Firm Management
Chat GPT and Prompt Engineering Tips – The Accounting Technology Lab Podcast – August 2024
Brian and Randy discuss how to get the most out of Chat GPT and other generative AI by using certain strategies in your prompts.
Aug. 16, 2024
![](https://www.cpapracticeadvisor.com/wp-content/uploads/sites/2/2024-Accounting-Tech-Lab.png)
Hosts Randy Johnston and Brian Tankersley, CPA, discuss how to optimize generative AI functionality like Chat GPT and CoPilot, and how learning how to ask the right questions or give the right directs (or prompts) to these language generators can greatly enhance the quality and accuracy of your output. Some great tips and tricks.
Use the podcast player below to listen to the podcast.
–
Transcript (Note: There may be typos due to automated transcription errors.)
Brian F. Tankersley, CPA.CITP, CGMA 00:00
Welcome to the accounting Technology Lab sponsored by CPA practice advisor, with your hosts, Randy Johnston, and Brian Tankersley.
Randy Johnston 00:10
Welcome to the accounting Technology Lab. I’m Randy Johnson here with my co host, Brian Tankersley. And you know, everybody’s becoming an AI expert. I just can’t turn around without somebody else claiming they’re an AI expert. Oh,
Brian F. Tankersley, CPA.CITP, CGMA 00:23
my God. It’s like social media experts were 15-20 years ago. You know, everybody, everybody that graduated from high school suddenly was an enterprise a enterprise, you know, social media expert. Yeah.
Randy Johnston 00:37
And making our listeners cringe. It’s like fingernails on the chalkboard. When I watch people talk about AI. And it’s like, oh, my, as my associate and friend George Takei would say, Oh, my. But today, Brian, and I would like to talk a little bit about Chat GPT and co pilot prompting or sometimes called prompt engineering. Now, there’s other platforms that we also can do this for. With Google’s Jim and I are clods, you know, anthropic Claude, all of those who are very interesting, but prompting is more about getting what you want. So Brian, give us an idea what we should want. Well,
Brian F. Tankersley, CPA.CITP, CGMA 01:22
you know, and I think you’ve got to decide, you got to decide how you want to handle it, you’ve got to decide, you know, there are a lot of things that you can get out of this, okay. So it’s a very powerful tool in like, like any powerful tool, like, you know, Microsoft Access or Microsoft SQL. If you don’t know what you want, when you sit down to do it, you’re not going to get what you want, you’re gonna get something else. Okay, so, so I want to talk first about prompt matters. Okay, so So and I’m gonna give you an example from my personal life, okay. My son, one of the is my son is, you know, lives with us. And he’s an adult. Now he’s in community college. And one of the ways he earns his keep is he does laundry, he takes the garbage cans to the street, he’s got a lot of chores, okay? And but unfortunately, we have to manage him to get those chores done. And so I want to talk about the prompts that I use with him, okay, because I, I don’t want to be too direct too early with him. But I rather want to, I want to encourage him to to so that he can get that sense of accomplishment that comes from taking care of things yourself. Okay? So so the prompts absolutely matter. Now, for our perspective here Monday morning is garbage day, Sunday at 5pm. I say to my son, hey, tomorrow’s garbage day, don’t forget to collect everything and put it in the CANS on the street. Okay. Of course, he keeps playing his video game and ignores me. Seven o’clock, I look out what’s the plan on the garbage cans, okay, nine o’clock, I look out at the camera that you’re actually looking at the view from it’s on the corner of the house. And it’s they’re not there. And I say garbage cans now. Okay. Now, if I have to go to Monday morning, I get his phone and his headphones, and I wake him up from a dead sleep. And I say, cans are still not on, you’re still on the street, I have your phone, I’m going to hang on to it until it gets done. I start reading text messages in 10 minutes. I’ll wait. Okay, and, and so he doesn’t want any part of me getting in the middle of that. And so I just kind of I mentioned that to you here because again, in this context, my son and I have a history on this, okay. And if I come in, if I come in with that last that last final, you’ve got to do this, or I’m gonna be a tyrant. He’s gonna think I’m a target. Okay, and so that’s so I’d rather do this, I know, he’s in the middle of a video game, nobody’s gonna die. If he doesn’t put it out until 829. Monday morning, you know, but it must get put on the street. Okay. And so the context is critical in here to how I do the prompt, okay, to what I want, and what is really required. Okay, and so I think this whole, you know, so putting the context in place here is important because all of these large language models learn from statistical relationships between text. So the more context you give, the easier it gets. Now, there is a conflict between confidentiality and usability inherent, okay. And so we have to be very careful about when we’re using these tools, because we don’t necessarily want our interactions to be retained for use use in future content. Okay,
Randy Johnston 04:35
Brian, I’m going to stop you there for just a minute because I wanted you to tell your story uninterrupted with the sun and the garbage cannon. And, you know, I know Brian and I know his son, I’ve known him since he was born. And so you know, the son is actually pretty responsible, but he does, like plays games. And, you know, you heard Brian say, I’m going to read your texts. You know, I’m thinking about confidentiality there as you were saying that Brian, and Brian’s not much of a hard nosed, he’s about as laid back as he can get. But you know, when when he gets aggravated, he’s aggravated. And so
Brian F. Tankersley, CPA.CITP, CGMA 05:08
the problem is you have, when Mama’s not happy, nobody’s happy. And if those garbage cans do not get taken away, Mama is hard to live with.
Randy Johnston 05:18
Understood. So I just wanted to kind of frame this, you know, Brian and I are both pretty laid back in our lives and career. And we’re pretty lucky with all of our families. But the story here as it continues to unfold, you’ll see why it’s so germane to the prompting that has to happen with AI. So I’m going to turn you loose again, as you’re talking about Groundhog Day and confidentiality. Now,
Brian F. Tankersley, CPA.CITP, CGMA 05:43
again, there’s a conflict between confidentiality and usability. Okay, so it is harder when you’re starting out with jet GBT and these other things, because you don’t know what you want. And you don’t know what the privacy is, we’re going to do a whole session on privacy, that I think is going to be our next episode of the of the technology lab. And we’re going to talk about this, and I’ll tell you, which plans have privacy and which plans don’t and which copilots have privacy and which ones don’t. But these large language models are very good at surfacing and summarizing content, okay, so you don’t want to put confidential stuff into them. That especially stuff that’s regulated under regulations like HIPAA, or FERPA or, you know, for educational institutions, or, you know, again, GLBA, or, or, you know, getting anything tax related, you don’t want to put into these things, unless you know, that the data is not going to get used for training and here, okay. So I want you to get here that this, this conflict between confidentiality and usability is really important. Because there are things that you can do from a usability perspective and problems you can solve with jet GPT, that you don’t want to, okay, great example of this bank, turning bank statements into CSV files, okay. You can absolutely feed images of bank statements into into chat GPT and have it create CSV files from it. Okay. The problem is, you’re putting all that information into Chet GPT. And potentially, that can be used for training purposes, and so forth. And if you don’t do it exactly right, you you, you may have may have some big compliance problems and data breach responsibilities to deal with. Now, let’s talk about prompts here, okay. And you know, the ultimate prompt is a pickup line. And so the as we’re looking at this now, this is a, this is a piece of this is actually a graphic that I created here. But again, it’s a way of starting a conversation. And again, there are threads, and you go back and forth, like a chat with, with chat GPT. And these tools. So, you know, again, with prompts, it’s like the pickup line into singles bar or an information or again, giving guidance, you know, the more you know, you’re what you’re really trying to do in a singles bar, when you’re meeting, meeting those folks is, if you’re looking for that for life’s next partner, you’re actually trying to start a conversation, and then to figure out if you have anything in common and to qualify this person and see, do they want the same things in life I want, okay, so it really is a sales process. As we look at this now, this is a graphic that I had it create for me. And here’s the prompt I used, okay, create a drawing, which shows a man using a pickup line on a robot who looks like a woman at a bar, the bar should include zero and one characters in a waterfall similar to the matrix movies, and the woman should be drinking from a can labelled data. Okay, so this is the first attempt in here. And one thing I’ll say about this is that I thought it was actually a pretty good, pretty good thing, you know, the woman seemed to be a little bit overly sexualized the robot. But, you know, other than that, you know, I thought it was a particularly good, particularly good graphic now, one of the things you got to learn about prompts is that you could submit the same prompt over and over to chat GPT or to copilot and we’ll come up with slightly different answers every time. So now I came through and I said and the next thread in the conversation create another version except the man and the robot should both be accounting professionals and look at that they both have on jackets okay. And so so there it is, now. Then I did another one in here. And in this one I said create another version of the graphic for the robot in the person are sitting at the bar behind them as wall of liquor bottles, all labeled data in the background behind the characters. And you can see by the way, that not it didn’t spell data right on any of those individual liquor bottles. If you zoom in on him here, some of them say DTA and da, da, ti di, you know, again, it’s it’s like they’ve already started drinking when you put spell the stuff on there something but the shelves again, have the labels and then I describe some other things in here. Notice it didn’t do some of the things I asked it to do. So I asked it to put in there web crawling social media scraping and confidential financial information, labeling the levels of bar. And it didn’t do that. Okay, I had an idea. And it just said, No, I’m not going to do that. Then I ran some additional ones in here, some of the other ones, again, that it came up with notice the GPT, rule Street and things like that. Here, here’s another one we have, I then took all of the previous guidance in here that I had. And I said, Okay, let’s create the one to the one to rule all all answers. And so I had this huge prompt in here. And I want you to see this, because there are a few lessons for you to take away from this. Okay, the first one is, this is an iterative process, okay, so you’re going to try something, and it’s not going to be exactly what you want, then you’re going to figure out how to do it differently. And then you’re going to, you’re going to ask it to revise and do things do things differently. Okay. The second thing here is that you may want to save the prompt that you create. And I actually have a OneNote notebook I use with this, where I, when I have a prompt that works particularly well, I copy and paste it and throw it into OneNote. So I can search it and use it later. Now, so you can see it created this graphic. Now, here’s the and again, this the same one in here, I then said, Hey, I really liked that. And I said, Okay, now I need you to make it in a in a widescreen format. And here is the widescreen format of it that it did notice it was significantly different than the previous one. But what I want you to get about this is that what it’s really doing for you is it’s it’s really giving you ideas, okay, and you know, and by the way, if I didn’t like this one, I would have made 10 other ones, and then picked the one out of them that I liked. Okay? Again, you have to note here, when it comes to the graphic creation stuff in here, even on the most modern version of this, the spelling, there will be typos in the spelling, okay. And again, if you look at these, it’s not quite lorem ipsum, you know, those that texts they put in when they’re demonstrating fonts, but nothing is spelled correctly in here. So if you’re particularly OCD about spelling, like many of us are in the accounting profession. That’s that’s something to note. So. So anyway, we’ll talk more about prompts here in a second, Randy, but any comments on that on that the store? Can’t hear you.
Randy Johnston 12:20
So sorry, what I wanted our listeners to hear and understand is Brian was going through a series of prompts trying to tune up the conversation. And you can teach these prompts to your teen. And you can tune up your prompts, and you can save your prompts. And it’s going to take a little bit for the engine first to know who you are and what you’re like in chat GPT, for example, there’s a place where you can tell it, you know, your preferences, and you should do that. But over time, you will find that the engines will actually become a little more refined as you go forward. So, in effect, prompt engineering, the way you talk about the prompts is important. There’s debate in the marketplace, if politeness gets better results or not, or whether brevity gets better results or more completeness gets better results. And you know, my experience on that has been it just seems to vary over time, even within different AI engines. So, you know, just because of my personality, I tend to be a little more polite, the way I asked things, but I also tend to be kind of brief. And I think the engines do a little bit better with more detail. And I tend not to supply that detail early on, I want to get it in the zone first. So just like the first prompt that Brian said that to the the female, the borrower was little over sexualized, that was probably true. And you know, he might have been more specific, if he had thought about it saying, you know, a very smart female, that
Brian F. Tankersley, CPA.CITP, CGMA 14:08
attractive, but conservatively dressed, you know, ugly dress, those
Randy Johnston 14:12
types of things would really help. But the key here is as accountants, we have to be able to specifically say what we’re looking for, and we have to recognize that the accuracy of certain things are still going to be messed up. But all of these things once you learned how to use the prompts, kinda like you’ve learned how to use search engines now that they’ve become a lot more popular. You kind of know how to search to get what you want, but my bet is that most of you aren’t very refined that you’re searching. I know many people that are very refined it they’re searching to get much better results much more quickly. But again, I tend to be kind of loosey goosey on the searching too. I know a few tricks that helped me but I’m not really really, really refined on my search engine prompts either.
Brian F. Tankersley, CPA.CITP, CGMA 15:05
And I think one of the one of the things about this iterative, this iterative thing with prompting Is that is that generally, we don’t know exactly what we want, we know we have an idea of kind of the general territory, you know, when I was doing those graphics that I was that I showed you there for, for example, and that was actually doing them for this presentation, because I wanted to have those, you know, that conversation as a graphic. However, I didn’t know exactly what I wanted until I got through here. And so that’s in here. Now, I want to remind you here that large language models are different from the other other things you’ve done with computers in the past, okay? In the past, when you when you did, when you had ran Excel or you ran, you ran a report out of QuickBooks, or you rent, you created the document in Microsoft Word, it had the features to go through and handle all of those things, all of those things that we had going on. So it would have it would make sure that everything was absolutely accurate all the time. And so you need to understand with respect to generative AI, that it will not only be an accurate, okay. And that’s one of the reasons I pointed out the spelling on those on those graphics, it will not only be an accurate, sometimes it will just make up things out of whole cloth. Okay, so we were running bios on people through Chet GPT, just to see how well it would create bios for for a conference. And we ran a bio on our friend Doug Sleater. And chat GPT 3.5 told me that Doug was dead. Doug thought that was hilarious. Okay. But I’m share it with you here. Because you have to understand that this is not like a traditional when you’re doing stuff with with AI. This is not like a traditional computer application where you can trust that it is absolutely correct. And that the data you put in is absolutely what came out. Notice that when I asked it to do certain things, it just ignored me altogether. Okay. And in other places, it just made up new things that I didn’t ask it to do. Okay, so
Randy Johnston 17:10
just a reminder, in prior Technology Labs, we’ve talked about hallucinations. We’ve talked about what different language models can do. Today, most of our context is around Chad GPT, and Microsoft 365 co pilot, which happened to be the same basic engine, but large language models. So this idea of making stuff up, you can learn more about that in a prior accounting Technology Lab. But Brian, you know, as you’re thinking about this, Microsoft has actually put out pretty good guidance on copilot where they suggest that the best ingredients and through the years you’ve used lots of cooking analogies in our presentations you’ve created you did that for business intelligence here recently as an example. But they suggest that you start with a goal, what output you want from co pilot, the context, why do you need it and who’s involved? The source which information sources or samples should copilot use? pointing to the right sources could be good. And for accountants going to authoritative sources is quite good. Now in other technology labs, we’ll talk about the tax research tools, for example, Thomson Reuters, edge, Koch, co counsel, or we’ll look at what Wolters Kluwer has done with the answer connect in their AI additions, but they are authoritative sources, which is interesting. And then of course, preferences, you know, what do we want COPPA do to meet our preferences? So that those ideas, goal, context source and preferences is really solid guidance.
Brian F. Tankersley, CPA.CITP, CGMA 18:51
Yeah, and absolutely, Randy. And I think that’s, that’s the critical piece here. Now, this is open AI also has a guide to prompt engineering, okay, and I’ve got the link there. But again, you can, you can actually, you can actually, again, go out to the open AI website and see this. But their their principles are to write clear instructions, provide reference text, split tasks into subtasks give the model time to think use external tools and then test changes systematically. Okay? And so there’s there’s much more in that particular guide. But there are also an open it there’s an open AI cookbook that talks about different things you can do and different kinds of prompts you can create. And there are many prompt examples out here. So if you go out now to for example, this this open AI tool in here, you can see that they have different ideas of things here. You know, when when the reason I do the recipes, Randy is that when my mother sent me off at the college, when I got my first apartment, one of the gifts one of the housewarming gifts she gave me was A book that I still have in my kitchen called The Joy of Cooking that had a lot of recipes. Okay. And so I would just encourage you to look at these prompt examples. And in other pieces like this, I mean, look at this rap battle rider. Okay, you know, I’ve never needed that in my entire life. And in all the years all those laps around the sun that I’ve done, I’ve never needed a rap battle writer. But look, there’s an example of one if I need it. Oh, look an emoji Chatbot. Okay, so more of the sarcastic Chatbot. Okay, so I’m just kind of showing those to you here. Because I think it’s important that you look at what other people do the same way that, uh, you know, again, I didn’t decide I was going to make I was going to follow Julia Child, one of the great masters of cooking. But But I decided that I was just going to try out things with effectively this recipe book.
Randy Johnston 20:50
Yeah. And you know, as you’re thinking about these prompt libraries, and you showed a lot of the pictures in the video version of this podcast, the results. But it turns out that we’re going to have aI solving all sorts of situations. For example, a new generative AI for music is called sudo, su N O, and it just generates music from prompts, like Brian’s been teaching you. And that is amazing, to me, simply amazing. Well, we will let you continue with Prop libraries.
Brian F. Tankersley, CPA.CITP, CGMA 21:23
So with libraries, I have a, I have a spreadsheet actually, that I use for these. The idea here is that you can save these into, you can save these, you can create templates for you here, I’ve actually got a spreadsheet, I’d be glad to send to you if you want it. But the idea here is I’ve taken those gold context source preferences, and then I write a sentence for each of those. And then it feeds it together. And then I can copy and paste that into the chat GPT to have that creative prompt. Another one we have here. This is another framework from somebody else. And so there I’ve got a list in this of different roles, I have a list of results I want, I have a list of explanations and audiences and tones and explanations. Again, there, you can customize this and make this wherever you want. But the idea that I’ve got here is I’ve got it set up where you can just pick these from the list from the drop down, and then it will go through and actually revise and create a sentence. So you can see we have portfolio manager right now, let’s just trade this one out. And let’s pick a different context. Let’s say you are an accounting professional. And then the result that you want to have here is, is going to be, let’s go down here to edit. Let’s make this. Let’s make this explain. Okay, and so now notice that it has gone through here, and it’s created these six sentences in here. And I know that this seems overly simplistic, but this is an approach to get the conversation started. Okay, so now that GE pickup line generator, yeah,
Randy Johnston 23:03
so Brian, at the risk of stepping on you there, I just wanted to make sure our listeners knew that you offered to send this prompt template, it’s a well done prompt template, it’ll save you hours of creation, we’ve been talking about, uh, you know, using templates, and we’ve got complex ones, this one is a simple starting point for you. And I would encourage you to make that request to either,
Brian F. Tankersley, CPA.CITP, CGMA 23:30
let’s let’s do this. I’ll I’ll actually hang it off of my website, a CPA tech blog.com. I’ll upload it to this before this, this thing goes live. So you can download it from there. And I’ll also send it to Isaac so he can put it with the magazine.
Randy Johnston 23:49
That’s beautiful. Well, Brian, I know you’ve got more you could be teaching about prompting, what are some closing thoughts here, I know you’ve got a couple of more that you wanted to make.
Brian F. Tankersley, CPA.CITP, CGMA 23:59
Again, I think this I think this idea of using an Excel workbook or saving these things is important. You’re going to look this is this is like any skill you have, okay. And you know, my my son right now is, is taking English 101 for the second time and community college. Because he didn’t he’s not great at writing essays and expressing his thoughts and things like that. Okay. And it’s the same thing I’d say about prompting is the same thing I’d say about him, Okay, you’re not going to be good when you start, okay. But the important thing is to start, okay, you’re gonna start out and you’re gonna make some really bad sausage, okay, you’re gonna get closer to filet mignon as you develop your skills, but I want you to get that, that when you’re creating these, these prompts in here, you’re gonna get better at it and then you’re gonna come up with prompts that work for you, and they’re gonna work great. Okay, and then open AI or Microsoft is going to have changes to the model, and you’re gonna have to change them up. Okay, so I want you to get that We’ll change around so, but that’s your that that’s what I just suggest your you know, and again, St. I’d say the same thing to you that I would say to him, you know, it doesn’t matter how bad the first one is, it just matters that you get started. Okay. So Randy, I’ll let you close this out here. You
Randy Johnston 25:18
know, we’ve I’ve let you run today for a lot of different reasons. Because, you know, I knew that your preparation in this area, your research in this area could help all of our listeners, the template that’s been offered, also is a big gain, I’m going to tell you will continue to refine our knowledge and pass it on to you in this area. And again, watch who you’re listening to on AI out there. Because unfortunately, there’s too many social media gurus and snake oil salesmen trying to tell you that you can have whiter teeth and fresher breath if you just use their AI methodology and
Brian F. Tankersley, CPA.CITP, CGMA 26:01
nobody’s going to pimp out your private data. But you need to pay very careful attention to that. Because you cannot take back that private data that you put into the models. And there are there’s research that shows that the there’s memorization that happens with the training that that really can can be devastating to your ability to keep things private. Thank you for sharing your time with us. We’ll be back next Saturday with a new episode of the technology lab, from CPA practice advisor. Have a great week.
= END =