Salesforce Prompt Builder Features Every Admin Should Know

Salesforce Admins Podcast

Today on the Salesforce Admins Podcast, it’s time for a deep dive with Josh Birk, who talks to Raveesh Raina, Principal Solutions Engineer at Salesforce.

Join us as we chat about what Prompt Builder can do and how to write effective prompts.

You should subscribe for the full episode, but here are a few takeaways from our conversation with Raveesh Raina.

Intro to Prompt Builder

Raveesh is the perfect person to talk to about Prompt Builder and all the cool stuff you can do with it. As a Solutions Engineer, he works with customers to implement the latest and greatest Salesforce innovations and right now, that’s Prompt Builder.

Prompt Builder takes all the power of LLMs and combines it with your Salesforce data. You can use it to help you write personalized emails to customers, build out records with more information, and much, much more.

The four prompt template types and what they do

Right now, there are four prompt template types in Salesforce:

1. The sales email template drafts an email you need to send to your customer. Some examples include introducing them to a new product, or reminding them about an appointment.

2. With the field generation template, you can quickly and easily populate or complete a specific field on a record page with a summary or description created by an LLM.

3. The record summary template does exactly what it says it does: summarizes a record. This one is really easy to use via Einstein Copilot, and his clients love it for meeting prep.

4. Finally, there’s the flex prompt template. This can be used anywhere and everywhere on the platform to create a customized prompt template that incorporates records from multiple objects simultaneously. An example would be to create a personalized product recommendation for a customer based on an Einstein Next Best Action.

With all of these prompt templates, you can dynamically ground them with data from Salesforce or Data Cloud. That gives the LLM the power to pull data from records or, with flows, from pretty much any object—standard or custom—in your Salesforce org.

How admins can write effective prompts

So how do you write effective prompts that do what you want them to do? Raveesh has four tips to share with us:

1. Be explicit about your expectations and goals. What are you hoping to get out of the response? The AI needs a clearly defined goal in order to generate a good response.

2. Contextualize the information. State if you want to add related records to contextualize the response.

3. Specify your role. What is the persona for which this prompt template is built?

4. Add limitations and set boundaries. For example, “do not exceed 500 characters.” The AI needs to be told, and sometimes told again, what not to do. This is especially important to think about as you test and refine your prompt in order to get consistent results.

There’s a lot more great stuff from Raveesh about building better prompts and how Salesforce protects your data, so be sure to listen to the full episode. And don’t forget to subscribe to hear more from the Salesforce Admins Podcast.

Podcast swag

• Salesforce Admins on the Trailhead Store

Learn more

• Blog: The Ultimate Guide to Prompt Builder | Spring ’24

• Trailhead: Get Started with Prompts and Prompt Builder

• Video: Salesforce Admins Innovate with Prompt Builder

• Podcast: Get to Know Prompt Builder with Marissa Scalercio

Admin Trailblazers Group

• Admin Trailblazers Community Group

Social

• Connect with Raveesh

• Salesforce Admins: @SalesforceAdmns

• Mike on Threads: https://www.threads.net/@mikegerholdt/

• Mike on Tiktok: https://www.tiktok.com/@salesforce.mike

• Mike on X: @MikeGerholdt

Full show transcript

Josh Birk:

Greetings everybody. Guest host, Josh Birk here, to take another deep dive into a topic. And this week, the topic is going to be Prompt Builder and Building on Prompts. I am joined by my good old colleague, Raveesh Raina, who is one of the people who wrote, for instance, our Ultimate Guide to Prompt Builder. Now, with no other hesitation, let's get over and talk to Raveesh.

All right, today on the show, we welcome Raveesh Raina to tell us about all things Prompt Builder. Raveesh, welcome to the show.

Raveesh Raina :

Thank you very much, Josh. It's an honor to be here and thank you for having me.

Josh Birk:

Excellent. Well, let's start off with a little bit about you. Tell me, what is your current role at Salesforce?

Raveesh Raina :

So I am a Principal Solutions Engineer. I work predominantly with our account executives and account directors to helping customers get inspired with our latest and greatest innovations that we are publishing and that we are releasing into the market. So I have a specific focus towards financial services. I'm based out of Toronto, Canada, so I work with all of our major enterprise banks with a lot of the different innovations and as far as financial services cloud is concerned, which is our flagship industries product. So helping them get inspired and sharing ideas on how we can help them better meet their goals.

Josh Birk:

And was computer something you always wanted to get into?

Raveesh Raina :

No, it's definitely an area that I personally have tried to get better at.

Josh Birk:

Okay.

Raveesh Raina :

And it's definitely an amazing opportunity and an amazing space to be.

Josh Birk:

What did you go to college for?

Raveesh Raina :

So funnily enough, even though I am based out of Canada, I actually went to the United States for my undergrad and I did mechanical engineering.

Josh Birk:

Ah.

Raveesh Raina :

So I graduated from New Jersey Institute of Technology, NJIT, with a mechanical engineering degree. But life had a different set of actions for me and somehow, I ended up aligned to the Salesforce ecosystem a long time ago.

Josh Birk:

Nice. Okay. Now, today we're going to talk about Prompt Builder. And when customers ask you about Prompt Builder, they're not familiar with it, what's the early elevator pitch you give them?

Raveesh Raina :

I essentially tell customers that Prompt Builder is an easy to use tool, which allows them to send a predefined set of instructions to an AI model, with the expectation of getting a response that has both value and meaning to them. And by value and meaning, I'm specifically referring to within the context of the data that they are providing via their CRM for Salesforce.

Josh Birk:

Okay. And walk us through the four templates we have right now that frame that data and that value add?

Raveesh Raina :

Yeah, absolutely. So right now, as of today, we have four prompt template types, sales email prompt templates, that's the first one. The second one is going to be field generation. The third one is what we call flex prompt templates. And the fourth one is record summary prompt templates.

Josh Birk:

Yeah. Now, I think the first two there are pretty straightforward. We're generating an email to send out to somebody, and if we're summarizing or we're generating a field, it means instead of a human putting in things manually, the GenAI is going to create it. Tell me a little bit more about the second two because I feel they're off to the corners and they're also very different from each other. Right? What's the purpose of a flex template and what's the purpose of the record summary template?

Raveesh Raina :

Yeah, so what I'll do is let me tackle record summary template first.

Josh Birk:

Okay.

Raveesh Raina :

Just as its name suggests, a record summary template is used by an admin to summarize a record or to get a succinct overview of a particular record. It is best served, best delivered in the UI via Copilot. So that is what we've seen greatest success in, in that record summary prompt templates can be surfaced up via Copilot, but it allows the admin to summarize a record based on whatever data model that they have already implemented.

Josh Birk:

Got it.

Raveesh Raina :

So that's a record summary template. Now, flex prompt templates, they are really powerful and versatile because what they allow an admin to do is to connect up to five objects together in a prompt template for the purposes of grounding. Those five objects could be standard objects or custom objects. They don't have to be related to each other in any way, shape or form, but through the power of the UI that the admin is surfacing, they can ground the flex prompt templates with each of those different objects with the expectation of getting a response from the LLM.

Josh Birk:

And for listeners at home, if you have somehow missed all of the wonderful Salesforce marketing about Einstein Copilot, it is a conversational AI for your platform and we will discuss that in another episode. Now, when it comes to these kind of templates, let's focus on the sales email, as I think it's the most straightforward. Some people listening might be like, "Well, I already know how to generate an email. Why do I need generative AI to do this?" What do you think the value add specifically that these are being pulled in through an AI model and having a generative response as opposed to a human created one? Where do you think the power lies there?

Raveesh Raina :

So the way I like to think of it where generative AI sales email templates provide value is the fact that it provides an easy mechanism for a user to craft an email in a conversational style that includes human and natural language text into it. If I were to use a standard email template as an example, that is static text, but replaced with either merge fields or even sometimes even with related records. But it removes that humanality sort of, from crafting an email. And the recipient may recognize that this is static text, but with generative AI, it adds more humanity towards that email template type. So that's where I feel is some of that value added.

Josh Birk:

And I think to some people, they might be scratching their eyebrows there because they're like, "But Raveesh, I'm using a computer to generate this email." But I think that's exactly spot on because we've all gotten that email that's like, "Hello, first name at company name. Congratulations on your date." And it's like you know instantly, this isn't from a human, but what these GenAI models are really good at is mimicking a human-like response, this human-like writing. And so it's like you get something that's acting like it's somebody to draft an email every single time you're sending this out to somebody.

Raveesh Raina :

Absolutely. Absolutely. It adds a little bit more personalization to a piece of communication that was missing with your static email templates in the past.

Josh Birk:

And also, I think it's hard to do it on a podcast, right? It's hard to show it on the podcast, the power of being able to ground this in Salesforce data. Tell me a little bit about... For each one of these templates we're talking about, what kind of data can we bring into the AI model to let it represent whatever it is that we're trying to generate?

Raveesh Raina :

Absolutely. So the real power of our generative AI tools with Salesforce, in my personal opinion, is the fact that we allow admins and customers the option to dynamically ground these prompt templates. What we mean by dynamically grounding is adding placeholders in these prompt templates that will be automatically replaced with either merge fields directly from a record or related lists that trace back to the origin of that record, or even through the power of flows. Once again, get a subset of related records associated to that original parent record as well.

So dynamically grounding can be done using merge fields, related lists, flows to get subset of records. And you can take it two steps further by grounding with Apex classes or even through the power of Data Cloud data model objects as well.

Josh Birk:

And what I think is also going back to the old way of doing things and doing things with this way, using flow, but also just because it's a conversational model, you can instruct the AI to do things based on the data. So it's not just, "Say something about this person's description." It could also be like, "Based on the fact that this person has no survey scores, change your response based on that." And what are some things that people who have flow skills can do to apply that extra layer of logic? Give me a good example of something that would change the course of the outcome based on a flow?

Raveesh Raina :

Yeah, absolutely. So I did mention that the power of flows is the fact that you can get subsets of records. So for example, if I want to surface up using a generative AI response, a summary of all open cases associated to a client or an account. If I were to use the standard related list component or resource picker, then it would give me a summary of all the cases associated to an account, whether it's open, closed, escalated, high priority, so on and so forth. But with flows, if I just want to surface up open cases only, that is what I can do through the power of dynamic grounding as well.

Josh Birk:

And what I love about it, if anybody's here, maybe for some reason, you haven't dipped your toe into the flow pool, the flows we're talking about can be very... I can do these flows, and I'm not Jen Lee. Flows are her job. She's the one who goes like, "Oh yeah, flow can do that crazy, impossible thing. Wait a minute, I'm going to show you how." Going on the other side of the spectrum, tell me a little bit about how data cloud can help complete this picture?

Raveesh Raina :

Yeah, absolutely. So as we all know, Data Cloud is an amazing tool that unlocks an organization's ability to tap into their enterprise data. So what that means is that for an organization that is using Salesforce, they may have a good portion of their data sitting in CRM, sitting in Salesforce, but a lot of it could also be sitting off platform in, for example, a lot of organizations, they have a separate dedicated app for their orders or their accounting, or even if they have a public website or mobile app, then some of that web traffic is logged in those third party systems.

Josh Birk:

Yep.

Raveesh Raina :

So through the power of Data Cloud, if they want to bring those insights and have a summarization or generative AI information associated to that data, customers are able to take all of that third party data that is sitting in those systems and have that surfaced up in Data Cloud. And the power of Prompt Builder allows for grounding with those Data Cloud objects or data model objects, I should say, and then surface up that information directly within the flow of work in Salesforce.

Josh Birk:

Yeah. Now, let's take a couple steps back because we've been talking a lot about data and grounding the data, taking the data from sources like Data Cloud, and I think that a lot of people have cautionary tales about AI. Describe the Einstein Trust Layer to me and how does it keep all of this stuff safe?

Raveesh Raina :

So one of the biggest messages that we have is the fact that we have zero retention policies with all of the public AI model providers that we have partnerships with. Whether those are, as an example OpenAI. So what that means is that I think we have to also take a step back and appreciate the fact that Salesforce's number one value is trust. And what we mean by that is the fact that we value our customers, we value that they are trusting us to securing their data. And we also take it one step further in ensuring that we are not sharing their data in any way, shape or form with any of these third party AI model providers.

So a big part of the Einstein Trust Layer is the fact that we are not sharing customer data directly or indirectly with AI model providers there. And also at the same time, those AI models are not learning from Salesforce's and customers' data as well.

Josh Birk:

And to be clear there, it's not like some generally handshake between us and Sam Altman. We filter that stuff out. OpenAI does not have the chance to train their models based on our data because they never see it.

Raveesh Raina :

Correct. Absolutely.

Josh Birk:

Okay. And then just to bring this to another level, because we think about things like in P-term, P2 and privacy and stuff like that, that's also stuff that's not being shared with entities that don't need to see it.

Raveesh Raina:

That is correct. Yeah, absolutely. We as a company have gone the extra mile to ensure that PII, personally identifiable information, or PHI, that is not shared, and it is actively masked and obfuscated before it is even sent to an AI model for processing as well.

Josh Birk:

Got it. Now, you've written a blog post, Ultimate Prompt Builder Guide, thank you for that, it's very extensive, and in it you include some very good tips on how to effectively write a prompt. Give me the high level like what are some really good tips to let me be more effective with that?

Raveesh Raina:

Absolutely. So a couple of pointers that I will suggest as admins and customers are looking to design prompts are that there are a few different areas that you should try to focus on. First is be explicit in your instructions in terms of the expectations and goals as far as what is it that you're hoping to get out of the response. So making sure that you are setting those expectations. Second is contextualizing information, meaning that state if you are adding in some related records to further contextualize the set of instructions. Third would be to specify your role. What is the persona for which the prompt template was built out for, whether it was a customer support manager or an account executive?

An important aspect of the prompt template that I would suggest is also ensuring that you add limitations and setting boundaries, so by being explicit with instructions such as, "Do not exceed past 500 characters," as an example. That is an explicit instruction and a guardrail, ensuring that the prompt template is not going to give a response longer than 500 characters. So those are some areas that I would suggest that folks go in on as far as designing prompt templates.

Josh Birk:

Yeah, one of the earliest things I realized in working with AI was, and this is in general, AI has this sort of default setting, shall we say, like a default tone, a default style. It thinks its response should be two paragraphs long, paragraphs should be three to four sentences. It's almost exactly what we'd learned back in high school for what a good writing style was kind of thing.

Raveesh Raina:

Right.

Josh Birk:

And it will default to that all the time unless you tell it not to. My favorite thing what you just said is explicit.

Raveesh Raina:

Correct.

Josh Birk:

I would actually add to that explicit and occasionally repetitive.

Raveesh Raina:

Yes. Yes, absolutely. It doesn't hurt being repetitive with your instructions because it's about ensuring that the AI model is going to meet your expectations of a response that has both value and trust.

Josh Birk:

Right. Because this is dynamically generated, which means even if you see the thing that you wanted to do once, run it three more times.

Raveesh Raina:

Correct. Exactly.

Josh Birk:

Because it might break one of your rules at one of those points. And then one of favorite anecdotes I give people is, if you saw the Dreamforce '23 keynote, we use Prompt Builder to do a sales email, and to have some fun, we added emojis to the sales email, which worked about three times out of five.

Raveesh Raina:

Yes.

Josh Birk:

Right? The team kept on coming out, they're like, "How do you fix this?" I'm like, "I have no idea. I'm going to go talk to the AI and figure it out." What it turned out was I had to beat the AI over the head and be like, "Not just use an emoji with this email, make heavy use of emojis all through this email." Be very, very explicit.

Raveesh Raina:

Absolutely, [inaudible 00:03:52].

Josh Birk:

Now, unlike myself, you have actually had the advantage of working directly with customers on some of these issues, and feel free to nerd out on fintech and banks in general, but what are some specific problems that you've seen from customers? Obviously, if you need to not name names, that's great, but what are some specific problems that you're seeing Prompt Builder help solve?

Raveesh Raina:

Yeah, absolutely. I think one of the reasons why I have been able to talk to a lot of customers is just because I've been also listening to a lot of customers as well. I think one of the most glaring areas where generative AI and Prompt Builder can provide most value is the use case for meeting prep. So I'm a rep, or I'm a customer support manager, and I need to prepare for an upcoming engagement with a client for a meeting. Using generative AI, I can quickly get a succinct overview on a particular client with all of the different data points that has both value and meaning to me, which I can use as part of my conversation in that meeting with that client as well. So meeting prep or account summarization is of immense value where we are seeing a lot of success with sharing this value with customers, and they are absolutely head over heels signing up for using Prompt Builder to address this need.

Josh Birk:

Got it. Well, to learn more, everybody, we're going to point to Raveesh wonderful blog post. We'll point you to a couple of Trailhead resources, some help documents. Raveesh, thank you so much. I do have one final question for you. What is your favorite hobby?

Raveesh Raina:

My favorite hobby at the moment right now is cycling. I use that as an opportunity to bond and to spend time with my kids. I have a seven-year-old and four-year-old, so any chance we get, especially with the weather that we're having this summer, we go outside, we go for cycling around the trails and around our neighborhood. So yeah, right now it's cycling for me.

Josh Birk:

Wonderful. Raveesh, thank you so much for the great conversation and information. That was a lot of fun.

Raveesh Raina:

Thank you very much, Josh.

Josh Birk:

I want to thank Raveesh for sitting down with us this week. And as always, I want to thank you for listening. Now, if you want to learn more about this show and stories about being a Salesforce admin in general, go down over to admin.salesforce.com. Thanks again, everybody, and we will talk to you next week.