ChatGPT for conversation designers: generating sample dialogs with personas

Previously, I shared an exploration of how conversation designers can use ChatGPT to edit conversational AI dialog. Let’s move on to something a little meatier, with some prompt-within-a-prompt action: creating sample dialogs with user personas.

Disclaimer: don’t use AI-generated personas

This article is not about creating personas with ChatGPT.

Ideally, you should be using real personas created by user researchers — you know, the folks whose bread and butter is researching users — to be anywhere near the realm of accuracy.

The persona you copped from your team’s UXR asset library? That’s legit. The persona you asked ChatGPT to create from scratch? That’s a hard pass.

Don’t expect ChatGPT or other LLMs to replace actual user research. AI is exciting, but we shouldn’t reshuffle our design process around it just because it’s there.

“But I’m a designer, I already know who the user is.”

You’re a designer, not a researcher. Giving ChatGPT an amateur interpretation of your user’s behavioral traits, frustrations, and motivators will result in something you won’t be able to use to make good design decisions.

ChatGPT and LLMs like it will reformat, rephrase, regurgitate words that may result in an imprecise, inaccurate portrayal of your user. As much as you refine and reprompt, you’re just not going to get a real, reliable persona out of an AI this way.

TLDR; don’t make personas with ChatGPT. 🤷🏻

Caveat to simulating dialogs with personas

If a persona is a representation of a user, then anything an AI builds on top of it would be a representation of a representation of a user. (🤯) This applies to sample dialogs based on a legit persona. Once you plug your persona into ChatGPT, it becomes the AI’s interpretation of that artifact. It can only generate text that it predicts would correlate with the text you used to communicate your persona.

So while this is a fun exercise to create sample dialogs and explore potential conversational pathways, always keep in mind that there’s no substitute for a real user.

Wait, what are sample dialogs?

Sample dialogs are exactly what they sound like — quick examples of back-and-forth conversations between an AI and a user. As conversation designers, this is a critical part of exploring the different routes a conversation might take and inspecting which of those branches are worth accounting for in our final designs.

For example, the following sample dialogs explore different ways a conversation could go down stemming from the same utterance.

Sample dialog 01.

🧑🏽‍🦱 “Turn on the lights, please.”
🤖 “Which lights do you want to turn on?”
🧑🏽‍🦱 “All the lights downstairs.”
🤖 *earcon*

Sample dialog 02.

👱🏼 “Turn on the lights, please.”
🤖 “Is that for the kitchen light, or the bedroom lights?”
👱🏼 “Bedroom.”
🤖 ”Alright, turning on the bedroom lights.”

Creating persona-driven dialogs with ChatGPT

We don’t always have a particular persona in mind when writing sample dialogs — they’re more commonly centered on use cases.

Designers can upscale sample dialogs by asking ChatGPT to reference a persona and turn factors like behaviors, sentiments, and motivators into conversational variables. But more on that in a bit.

OK, let’s jump in.

1. Share your persona with ChatGPT

As a reminder, ChatGPT can’t read images. Personas often contain visual elements that may require transcription, so you’ll want to do that before getting started.

My prompt:

Here’s a persona description for Jenna.

Name: Jenna
Age: 35
Relationship status: married
Occupation: stay-at-home mom
Location: California, USA
Interests: health, nutrition, cooking, fitness, technology, social media

Bio: Jenna is a tech-savvy stay-at-home mom of two young children. She takes pride in managing the domestic errands, grocery shopping, and cooking nutritious meals for her family. Jenna is always on the lookout for new recipes and ingredients that are healthy and delicious. She keeps up with the latest health and dietary trends by actively seeking information through online and social media channels. She enjoys using technology to make her life easier, from meal planning apps to smart kitchen appliances. In her free time, Jenna likes to stay active by doing yoga and going for runs.

Jenna’s motivations:
1. Her children
2. Her physical health
3. Fear

Jenna’s goals: keeping her family healthy, becoming knowledgable about nutrition, spending less on groceries and home expenses.

Jenna’s frustrations: hidden ingredients, conflicting online information on dietary trends, her kids not liking healthy food, expensiveness of organic produce

A quote from Jenna: “I would like to provide my family with healthy, budget-friendly meals that are nutritious, organic, and easy to prepare.”

Note: I’m not a user researcher, and I can’t share real personas I’ve used in the past — so for the purposes of this article, a half-assed persona will do.

Now, ChatGPT should be able to recall your persona throughout the rest of your conversation. If it can’t, offer a reminder to jog its memory — this can be as simple as mentioning the persona’s name or occupation.

2. Include a description of your AI’s voice and character

Virtual assistants have personas, too!

To make your test dialog more accurate, your stand-in AI should reflect your real AI’s voice and character.

My prompt:

Here’s a persona description for Bottie:

Bottie is a gender-neutral voice assistant who is helpful, friendly, and polite. Bottie’s main priorities are giving people quick and reliable responses to their question and helping them complete tasks around the house. Bottie is friendly, but not too energetic. Bottie is kind, but not a pushover. Bottie is intelligent, but not a snob. Bottie’s responses are never longer than 350 characters. Bottie’s speech is always aligned with Grice’s Maxims of Conversation.

Try refining your AI’s description as you go along — there may be something you forgot to add that could change the course of these sample dialogs. Or maybe your AI’s generated script uses too many exclamation points and you want ChatGPT to cut that out.

It can be annoying to keep track of all these moving pieces at first, but the more you iterate, the more accurate your output will be.

3. Generate ideas for a scenario (optional)

If you don’t have a scenario or use case in mind, you can ask ChatGPT things to reveal more about your persona’s life:

• What does this person struggle with?
• How do they go about their day?
• What are their daily touch points with technology?

This exercise may help you generate a good context to use when setting the scene for your dialog. But you can skip this if you can think of a context that’s more precise to your design exploration needs.

My prompt:

How does Jenna interact with technology in their daily life?

ChatGPT’s response:

Jenna interacts with technology in various ways to simplify and streamline her daily tasks as a stay-at-home mom. She uses her smartphone to manage her family’s schedule, stay connected with friends and family, and access online resources for health and nutrition information. Jenna also uses meal planning and grocery shopping apps to help her save time and money.

In the kitchen, Jenna uses smart appliances like a smart oven or smart blender to help her prepare meals more efficiently. She also uses social media and online resources to discover new recipes, cooking tips, and meal ideas.

Overall, Jenna values technology as a tool to help her manage her household and stay organized, while also providing her with access to valuable information and resources related to her interests in health and nutrition.

I already knew that I wanted Jenna to talk to her virtual assistant on the smart fridge. So that’s where I’ll start.

4. Set the scene and hit “send”

Here are some details you can give ChatGPT to establish the scenario you’d like to explore:

Situational contexts

• What your persona is currently doing
• What devices your persona is close to or currently using
• What your persona wants to know
• What your persona wants to do
• What’s frustrating your persona

Conversation sentiment and success

• “In the end, Jenna is happy with her conversation with Bottie.”
• “At first, Bottie is unable to answer Jenna’s question because Bottie couldn’t hear her. Bottie lets her know that and then continues to help resolve Jenna’s query.”

Targeted modalities (especially important for multi-modal designers)

• “Whenever Bottie’s response includes an image, please note the image in asterisks *like this.*”
• “Throughout the conversation, Jenna only communicates with Bottie by voice.”

Tip: if ChatGPT forgot the details of your personas, or if you don’t feel confident that it remembered, you can write a longer prompt that contains your user persona, your AI’s character description, and the setting you want to them to interact in.

My prompt:

Using the persona descriptions of Jenna and Bottie, create a conversation between them in which Jenna is looking for healthy recipe recommendations and asks Bottie the nutritional value of at least one recipe that Bottie suggests. Bottie is responding from Jenna’s smart fridge screen. End the conversation with Jenna choosing a recipe and feeling satisfied with Bottie’s responses.

ChatGPT’s output:

Jenna: “Hey Bottie, can you suggest some healthy and easy-to-prepare recipes for tonight’s dinner?”

Bottie: “Of course, Jenna! How about grilled chicken with roasted vegetables or quinoa salad with roasted sweet potato and kale?”

Jenna: “Those both sound great! Can you tell me the nutritional value of the quinoa salad?”

Bottie: “Sure thing, Jenna! The quinoa salad has 310 calories per serving, 16g of protein, 10g of fiber, and 25g of carbohydrates. It’s also high in vitamin A, vitamin C, and iron.”

Jenna: “Wow, that’s impressive. I think I’ll go with the quinoa salad. Can you send the recipe to my phone, Bottie?”

Bottie: “Absolutely, Jenna. The recipe has been sent to your phone. Let me know if you need any help with the ingredients or cooking instructions.”

Jenna: “Thanks, Bottie! You’re always so helpful. I appreciate it.”

Bottie: “No problem, Jenna. Enjoy your healthy and delicious meal!”

Using this approach to create sample dialogs can help conversation designers inspect and explore conversations across different device types and interaction modalities, various contexts, and even between personas.

For example, I didn’t expect Jenna to ask Bottie to send the recipe from the fridge to her phone, but it’s an interesting and plausible cross-device interaction to consider for this scenario.

Building sample dialogs around pre-existing responses

By plugging in your persona and a particular response you want to test, ChatGPT can demonstrate how it might play out in an almost-real conversation. Instead of asking for a fully fabricated conversation, as shown previously, you can prompt ChatGPT to use a response you already made to see how it might perform.

My prompt:

Create a new conversation between Jenna and Bottie, in which Jenna asks Bottie for cooking oils she can use as a healthier alternative to coconut oil. Then Bottie gives the following response before the conversation continues:

“I found a few alternatives to coconut oil. There’s olive oil, butter, and sunflower oil. Want to know more about these?”

ChatGPT’s response:

Jenna: Hi Bottie, can I ask you a question about cooking oils?

Bottie: Sure, Jenna. What’s your question?

Jenna: I’ve been using coconut oil for cooking, but I heard that it’s not the healthiest option. Do you have any suggestions for a healthier alternative?

Bottie: Yes, I found a few alternatives to coconut oil. There’s olive oil, butter, and sunflower oil. Want to know more about these?

Jenna: Yes, please. Can you tell me more about each of them and how they compare to coconut oil?

Bottie: Sure thing. Olive oil is rich in monounsaturated fats, which are known to lower cholesterol levels and reduce the risk of heart disease. Butter is a good source of healthy saturated fats and is a great option for baking or sautéing. Sunflower oil is high in vitamin E and is a great option for frying and roasting.

Jenna: That’s great to know, thanks, Bottie. I think I’ll give olive oil a try. Do you have any recommendations for a specific brand or type of olive oil?

Bottie: I recommend using extra virgin olive oil as it is the least processed and contains the highest amount of beneficial nutrients. Some good brands to try are California Olive Ranch, Pompeian, and Lucini. Just be sure to store it in a cool, dark place to preserve its quality.

Jenna: Thanks, Bottie. I’ll keep that in mind. You’re a great help!

Bottie: You’re welcome, Jenna. Let me know if you have any other questions.

Generating conversations with visuals

I mentioned previously that you can ask ChatGPT to denote whenever a visual appears in your assistant’s response. Here’s one example of how you can do that.

My prompt:

Write another conversation between Jenna and Bottie, this time Bottie is an AI speaking from the screen on the fridge and can show pictures. Every time Bottie’s response has a visual, describe the visual in Bottie’s response surrounded by asterisks.

Example:

Bottie: Here’s what I found. *List of recipes with pictures of food.*

ChatGPT’s response:

Jenna: Hey Bottie, I’m in the mood for something sweet, but I don’t want to eat junk food. Do you have any healthy dessert ideas?

Bottie: Sure thing, Jenna! How about making some chia seed pudding? It’s a healthy, satisfying dessert that’s easy to make. *Displays a picture of chia seed pudding with fresh fruit toppings.*

Jenna: Hmm, that looks interesting. Can you tell me more about it?

Bottie: Chia seed pudding is made by combining chia seeds with a liquid, such as almond milk or coconut milk, and letting it sit in the refrigerator for a few hours until it thickens. You can sweeten it with honey or maple syrup, and top it with fresh fruit or nuts for added flavor and texture.

Jenna: That sounds simple enough. Do you have a recipe I can follow?

Bottie: Absolutely! Here’s a simple recipe to get you started. Combine 1/4 cup of chia seeds with 1 cup of almond milk and 1 tablespoon of honey in a bowl. Whisk everything together, cover the bowl with plastic wrap, and let it sit in the fridge for at least 2 hours, or overnight. When you’re ready to eat it, top it with your favorite fresh fruit or nuts. *Displays a picture of the ingredients and step-by-step instructions.*

Jenna: Great, thanks, Bottie! I can’t wait to try this out.

Bottie: You’re welcome, Jenna. Enjoy your healthy dessert!

Overall, this is a good exercise if you’re brainstorming dialog, exploring potential use cases, or building on sample dialogs you’ve already written to see how they play out in a full conversation.

But remember, these are simulations. There’s no replacement for real user interactions with your AI.

Have you tried creating and testing sample dialogs with ChatGPT? Share what works for you in the comments! 💬