Generative AI Prompt Engineering and Education

Welcome back to another edition of AI Insiders from the AI Marketers Guild. We have a special guest today—Erin Reilly, the founding director of the Texas Immersive Institute and Professor of Practice at the School of Advertising and Public Relations. Erin has a remarkable background and will introduce herself more thoroughly. Erin, you have developed practical, actionable ways to use AI, and I think our community will find your insights very valuable. We have a highly interactive group here, so please feel free to encourage participation at any time. I’ll also monitor the chat for questions. Erin, please take it away.

(01:13)

Thank you, David, for inviting me to share with you all. Let me provide a little background on my experience with AI. I’ve been working in the AI space for over a decade. My first significant project was with IBM while I was running the Annenberg Innovation Lab at USC. We focused on fan engagement, specifically on understanding what motivates fans to participate in their passions.

(01:50)

I developed a framework called “leveraging engagement” using natural language processing to analyze large amounts of online data. This allowed us to identify different types of fan motivations, such as learning, representation, or even debating—especially in sports, where arguing with referees is common. This approach helped us understand the ways fans connect with their interests.

By building both unsupervised and then supervised learning models, we created a system that’s now integrated into IBM Watson. It helps identify fan motivations, reduce churn, and inform strategies like recommending the next set of stories for a fan group. This was part of our early work in the field.

(02:57)

Currently, I teach a course on Creativity and AI at UT Austin, which fills up quickly every semester. Many have requested an online version so more people can participate. The course explores generative AI and its role in the creative workflow, which I’ll discuss with you today.

My focus has shifted from just audience engagement to how audiences now play an active role in immersive storytelling, especially with emerging technologies like AI as the backend infrastructure. We’re entering a spatial era where persistence and interoperability are key. The challenge is using these tools effectively to create real-time, innovative brand engagements. I work closely with brands on experiential marketing. Unless there are any questions about my background, I’m happy to jump right into the presentation.

(04:15)

Sounds good? Let’s begin. First, I’d like to thank Nicole Quail, one of the founders of AMG APAC, who recently hosted an event. I appreciate her for the introduction.

(04:48)

Are you seeing the presenter view? I want to make sure the right screen is shared. It looks like you can see my talking points. Let me adjust that.

 (05:15)

We’ve all been there. How’s this? Okay, here we go. Let’s get started. Before we dive in, I’m curious: How many of you have experimented with generative AI before? I believe most of you are already familiar with it.

---

### (05:37)

Since this group has experience, I won’t go into the backend technicalities of how AI works. Instead, I’ll focus on new tools and strategies for better prompting and using AI as a creative partner. This session will cover how to master prompts, parameters, and personalization.

---

### (05:59)

I’ll take about 30 minutes to review these topics and share some case studies using OpenAI, my go-to tool. But first, let’s step back and consider generative AI more broadly.

---

### (06:19)

Are any of you familiar with lifearchitect.ai? This website, created by Dr. Thompson, provides in-depth analysis of major AI models like GPT-4, Gemini, and Claude. Dr. Thompson compares their architectures, capabilities, and implications for the future of AI.

### (06:45)

This image shows the data used in two different models. On the left is GPT-3, which uses a proprietary dataset from OpenAI, and on the right is a more open, transparent dataset called The Pile version 1. I like to start here because the data your models are trained on is crucial.

### (07:22)

It’s important to understand that you shouldn’t rely solely on one tool. I provide my students with worksheets listing all the different tools I use and what I use them for. Because of differences in datasets and models, the results can vary significantly depending on your goals.

### (07:54)

Large language models are trained in two main phases. My early work with IBM focused on unsupervised, pre-trained models, while my later work involved supervised, post-training with human feedback.

### (08:14)

This is where we fine-tuned models using specific examples and human guidance. This process isn’t new. Many of us remember when Google started, and they gamified tagging images to help build their search engine and image recognition.

### (08:33)

There was even a game where you’d sketch an image, and the AI would try to guess what you were drawing. These early efforts helped foster collective intelligence and were crucial in building modern AI tools.

### (09:01)

That was early post-training with supervised human feedback. When people worry about AI replacing humans, I remind them that these systems are built by us. Human expertise and oversight remain essential.

---

### (09:43)

Large language models are incredibly promising and can generate text, images, music, and even 3D models. For example, Playbook AI is a great platform for 3D modeling, and Unity has built-in Muse AI for integrating AI into game engines. There are also new developments in creating AI “avatars” of yourself.

---

### (10:31)

These are just some examples of how AI is enabling new creative possibilities. However, there are challenges you need to be aware of, such as weak integration between tools, limited contextual understanding, data privacy risks, and AI hallucinations.

---

### (11:09)

Integration across tools is still limited, but that’s likely to improve over the next few years. AI may sound smart, but true understanding is still a challenge. I appreciate recent improvements in ChatGPT’s memory and personalization, but data privacy remains a concern, especially if the AI was trained on sensitive or proprietary information. Always be cautious with your data and check for hallucinations.

---

### (12:29)

If you encounter suspicious outputs, ask for sources. If the AI can’t provide credible citations, try using tools like Perplexity, which tends to give better references.

---

### (13:34)

There’s a whole field of study around how AI interprets data, and even engineers don’t fully understand these models’ inner workings. For example, Codex uses code to build solutions, but if you lack foundational knowledge in programming, you risk generating unreliable results or code hallucinations.

---

### (15:03)

AI is a tool, not a replacement for human expertise. It can provide basic understanding and outlines, but higher-level thinking still requires expert input. These tools enable a new kind of distributed cognition, much like spellcheck for writing.

---

### (15:44)

If you want to progress beyond beginner-level outputs, you must develop advanced skills and deeper understanding.

---

### (16:07)

I created a RAG GPT and sometimes the AI doesn’t return the expected answers, even when the content is correct. This highlights the challenges in engineering AI systems and the importance of prompt design and data organization. Using frameworks like WISER can help improve prompt effectiveness.

---

### (17:21)

I work with multiple AI tools for branding and strategy and would appreciate insights on how Claude differs from other models. Sometimes the results aren’t as strong, which might be due to the approach or prompting. Any tips on using Claude would be helpful.

---

### (18:25)

Anthropic, the makers of Claude, are very transparent. They recently shared insights into how AI models process thoughts, which is rare in the industry. I have a document comparing generative AI tools, including Perplexity and Claude, which I’ll share once I locate it.

---

### (19:46)

Understanding the data used in training models is important because it affects bias, toxicity, copyright issues, and privacy. Partners and clients may be concerned about where information goes and what the AI can infer. For example, repeated interactions with ChatGPT can lead it to make inferences about your identity and preferences as it builds a long-term memory.

---

### (22:13)

Privacy concerns vary by generation—Gen Z and Alpha tend to be less worried about certain aspects and more about others. As AI memory improves, these privacy questions will become even more relevant.

---

### (22:48)

Text-to-image models like DALL-E, Midjourney, and Stable Diffusion generate visuals based on text prompts. These models are powered by architectures such as diffusion models, GANs, and autoencoders. The more detailed your prompt, the more specific and useful the output.

---

### (23:48)

Think of prompt-writing as a new programming language. I use the WISER framework:

- **W**: Who is the AI? Assign it a role (e.g., marketing strategist).

- **I**: Instructions. Specify what you want the tool to do.

- **S**: Subtasks. Break down the task into clear steps.

- **E**: Examples. Provide specific examples, templates, or references.

- **R**: Review. Iterate with follow-up questions, clarifications, or requests for citations.

---

### (26:14)

You can also use your own draft as an example and ask the AI to improve it. After your initial prompt, review and iterate by clarifying sections, checking for hallucinations, expanding responses, or refining for the target audience.

---

### (27:16)

Be as specific as possible about your target audience and their behaviors. Instead of relying on broad categories like “millennials” or “Gen Z,” focus on specific motivations and emotional drivers, using your motivation framework.

---

### (27:40)

Consider the “temperature” setting in your prompts, which controls the model’s creativity. A temperature of zero yields safe, factual responses, while a higher temperature encourages more creative, novel outputs. For example, a low-temperature prompt for a coffee shop name might yield “The Coffee House,” while a higher temperature could result in more original ideas like “Bean There, Sipped That.”

---

### (29:02)

If you’re using tools with backend access, you can adjust the temperature directly. Otherwise, specify in your prompt whether you want a deterministic or creative response. The WISER framework helps you communicate these needs.

---

### (31:16)

The clearer your prompts, the better the results. Generic prompts yield generic outputs. Use the WISER framework for more personalized, precise responses.

---

### (31:47)

This approach aligns with instructional tuning, where you assign the model a role, adjust temperature, and combine prompts for the best results. I use prompt recipes—predefined templates for effective prompts—which I recommend saving and reusing in custom GPTs.

---

### (33:04)

Here’s an example of a prompt recipe: Assign the role, provide instructions, specify output format, clarify context and perspective, define the target audience, and list any parameters or inputs that might change.

---

### (33:51)

For example: “Act as a digital marketing specialist for an online publication. Provide a list of potential campaign ideas and strategies to increase sales and customer engagement. Use bullet points and headings, and make your suggestions specific, actionable, and tailored to different audiences. Do not include implementation plans or generic concepts. Write the content for a general audience.” Be explicit about what you do and don’t want.

---

### (35:43)

Like coding, prompting is an iterative process. Try a prompt, evaluate the result, tweak for clarity and relevance, and repeat. Evaluate outputs based on predefined criteria to ensure they meet your goals.

---

### (36:02)

There are five main ways to evaluate prompts:

1. Human evaluation—does the response meet your needs?

2. Benchmarks—test across different case studies.

3. Word-level metrics—useful for large-scale text comparison.

4. LLM-assisted evaluation—use one model to evaluate another.

5. Real-world application—continuously refresh examples and test in practical scenarios.

---

### (39:27)

For students, fostering reasoning and questioning skills is crucial. I believe our education system needs to update its approach, moving away from traditional models. I use Socratic dialogue and in-person discussions to develop students’ critical thinking. Peer-to-peer teaching and real-world projects, like working with actual clients, help students build authentic skills and accountability.

---

### (44:12)

Now, let’s explore the OpenAI ecosystem. It includes ChatGPT for text, DALL-E for images, Codex for code, Sora for video, and customizable GPTs. I use custom GPTs extensively, as they can save significant time and allow for deep personalization.

---

### (45:47)

For example, Instacart uses ChatGPT to enhance the grocery shopping experience, helping customers plan meals, save time, and receive personalized recommendations. This demonstrates how brands can leverage GPTs for tailored customer engagement.

---

### (46:39)

In the creative space, Alexia Donna, Director of Creative Technology at Edelman, uses DALL-E to prototype new products and generate mood boards. While AI doesn’t replace designers, it serves as a valuable starting point for creative teams to brainstorm and iterate ideas.

---

### (47:51)

My friend, a film director, uses AI to storyboard scripts and communicate his vision to art departments. AI-generated images allow non-artists to visualize and share concepts quickly, making it a powerful tool for collaboration.

---

### (48:13)

It’s different from traditional mood boards because you can now fully realize your ideas visually using AI. This capability democratizes the creative process and speeds up collaboration.

---

### (48:57)

Finally, Codex enables natural language commands to generate 3D scenes in engines like BabylonJS. This makes 3D modeling and spatial web design more accessible, allowing quick prototyping without advanced technical skills.

---

### (51:44)

Through prompting, you can render 3D models for brand websites or other projects, making it easier to communicate ideas before handing them off to specialists.

---

### (52:08)

We’ve covered a lot today. I’ll share links and resources after this session. Based on the great feedback in the chat, we might need to schedule a part two, as we didn’t even touch on music generation or other creative applications.

---

### (53:09)

Thank you all for joining. Please feel free to reach out if you have more questions or want to connect further. I appreciate the opportunity to share this master class on creativity and AI with you.