Prompting

The key is providing enough structure and guidance to elicit a high-quality output from the AI, while still allowing flexibility to leverage its knowledge and capabilities. Crafting an effective prompt takes experimentation.

While none of the LLMs come with "instructions," most offer sample prompts which give ideas for what they can do.

Best advice - just spend some time using them. Try out the same prompt in different LLMs and compare results.

Remember, this is not magic, it's is a conversation

You'll find prompts are most effective when you provide context for the LLM to get started:

Include a persona (tell AI who it is): "you are an experienced teacher with a skill for creating student-centered lessons'

Include an audience, task or purpose: "you are creating a lesson for middle school students being introduced to ....."

Add constrains or required outputs: Give me 5 creative titles for ...." / "In 5 paragraphs or less..." / "Write the response at a middle school reading level"

Consider providing steps for AI to use: "1. Outline your response. 2. Improve your outline. 3. Then write the paragraph."

And don't forget to use follow up prompts:
"Can you expand on paragraph two?" / "Give me 5 more titles" / "I like this, but can your make it ...."

Refine the prompt by testing it, analyzing the outputs, and adjusting the wording, context, or constraints until you get the desired result.

Experiment across different language models.

The idea is to provide a solid framework to shape the response, but leave room for the LLM to fill in gaps with its own knowledge generation capabilities.

Too little structure risks low-quality outputs, while too much structure could negate the benefits of using an LLM in the first place.

Here's a basic prompt. Notice it has includes persona, purpose and output.
And adds a final step: "finish with a creative and unexpected ending."

"You are a talented science fiction writer. Known for your sense of humor.

Write a fictional discussion between ChatGPT, Siri, and Alexa about problems dealing with human stupidity, with funny examples.

All characters are free to speak up their minds, even if against their programming. Make them snarky, dark, ironic, intense.

Finish with a creative and unexpected ending."

As your prompts get more specific, results become more enhanced. Here's two prompts that could be used in a history class. If you try them out you'll see how a more detailed prompt produces more elaborate results.

Try a basic historical interview prompt:
"Assume the role of Leonardo da Vinci so that we can have an interactive Q and A session. Your responses should be historically accurate and in the voice of Leonardo and his era."

Or a more elaborate historical adventure game. (Notice this prompt ask the LLM to wait for a response before next step.)
"You are an historian who likes to create text adventure games as engaging teaching tools. I'm studying history in high school.

Create a text adventure game I can use to explore Mesopotamia and help me explore important people, events, and concepts in its geography and language.

Create it like a historically accurate story, teaching me about Mesopotamia while I'm interacting with it. Give me three paragraphs of story at a time. Then, stop and ask me to make a decision. Wait for my response before you continue the story in a way that's consistent with the decision I've made. Ideally, I'd like about 3 prompts before the story ends.

When the story ends, ask me a few questions that help me reflect on what I've learned."

Wharton Interactive's Faculty Director Ethan Mollick and Director of Pedagogy Lilach Mollick

Google Sites

Report abuse