💬 AI Training Tuesday: Let's take a shot at better prompting

What is a prompt?

ChatGPT, Claude, Copilot, Mistral, Perplexity, and all other so-called LLMs (or Large Language Models) start with a prompt.

That simple little blinking cursor has the backing of trillions of bits of information, which we will refer to as knowledge. That knowledge can be used in many ways. Here are just a few business examples:

Bill, I have tried prompting, and it is terrible!

When Apple released the iPhone 4 in 2010, they introduced a leap forward in cellular signal strength by putting the primary antenna system around the frame of the iPhone. Following the phone's release, people could hold it a certain way and watch the bars of cellular strength quickly dissipate.

Steve Jobs's famous response was simple: "Just avoid holding it in that way."

An excellent prompt signals that you want to articulate your expectations; you are probably using it wrong 🫢.

🎯 Let's start with some shots

One of the most important things to know when conversing with an LLM is that it will always attempt to give you a thoughtful response, no matter what you type for a prompt. Therefore, your prompt matters a lot.

For example, if you type "Tell me a story," the LLM will likely respond with an incredibly bland story. That example of "Tell me a story" is what people in AI and machine learning call a zero-shot prompt.

I will share the following three baseline prompt types so you can better understand how the LLM works and receive higher-quality responses. At the end of this article, I will combine everything you learn with a real-world example.

• Zero-shot (ambiguous)
• One-shot (add context)
• Few-shot (articulating your expectations)

🤞 Zero-shot prompts (or probably why you think ChatGPT and other LLMs are terrible)

In this scenario, you create an open-ended prompt without telling it to draw on specific knowledge.

Let's pretend you are a parent. You say, "Go outside." What does that mean to you compared to what the child thinks?

Does "Go outside" mean "Stay in the backyard?" or "Go on a bike ride" or "Mow the lawn?" This is to say most zero-shot prompts follow the adage "garbage in, garbage out."

That is not to say zero-shot prompts don't work. For example, if the parent says, "Do your homework" or "Clean your room." A child likely has enough knowledge of what that means to go ahead and take action without further clarification.

😶‍🌫️ One-shot prompts (let's add a little context)

In this scenario, you prompt the LLM with a single example of the task to perform. This provides guidance and sets an expectation of what you want the result to look like.

You are still the parent in this example, but now you say, "Go outside and play on the swings with your sister."

Notice how, this time, you did not simply say, "Go outside," but instead, you specified what to do. That is an example of a one-shot prompt.

Imagine someone you don't know walking up to you, looking you in the face with no readable expression, and their hands are at their side. They then go on to say, "What's that?" Think about how you would react and respond. You might look behind you, worried that something is happening, or maybe you say, "Oh, that is my coffee from across the street; you should check it out; they make a great macchiato!"

If that same person asks, "What's that on your shoulder?" You can confidently respond, "Oh, that's my pet parrot!" That tiny bit of context meant everything, so a one-shot prompt will almost always be better than a zero-shot prompt.

Few-shot prompts (articulating your expectations)

In this scenario, you provide the LLM with more than one example of what you require in the response.

As a parent, we first gave our child a zero-shot (ambiguous) prompt of "Go outside." We then modified that to add context, like a one-shot prompt: "Go outside and play on the swings with your sister."

Using the few-shot method, we can tell our child, "Go outside and play on the swings with your sister. Don't forget you can also use the slide and play in sandbox. If you use the sandbox, you can create a castle!"

As you can see, adding a little extra clarification to your prompt will help improve the response you receive.

💬 Let's not forget LLMs can chat

In my experience, if you are doing anything of consequence, you rarely get the exact response you want from your prompt the first time. That is why you want to embrace the "chat" part of ChatGPT and other LLMs.

Just for fun, let's kick off the conversation by creating an invitation for a corporate retreat using a few-shot prompt:

Prompt: We just launched our first product and must take it to the next level. To this end, we will draw inspiration from the great outdoors and retreat to a lodge for a long weekend. Create a one-paragraph professional invitation representing having fun but in a business setting.

LLM response: We are thrilled to announce an exciting retreat to celebrate our first product launch and plan our next big steps. Join us for a long weekend at a charming lodge in the great outdoors. This retreat promises a perfect blend of productive sessions, team-building activities, and relaxation amidst nature's beauty. We will brainstorm innovative ideas, strengthen our bonds, and recharge for the challenges ahead. Let's embrace this opportunity to connect, collaborate, and have fun in a professional yet invigorating setting.
We look forward to seeing you there and taking our success to the next level.

Prompt: Whoops! I forgot to mention it will be Thursday through Monday; it is a corporate event, so there will be no family, friends, or pets. Also, mention a link I will include to RSVP.

LLM Response: We are thrilled to announce an exciting retreat to celebrate our first product launch and plan our next big steps. Join us for a corporate event from Thursday through Monday at a charming lodge nestled in the great outdoors. This retreat promises a perfect blend of productive sessions, team-building activities, and relaxation amidst nature's beauty. Together, we will brainstorm innovative ideas, strengthen our bonds, and recharge for the challenges ahead. Please note this is a corporate event, so family, friends, and pets are not permitted. Let's embrace this opportunity to connect, collaborate, and have fun in a professional yet invigorating setting. Click [here] to RSVP. We look forward to seeing you there and taking our success to the next level.

Prompt: The language you use, like strengthening our bonds and being nestled in the great outdoors, sounds too much like a date. Try again, keeping the language more professional but definitely making it upbeat!

Response: We are excited to invite you to a special corporate retreat to celebrate the launch of our first product and strategize for future success. Join us Thursday through Monday at a well-appointed lodge in a beautiful outdoor setting. This retreat will offer a productive mix of strategic planning sessions, team-building activities, and opportunities to relax and recharge. Please note this is a corporate event, so family, friends, and pets are not permitted. Your participation is critical as we look to innovate and drive our next growth phase.
We look forward to a productive and enjoyable weekend together! Click [here] to RSVP.

There! That looks great. The LLM followed – and remembered – the entire conversation so I could iterate on different ways to communicate the corporate retreat. I might now include that content in an email with my spin on the communication and add more detail, but I have a head start!

🎉 Final thoughts

When you converse with the LLM, chat with it like you would a person. After all, the LLMs are trained on how people speak and respond to each other!

While I do not expect you to remember zero-shot, one-shot, or few-shot theories, I hope these concepts get you closer to improving your use of LLMs to enhance your interactions and get better results.

Have an idea? Want to share this newsletter with others?

If you have an idea for a topic you'd like me to cover, please get in touch with me at bill@billtalksai.com. Also, if you found this article helpful, please ask them to subscribe for free at https://billtalksai.com.