13.12: Getting the most out of AI (prompting)- Describe what you want and keep trying

A chatbot is such an open-ended thing. The possibilities are endless; you can use any words you want to ask it for help. So where do you start? What are the best techniques? And how long will it take to learn them? Will it be too difficult if we’re not coders?

Here’s some good news: in my opinion, there’s no special technical knowledge needed to use chatbots well. Even though some are referring to prompting skills as “prompt engineering,” there is no evidence that “engineering” of any kind needs to be involved. Reading, writing, and critical thinking are the main skills needed.

You may see complicated formulas with acronyms for the kinds of things to include in your prompts, and there’s nothing wrong with those. But they’re absolutely not required. Very simple prompts that just say what you want in the words you would naturally use may serve you just as well.

I’ve come to this conclusion through my own practice, bolstered by the writing of Ethan Mollick, a well-known voice on AI in education and business, author of the New York Times bestseller Cointelligence: Living and Working with AI and the newsletter One Useful Thing. He reviews current research into prompt engineering and spends countless hours experimenting with the latest models. Mollick advises, “Don't aim for perfection - just start somewhere and learn as you go” (“Getting started with AI: Good enough prompting”).

Be specific

Think of more ways to describe what you are looking for. How can you be more specific? Below are some examples of ways you can get more specific with a chatbot. Again, none of these are required, so don’t let this list slow you down as you prompt!

• Context: What’s the project you are working on? For example, you could specify that you are preparing for a test on X or developing a scholarship application for a nursing program. You can copy and paste from any description of the test or application. The context might include the genre you are writing in, like a business memo, a medical research paper, or a political speech.
• Role: Is there a role you want the chatbot to play? If you could choose any kind of human to help you, what kind of person would you choose? You can ask the chatbot to (try) to play that role. Some possibilities: expert tutor, study skills coach, scientist, editor, curious skeptic, devil’s advocate. Notice that some of these describe a professional role (tutor) and some just describe an attitude (curious).
• Style: Is there a particular style you prefer or a style that’s expected? For example, when I was working on my professional bio, I told Claude, “I don’t want it to sound like I am full of hot air. I don’t like professional bios that seem generic and exaggerated like “innovative thought leader.’” You can describe the style by giving the name of a writer you want it to imitate or by throwing in adjectives like “academic,” “conversational,” or “simple but elegant.” One resource you can read and even attach to your chat session is the chapter Style: Shaping Our Sentences from my free and open textbook How Arguments Work: A Guide to Writing and Analyzing Texts in College.
• Tone: Style and tone overlap, but tone often refers to the emotional quality you are going for. For example, “optimistic,” “earnest,” “reverent,” “scathing,” or “melancholy.” For more, see Tone from my free and open textbook How Arguments Work.
• Examples: Just as you might with a person, give the chatbot an example of the kind of thing you are looking for. You can paste in an extended sample if you have one, or you can describe the example briefly or allude to something well-known. For example, you might say “I’m looking for constructive feedback on… Here’s an example of what I consider helpful feedback…”

Ask it to ask you questions about what you want

Sometimes we don’t have a clear idea of what we want or it feels overwhelming to try to articulate it. In those cases, one approach is to ask the chatbot to ask us questions about what we’re looking for. For example, when I wanted a sample argument to give my students to analyze for their final exam, I said “I need to create three new sample arguments with many of the same qualities as the samples below but with different themes. I would like two of the arguments to be about the future of AI in writing instruction. Ask me questions to help you develop the sample arguments.”

Get AI to take it step by step

Some have called this “chain of thought prompting.” (It’s kind of fun to have a technical term, but remember, people are just experimenting with these bots and making up catchphrases.) This can be as simple as literally telling the chatbot: “take it step by step.” Or you may get better results by telling the chatbot what the steps are. What steps would a human need to take to get to the answer? For example, if you want ideas on how you might responding to a letter from your health insurance company, you might give it the following chain of instructions:

1. Start by asking the chatbot to summarize main points from the letter.
2. Then tell it to consider what kinds of further information it needs,
3. Ask it to browse the internet to find credible sources that supply that information
4. Ask it to analyze those sources and summarize how they help us interpret the letter and respond to it.
5. Finally, ask it for a list of five possible strategic responses.

Prompting the system step by step provides you an opportunity to catch problems in a chatbot’s approach and correct them as you go. In the above example, if it misrepresents a main point from the letter, you want to catch that before you let it go on to come up with strategies for responding.

Push the AI to do better

Don’t judge a chatbot by its first response. Chatbots won’t tell you the same thing every time. And they don’t get their feelings hurt. So if you don’t like something about what the chatbot gives you, good! Tell it. It may be able to give you something better.

This is so simple, but somehow it isn’t intuitive for most people when we first interact with a bot. We might be tempted to turn away and later tell a friend “it got X wrong” or “it’s style is too bland” or “its answers are simplistic.” Instead, we could tell it, “That’s wrong because… Try again.” Or “Give me another version that is more nuanced.”

If you can’t put your finger on what’s wrong, then you have some key information that might help you get a better result. Just tell it what’s wrong. Ask it for another version that corrects the problem.

Even if you are not sure what is bothering you about what it gave you, if you are underwhelmed, you can ask it to give you another, more insightful response. Sometimes I just say “Please try again.”

Lean in to your reading and writing skills and keep building expertise

Many people are wondering how best to prepare for future workplace uses of AI. One thing we can be sure of is that you need basic understanding of your field and critical thinking skills to work with AI. To prompt well, you need to know what you are looking for and what sounds good but isn’t right. To know that, you need subject matter expertise.

That means what you’ve already been learning in school and what teachers already know how to teach will help you use AI. Writing and reading practice and understanding of a field will all help you get more out of AI. So will awareness of your own learning and thinking processes.

Here’s another way to think of it: the things you’re learning without AI will help you to use AI later on. All of your studies prepare you for capable prompting and iteration down the line.

Be playful and curious

Working with a chatbot is so open-ended–you’re in the sandbox, and the possibilities are endless. There’s not one way to proceed. One way to think about it is as a play space. When you’re stuck on a task or on how to get help, enjoy the freedom you have to try lots of different things, complain about what the AI is giving you, boss it around, try creative or strange prompts that strike you intuitively.

We can learn a lot about language models by trying weird things with them. For example, Ethan Mollick gave the Claude chatbot the text of a book and asked the bot to “remove the squid.” In another experiment, he kept prompting it with the phrase “garlic bread.” The results are entertaining, but the chat sessions also show how we can think outside the box when prompting, how the language model tries to find an existing pattern to match these odd prompts, and how inconsistent its responses are.

A playful approach to prompting is not just more fun; it means flexibility, curiosity, and openness, which are likely to lead to better results.

Further reading

• For a basic introduction with screenshots, see How to Prompt AI Chatbots by Joel Gladd and Liza Long
• For a quick list of prompting tips, see Getting started with prompts for text-based Generative AI tools from Harvard University Information Technology
• For a substantive, friendly discussion of what to try and what not to worry about, see Getting started with AI: Good enough prompting by Ethan Mollick
• For an approach to prompting that involves developing an imagined “relationship” with AI, see Ethan Mollick’s On the necessity of a sin: Why treating AI like a person is the future
• For a more technical look at prompt engineering, see Anthropic’s prompt engineering overview and prompt library
• For prompt engineering mini-courses of one hour or more, see LearnPrompting.org
• For a playful, humorous approach to AI systems, see Janelle Shane’s AI Weirdness blog and her book You Look Like a Thing and I Love You.