Prompt Like a Pro: Agentic Marketing Best Practices | Dreamforce 2025

All right, is everyone ready to talk prompting? Hi, everyone. Before we get started, just a quick love letter from our legal team to you. Please make all purchasing decisions based on publicly available information. But now that that's out of the way, thank you so much for being here. And thank you, everyone who. Who's tuning in on Salesforce as well. My name is Jiang Jiang. I'm a product marketer here at Salesforce for agentforce Marketing. And I'm joined here by my colleague Scott Robbins, who's a principal technical architect. And today we're gonna talk all things prompting. So who here in the crowd thought to themselves, when you were a child? Wow. You know what I really wanna be when I grow up? A prompt engineer. Really? I highly doubt that. But in today's age of AI and agents, it's really important for marketers to become pros at prompting in order to stay competitive. So today we're going to cover all things prompting. What makes a good prompt? Why does it matter to have a good prompt? How do you scale good prompts with prompt templates? And how do you customize prompts with agent force marketing? So let's go ahead and dive in. So why do good prompts even matter? Well, a good prompt or prompts are kind of like making a wish to genie. If you're vague and you say something like make me dinner, you might get a single boiled potato, probably not quite what you had in mind. But if you're specific and you say something like, hey, make me a spaghetti dinner with some sauteed zucchini on the side and some turkey meatballs, go crazy with the garlic, then you'll get exactly what you want. So in simple terms, a prompt is a set of instructions that you give to an LLM. And it's really important that you give GENIE exactly the right instructions so that it can deliver the best output for you. So you can think of prompts as the new creative brief. Agentforce is powerful, just like genie. But if you feed it great prompts, then it'll give you even greater outputs. Just like a great creative brief will give you great marketing outcomes, it'll lead to faster campaign creation. It'll lead to better personalization and more relevant conversations. Okay, so now let's dig into a few practical, general best practices for prompting. 101. So first, AI doesn't read our minds, at least not yet. So one thing you want to do is be really, really clear about what you're trying to achieve, what you want the AI to do, and give it context. And also what you want the AI not to do by giving it constraints, then you want to be really specific. Just as a good creative brief would contain information about your target audience, the channels that you want to activate on your brand tone as well as the cta, you want to do the same thing here. Then you also want to ground with data. The difference between grounding and not grounding with data can be a generic marketing email or a generic marketing message, and one that's really tailored to each specific individual audience. What's great about AgentForce marketing is that it's built directly on the Salesforce platform, so it can pull in data from across all of your departments, across service, sales, commerce, and with data cloud, we can pull in data from your internal business systems as well as external systems. So think data warehouses or data lakes as well. And finally with AI specifically you really want to iterate. So even if the prompt seems really clear to you in the beginning, it may not give you exactly the right results. So you want to test it out and see what it leads to until you're really happy with the results. Okay, so let's take a look at what this looks like with an example. So imagine I'm creating a campaign. An example of a not so great prompt might be something like make me a campaign for our fall sale. And here's the output that you might get from a prompt like this. So, dear valued customer, our fall sale is here. Explore fantastic savings. And here are all the different product categories and offerings that we have for you. It's okay. Like this is an okay marketing email, right? But it's also not great. It's generic. You probably get hundreds of emails like this every day and it doesn't really stick. So what went wrong here? Well, first, this prompt was very vague. We didn't provide any goals or objectives or context. We just said make me a campaign. We also didn't specify the audience, the tone or the channels. And finally, this email is not grounded in customer data. So let's see how we can do better. This is an example of a better prompt. Create a multi channel campaign for outdoor adventure event targeting loyalty members in California. The goal is to re engage customers who haven't made a purchase in 30 days. Personalize each message with the customer's most browsed product category and recommend the next best product offering 15% off draft email and SMS messages in a friendly, exclusive tone. Then create a customer journey that optimizes engagement. Now let's take a look at what kind of output we might get with this better prompt. So Right off the bat. Hi, Lauren. Right. This is customized to an individual customer. We recognize that she's based in California, and we're inviting her to a specific event based out of California. So outdoor adventure event. And because this email is grounded in data, we have all this information about her. We know that she's an outdoor enthusiast. She has browsed specific categories, probably like outerwear, and so we're recommending a very personalized product for her. Okay, so now that we've taken a look at prompts and what makes great prompts, I want to flip it over to Scott. Well, actually, before we do that, I'm going to cover up on why this prompt was great. So, first, we provided AI great objectives and great context. So we specified that, hey, the goal here is to target loyalty customers who haven't re engaged, and we want to drive to purchase. We also specified the target audience, those based in California who are outdoor enthusiasts. We specified the channel, and we specified the tone. And most importantly, you heard me talk about this already. We grounded this email in CRM data. Okay, so now we're going to pass it over to Scott to talk about how we can take great prompts and scale that with templates.

Thank you. All right, so I think we see how powerful the LLMs can be and how much they unlock for us for content generation. We also see that the results we get are directly related to the prompts that are sent to those LLMs. What we can't do as marketers really is make our customer experience just dependent on how well our users are able to prompt those LLMs. And that's where prompt templates come in. So within Salesforce, Prompt Builder is there, and this is your source to design and author reusable prompt templates that can create specific sets of instructions that are consistently applied every time a request requires it. You don't need technical skills to build these, and they unlock a ton of different benefits or advantages for you. We don't have the time today to go deep into each of these six advantages, but these are the ones that come to mind most often. First of all, you don't need a technical skill set. You don't need to be a prompt engineer to get in and make a highly effective prompt. Once you make the prompts, you're able to use them anywhere, so you can trigger backend processes with them using flows. You can get them directly in the UI through something like an embedded app or a component, and then you can expose them to agents so that your agents can use them as actions anytime a Request from a user might require it. Finally, the two that are really important here is that useanywhere. We'll see that in a moment with our live demo. And then the LLM flexibility, because as marketers, we like the creative aspects. And every LLM may handle the same request slightly differently. And so when you're putting in instructions around your brand tone and the content you'd like to see, you may have a preference for a certain model, or you might find that a certain model just does a really great job and you want to stick with that one. So let's look at how a prompt template comes together and how it helps. So here's an example of a prompt template. This is what it would look like in Salesforce. This particular example looks at different types of data and uses it to recommend the most relevant marketing journey for any contact. And it has a few things that I want to call out. So first, it provides a clear role. And this one, I think, is the easiest to overlook. And that's because you and I, as we all go into our job, we sort of have this inherent embedded understanding of our role and what we do. LLMs don't have that, so we need to give them some perspective to approach any task with. Next, we're inserting this dynamic grounding. So you see any of the blue areas in the prompt template. We're using dynamic grounding to put in things like, which customer are we referring to? So we saw in Jiyoung's example, Lauren's first name was coming in. We know that person's profile. Next, we're looking at the engagement that that person has had recently with our campaigns. So behavioral data that tells us exactly which things they have an affinity for. And the last piece there is, we're looking at all of our available marketing journeys. And as you retire, journeys, add new journeys, create new campaigns. That list will be constantly changing, and that's why we use a dynamic grounding there instead of just a static set of data. Next, we're giving the LLM a series of detailed instructions. So this gets down to the tasks that we're effectively delegating to it. It's as if we're handing off a task to a coworker. In this case, it'll be the model. You can structure this in different ways. You can put in a numbered list. But the idea here is that you're telling it how you'd like it to reason and what you're asking it to do. Beneath that, in a separate set of instructions, you can give it some direction on what you'd like the output to be. So this could be things like return HTML, because we're going to use this in a different editing system, like a marketing cloud content block. It can be things like how long or how concise to make it. So if we're using it for an SMS output, we need that to be a really brief message and impactful. If it's an email, it can be a little bit more longer, more verbose and include more characters. Finally, at the end, just reiterate the request so that when the whole package goes to the LLM, there's no confusion around what we're asking it to do, and we've effectively asked for the output. Then if this looks like a lot or it's a new skill set, don't worry. Down at the bottom, there's a pro tip. Any of your favorite generative AI tools, you can go to them and say, here's what I'm working on. Provide an LLM optimized prompt, because LLMs understand how other LLMs think and they will give you something like this structure. Now, I did mention agents earlier. Prompt templates are one of the foundational agent action types that you can use alongside other things like Flows and Apex and connecting to external systems through APIs. Once you have that template ready to go, you publish it. You can assign an agent action to it. And now, as your users interact with different agents, whether it's an internal employee or an external customer, that prompt template becomes available. So if you're having users summarize things like campaigns, segment demographics, whatever it may be for your business, the prompt can ensure that there's always a consistent output that contains the information that you expect to be surfaced. Now, let's see a live example of how we would apply this prompt template in a true marketing context. And we'll do that through the lens of a hospitality company called Coral Cloud Resorts. They want to acquire new customers in a highly competitive area that depends on a lot of seasonality, and they want to understand how an agent or prompt templates can help them to automate completely personalized content at scale, because they have so many variations that just using typical dynamic content is difficult to maintain. All right, let me flip over. So we might start with something like this, just a landing page. And we'll start out by saying, I'm interested in learning more about what a Coral Cloud trip could look like. That was the most stressful part because if I got my email wrong, this would all be over. And we're taking a family Trip. It's a very basic setup. A couple of adults, a couple of children, and then we have a selection here for what types of interests we have. Say we want some adventure, definitely want some good food, and maybe see some wildlife along the way. We'll submit this page. Okay, so in a moment we'll get an automated email back from that. But let's think about what we just did. We selected three interests out of 12. If we let every single person select three different interests out of 12, that's over 1300 different variations of content just from this one page. So imagine just trying to manage all of that within dynamic content blocks and still not have it look very slotted. Right. It becomes a very cookie cutter look. Like we want it to be completely fluid in a single message. So we refresh our inbox. We've got our new email and we see what awaits us at Coral Cloud Resorts. You can see in the styled text it's recommending the Canyon Zipline excursion, the culinary heritage experience in the Starlight Wildlife Tour. So it's great because this lines up directly with those interests that I just selected on the page. And what I want to mention is, has anyone ever said that we want to use AI, but our data isn't ready? So everything we just saw was data that we just collected in a standardized format just seconds ago. You can do these types of use cases and then at the bottom there's a dynamic offer. So we're using dynamic content targeted by content that didn't even exist because we generated it through the LLM. So let's toggle behind the scenes. This is the prompt template that just drove that email experience. So the role is a concierge. The objective is to give that personalized resort summary. We're inputting those preferences 1, 2 and 3 dynamically and then we're giving that series of instructions and output formatting and what this might look like if I. Well, real quick, the model we're using is GTV4 Omni. There's the whole list here. Like I mentioned, other providers, we're asking it for HTML because we do want to pass that to Marketing Cloud. And then on our preview side here, I ran a preview of hiking, dining and family time. So a different set of interests than we saw a minute ago. And you can see the output here is the Rainforest Discovery Hike, the Farm to Table Cooking Class, and the Family Adventure Quest. Now, the 1300 variations I mentioned a minute ago, if we think back to permutations, that's allowing for one option per selection. What if we allowed repetition and we said there are multiple options per interest category, so maybe every one of those 12 categories had five different experiences to it. Now the number of variations goes up to about 45,000. So imagine 45,000 different variations of a dynamic content when we could just pass this prompt template to an LLM and get a completely tailored response every time. Seaside Gourmet Dinner Chef's Table Culinary Journey Culinary Heritage Experience. So that's what we're able to do with a single prompt template for every single subscriber that you might have. All right, let's go back to our slides and just wrap up with a couple of final messages. All right. Hopefully you found this useful. Seeing how prompt templates can inform elevate your marketing through personalized content. A few takeaways or messages to leave with you. You can start checking this out. If you haven't activated any of your agent force capabilities, talk to your account team. They will help you get stood up, get you working on some basic prompt templates. You can play with them in a sandbox if you'd like. We do have the built in prompt templates and out of the box topics and actions that you can start to play with too. Those are really easy to activate. You can access all of our documentation right here. And then finally we do have a brand new agent force for marketing trail on trailhead that you can access that QR code that will start walking you through and getting your hands on with some of those capabilities we just looked at. And finally, don't miss the marketing keynote tomorrow. That's going to cover even more of the latest innovation that's going on with us. A lot of things to be really excited about that I can't wait to see and I hope to see many of you there. So thank you for coming today. Really appreciate the time and hopefully you found it useful.