bet you will have a lot of fun along the way!

---

2. When prompt engineering, divide a complicated prompt into discrete tasks.

When ChatGPT 3.0 became available, many educators noticed that the LLM faltered when asked to do multi-step processes. With the development of GPT 3.5 and 4.0, OpenAI has vastly improved their LLM’s ability to handle more complicated, multi-staged prompts. That said, researchers and experienced prompt engineers like Riley Goodside at Scale AI still agree that users retain the most control over the individual steps of complicated prompts by “coding” them discretely. Early versions of multi-step prompting took the form of simple unmarked lists (1., 2, 3,…), but more recently, prompt engineers have found that specifying the nature of the different kinds of labor within the larger list helps get better overall results. As I review the format of these new prompts online (check out’s Education for AI’s Prompt Library for more examples), I am reminded of my days teaching rhetorical analysis in first- and second-year writing courses as well as my doctoral work in rhetorical theory.

Conceptually, the prompt describes or, more accurately, concretizes a “rhetorical situation.” Hey, Purdue OWL, what is a “rhetorical situation”? Purdue OWL: “Anytime one ‘person’ using some sort of communication to modify the perspective of at least one other ‘person.’” (I added the quotes in hopes of provoking a philosophical debate.) Anyway, most rhetorical situations can be subdivided into different roles and resources—speaker, purposes, styles, tasks, contexts, and audiences—and these different sub-divisions become very useful when it comes to creating multi-step prompts.

For instance, let’s say you are using a LLM to create an activity for your 11th grade English class. Then, begin your prompt by precisely specifying the audience of the intended textual product. In many ways, the audience is the most important part of the rhetorical situation:

[Audience] 11th grade English class

Then, add some information about the context. In most representations of the rhetorical situation, the context is a circle that contains the speaker, message, and audience. The more specifics you can input the better:

[Context] Near the beginning of a unit on Civil Disobedience focused on the writings of Henry Thoreau, Ralph Waldo Emerson, Mahatma Gandhi, and Martin Luther King, Jr.

Next, Specify the task. Note that the user is not sending the message after each specification of role or resource. The user is compiling a long bracketed list of discrete operations. Here, task and message overlap in an interesting and productive way. Many rhetorical theorists conceptualize “message” as a directional, communicative action and thus, “task” seems a fitting way to extension of that concept.

[Task] Write four 500 word paragraphs. In each paragraph, describe a different everyday scenario in a person might experiences a challenge to their basic civil rights. Leave each story open-ended and without resolution.

If you would like, you can tweak the narrative voice a bit.

[Style] Highly descriptive third-person narrative voice, some interior monologue

Once you have compiled all of your individual elements, you can send the message.

If you like the results, you can easily get an individualized scenario for every student in class. Here, the follow-up assignment could be:

After reading introductions to and selections from about Thoreau, Emerson, Gandhi, and King. Jr., read your scenario closely. Then, write a response to your scenario that focuses on the distinctive action plan of one of the core thinkers of this unit. Cite one quote in your response. Feel free to imagine the conclusion to the scenario depicted in your individualized reading.

---

3. When prompt engineering, provide copious details about each discrete task.

We begin to see the value of providing more rather than less detail in the above exercise. This particularly become apparent when you use a LLM to generate a quiz, test, or rubric. Left to its own devices, a LLM will provide you the most generic instructional tools. And then you will have to go many, many rounds with it to get to an instrument that suits your purpose. Instead, start with the end-goal already mapped out. Yes, my friends, backwards design should continue to be an educator’s mantra… at least some of the time.

So how do you design a good quiz using a LLM?

There are many ways to proceed depending upon your purpose and your subject matter. I personally think LLMs do a great job creating the shell of multiple choice reading quizzes.

I’d recommend creating one question at a time.

[Text] “Our parents had decided to put an end to their calamitous marriage, and Father shipped us home to his mother. A porter had been charged with our welfare—he got off the train the next day in Arizona—and our tickets were pinned to my brother’s inside coat pocket.” Maya Angelou, I Know Why the Caged Bird Sings

[Setting] 10th Grade Honors English Class near the end of the 2nd Semester

[Style] Quiz language, clear and direct

[Task] Write a reading comprehension question that focuses on the word “calamitous” and the abrupt way the parents put the children on the train. This question should have 4 answers. 1 is obviously wrong. 2 are nearly correct. 1 is actually correct.

Needless to say, one may not be saving any time using a LLM to create quiz questions. The real benefit here is that a teacher can use a LLM to break out of their own implicit rhythms and insert a little randomness into their assessment content.

---

I think I am going to cut this post off at this point, and finish this overview in a second part. There is a lot to absorb and experiment in the above sections. What comes next are more advanced prompt engineering techniques: in-context learning and temperature adjustment. Please send on questions and comments as you experiment. Post sample prompts if you are willing to share. We are living in exciting times, and the more we lean into “group think” in the face of entities who are building up proprietary databases of such materials the better. Big shout out to Lance Cummings and Amanda Bickerstaff for leading the way by sharing out so many good materials. I will post links below.

One final note about the student side of the equation: Instruction on prompt engineering would fit nicely into any course focused on rhetorical and composition. Here, content could be modified for middle school all the way up to first-year and second-year writing. Through designing their own prompts, students get first-hand experience thinking about the rhetorical situation. Who will be the audience for the task/message generated by the LLM? What tone, voice, purpose, and type of narration should be used in the message/task? What discrete steps need to be carried out in order to accomplish the message/task? Essentially, the LLM becomes a rhetorical laboratory or playground.

Thanks for reading this week’s edition of Educating AI!

Nick Potkalitsky, Ph.D

Check out this additional resources:

1. Riley Goodside, expert Prompt Engineer, tries to break ChatGPT with his tricky prompts!

2. Lance Cummings, Cyborgs Writing, “How I Am Using Prompt AI Prompt Frameworks to Drive Deeper Learning”

3. Lance Cummings, Cyborgs Writing, “Why You Shouldn’t Be Writing a New Prompt Every Time.”

4. Oguz A. Acar, Harvard Business Review, “Prompt Engineering Isn’t the Future”

5. Amanda Bickerstaff, Education for AI, Prompt Library