Prompt Design's Potential, with Perpeta's Paul Carney

[00:00:00] Kundenservice kontaktieren? Für viele Menschen ist das der beste Weg, einen schönen Tag zu ruinieren.

[00:00:06] Aber bei Zendesk sorgen wir für eine bessere Customer Experience.

[00:00:09] Besser für die Großmutter, besser für die Floristin, besser für den jungen Mann in Hausnummer 3A, besser für sie, besser für alle.

[00:00:17] Denn während einige behaupten, dass der Kunde immer Recht hat, sagen wir, dass KundInnen immer Menschen sind.

[00:00:22] Und da wir auch Menschen sind, wollen wir etwas Gutes für uns alle tun.

[00:00:26] Zendesk. Customer Experience mit KI für Menschen gemacht.

[00:00:29] Brauchen wir wirklich noch einen Computer? Alle wahrscheinlich nicht.

[00:00:33] Aber wenn du Musik mit der Power eines Neuralprozessors neu erfindest oder unterwegs Migrationsmuster mit einem ganztägigen Akku analysierst

[00:00:39] oder deine Ideen mit dem KI-gesteuerten Co-Creator zum Leben erwächst, dann kann ein Copilot Plus PC einen Unterschied machen.

[00:00:46] Nicht alle brauchen einen leistungsstärkeren KI-Computer.

[00:00:48] Aber wenn du versuchst, die Welt zu verändern, auch wenn es nur deine eigene ist, haben wir einen für dich entwickelt.

[00:00:52] Microsoft Copilot Plus PC mit Snapdragon. Die bisher schnellsten und intelligentesten Windows-PCs.

[00:00:57] Die Akkulaufzeit variiert hier nach Nutzung und Einstellungen.

[00:01:00] Welcome to PeopleTech, the podcast of WorkforceAI.news.

[00:01:04] I'm Mark Pfeffer.

[00:01:18] Prompts set up real conversations with AI, if you know how to craft them.

[00:01:23] My guest today is Paul Carney, the founder of Perpetta.

[00:01:27] He's a CHRO who knows how to create prompts that get detailed results out of AI.

[00:01:34] There's a lot more to prompting than you might think, and that's what we're going to get into on this edition of PeopleTech.

[00:01:40] Hey, Paul. Welcome.

[00:01:44] You have been doing a lot of work with the prompts used to run AI searches, queries.

[00:01:52] Can you tell me about prompts and how did you get to be so knowledgeable about them?

[00:01:58] So, you know, it's funny because I'm a chief human resource officer for a bank also, which people always say, well, why is a chief human resource officer, you know, someone that's dealing with AI?

[00:02:06] I've logged over 1,100 hours with AI, but I'm a techie guy by trade.

[00:02:10] When I saw my first web browser, I said, this is going to change the world.

[00:02:14] And then, you know, it did.

[00:02:16] And then when I saw ChatGPT just about two years ago, I said, this thing's going to change the world again.

[00:02:23] And that's what got me into really delving in.

[00:02:25] And what happened was I used my technical and HR knowledge to say, what can this do for us?

[00:02:30] And as I started talking to my HR peers, I realized they had no idea this thing was coming.

[00:02:35] And someone had dabbled and done stuff.

[00:02:37] So I got on this path to say, I need to help HR people not only understand what this can do for them in their career, but they've got to guide the entire organization, policies, training.

[00:02:47] There's so much that they're going to have to understand that if they don't understand it themselves, it's going to be harder to guide your business partners to say, this is what it can do for us.

[00:02:56] So let's talk about prompts themselves, because I have this impression that a lot of people, especially in the upper reaches of the corporate world, they think of prompts as kind of like a simple search query.

[00:03:11] But it can really be a lot more than that, can't it?

[00:03:15] Absolutely.

[00:03:16] In fact, you know, when I show people the difference between what I call verbose or narrative prompts, where you just kind of write your thoughts and ideas and give a few instructions, I show a versus a structured prompt where you say, role, I want to assume this role.

[00:03:31] Here's my request.

[00:03:32] Here are the instructions, the output format.

[00:03:35] Here are some other things like keep it to this length, what I call snippets.

[00:03:39] When you start to format it in a little bit more constructive way, you get dramatically different results.

[00:03:44] The results change dramatically because when you just kind of write this verbose prompt, the AI is scanning all those words to figure out what do you want?

[00:03:54] What, what, which pieces and parts of this, and it'll start making assumptions.

[00:03:58] And the better it does at making, when you get it to stop making assumptions and show it exactly what you want, you drive it towards the path of, hey, this is what I want you to focus on.

[00:04:08] So that prompting technique is what I discovered just earlier this year as I started working with people that I broke it down into, you got to have templates, you got to have profiles or personas.

[00:04:18] The better you do this, the more you can reuse these things again and again, too, is just a powerful thing.

[00:04:25] So you're not recreating it every time.

[00:04:27] And then share with each other.

[00:04:28] You can share these reusable components with each other.

[00:04:30] So that structure, and again, a lot of people refer to it as prompt engineering.

[00:04:35] I like to call it prompt design because prompt engineering is designing prompts, but it's the technical people using those to train the large language models.

[00:04:44] They're, they're prompting it based on the data to say, here's how to respond to that type of data.

[00:04:48] That's a very deep technical, and the word engineer tends to scare people away anyway, going, well, I'm not an engineer.

[00:04:54] But when you call it prompt design, they feel more comfortable.

[00:04:57] They understand this is about me, the consumer, putting in the inputs to get the better outputs.

[00:05:03] So what do these prompts look like?

[00:05:06] I mean, they're probably not a couple of sentences, you know, providing detail on the topic.

[00:05:14] I'm assuming they're more organized than that.

[00:05:17] They're more organized.

[00:05:18] You know, my typical prompts can be two pages long, and that's why I store them offline.

[00:05:23] So I kind of put them all together, and then I copy and paste that into my favorite AI engine.

[00:05:27] Because if you try to type all that in the AI engine, it's going to get lost.

[00:05:30] And if you're like me at ChatGPT, when you hit the return key because you wanted a carriage return, nope, it submitted it, and you're halfway through your thought.

[00:05:37] And I do it.

[00:05:38] So now I do it offline, get all my thoughts together, and paste it in, say, now go.

[00:05:43] So yeah, so the prompts tend to be very structured.

[00:05:46] Like I said, the first thing I do is define the role I want it to assume.

[00:05:50] If I want it to be an expert marketing person, if I want it to be an instructional designer, whatever it is I'm the background I want it to be, that's first.

[00:06:00] And then the second thing is the request.

[00:06:02] Here's what I'm asking you to do, and it's very short.

[00:06:04] It's not very long.

[00:06:05] And then all of the extra context, instructions, examples.

[00:06:09] I'll give it examples, like if I'm asking it to write me a LinkedIn post on a topic, then I will say, here are some examples of my writing style.

[00:06:18] Make it look and sound like me.

[00:06:20] So that's the type of stuff that you can do.

[00:06:22] And like I say, where you can store these things separate and reuse pieces and parts so that you can kind of pull them in as needed.

[00:06:29] And that's the other power of the efficiency is being able to not have to recreate these things every time.

[00:06:37] And how do you go about forming these?

[00:06:39] I mean, this doesn't sound like something that you just sit down and dash off.

[00:06:43] It sounds like you've got to put some thought into it.

[00:06:46] You do.

[00:06:47] You know, so I use this description that if you're going to go have a difficult conversation with someone,

[00:06:51] you generally don't just walk into the conversation and just start talking like you would in a verbose prompt.

[00:06:56] You probably have structured your thoughts and put together, here's some of the points I want to make.

[00:07:00] Here's some of the things they may say that I can have counters ready for, that type of thing.

[00:07:04] It's no different with a prompt.

[00:07:06] So really, prompting should be a well-thought-out process before you really get to it.

[00:07:10] But here's what's cool about AI.

[00:07:12] You can actually help it, have it help you do what you want.

[00:07:17] So for instance, say you say, you do believe me and say, I want a role, request, and then maybe some context and instructions.

[00:07:24] And you can say to ChatGPT or one of the AIs, you can give it a little prompt and say, hey, I want to really discuss about this topic.

[00:07:32] What role would you be, what would be best for you to assume to help me understand this?

[00:07:37] And it comes back and says, oh, I should be a blah, blah, blah, a detailed marketing expert.

[00:07:41] Great.

[00:07:41] So then you say, role, detailed marketing expert.

[00:07:43] And then you say, you know what you're recording.

[00:07:45] And then you can ask it, and I've done this many times, what information do I need to give you to help us have a conversation about this topic?

[00:07:52] And it'll say, oh, here's some factors you should incorporate.

[00:07:54] Great.

[00:07:55] So the instructions now say, please be sure to include this, this, and this.

[00:07:58] So you can have it be like a brainstorming partner to help you put together the prompt that you can then use.

[00:08:05] Oh, interesting.

[00:08:06] Isn't that cool?

[00:08:07] Yeah, that is cool.

[00:08:09] You know, really, it gets toward the notion of conversations with the machine.

[00:08:13] Correct.

[00:08:14] Right.

[00:08:14] Which is another whole issue, you know, that I think we talked a little bit about before, is I'm starting to get into this as digital colleagues start to show up in our Zooms and our Slacks and Slack channels and every place.

[00:08:25] People have got to start to get comfortable with these entities.

[00:08:28] They're not people, but getting comfortable with these digital colleagues.

[00:08:31] And that's another, because in the end, it doesn't matter what it is, a chat bot, a Microsoft copilot, or in ChatGPT.

[00:08:38] It all comes down to how you interact with it, what you ask it to do and how you're asking it.

[00:08:42] And that prompt really matters.

[00:08:44] That's why I say sometimes people are going to get a little too casual because it's chatty with them, but you still got to get back to, I got to structure my thoughts because it'll go off in directions you didn't expect.

[00:08:54] I want to take a break real quick just to let you know about a new show we've just added to the network.

[00:09:02] Up Next at Work, hosted by Gene and Kate Akil of the Devon Group.

[00:09:08] Fantastic show.

[00:09:09] If you're looking for something that pushes the norm, pushes the boundaries, has some really spirited conversations, Google Up Next at Work, Gene and Kate Akil from the Devon Group.

[00:09:24] Hi there, I'm Peter Zollman.

[00:09:26] I'm a co-host of the Inside Job Boards and Recruitment Marketplaces podcast.

[00:09:31] And I'm Steven Rothberg, and I guess that makes me the other co-host.

[00:09:34] Every other week, we're joined by guests from the world's leading job sites.

[00:09:38] Together, we analyze news about general niche and aggregator job board and Recruitment Marketplaces sites.

[00:09:44] Make sure you sign up and subscribe today.

[00:09:48] What kind of output are you looking for or does it really depend on the query?

[00:09:54] I mean, are you looking for a lot of pages with a lot of detail or a simple answer?

[00:10:01] So, and that's one of the powerful things I teach people about the output formats.

[00:10:05] Don't just ask for text, first of all.

[00:10:08] You know, if you're asking it to compile a bunch of information, you can say, give it to me in a comma-separated file that I can then download and go take and put into Excel.

[00:10:16] You can ask it to put it into a type of table.

[00:10:20] Like if you're asking it to compare the pros and cons of these two types of things, tell it to give it to you in a table.

[00:10:25] And it'll give you a beautiful table that you can copy and paste into Word.

[00:10:30] Markdown is another powerful thing.

[00:10:31] So, if you're asking it to give you sort of an outline with headers and subheadings, do it in Markdown because you can take Markdown and paste that in and you keep all your underlines and bolds and all that type of stuff and bullets.

[00:10:43] So, there's a lot of ways you can ask it to give you executive summaries, bulleted points, outlines.

[00:10:50] You can ask it and it will pretty much give you what you need.

[00:10:54] So, how did you figure all this out?

[00:10:57] Now, there have been a lot of training sessions or books about this that I've seen.

[00:11:03] So, there are some training.

[00:11:04] You know, so I went out in September of 2023, about nine months into playing with this AI and got certified.

[00:11:11] I looked at the Blockchain Council, which is an international organization, and took a prompt engineering course and a prompt expert course and became certified in both.

[00:11:20] Because I said, I need to understand this using my technical background to understand the technical, but then to be able to bring that to the top for an HR person who doesn't want to be technical.

[00:11:30] So, that's what I did.

[00:11:31] And then over 1,100 hours of logged on these engines, playing with them, challenging them, getting frustrated with them, expressing my frustration to them, which causes them to do things that I didn't expect.

[00:11:43] Like, one time it said, let's step back and start from the beginning.

[00:11:46] And it asked me a question.

[00:11:48] I'm like, whoa, where did that come from?

[00:11:50] And we went down a few questions and we solved the problem.

[00:11:53] But we've been going in circles for so long.

[00:11:55] So, you know, it's really pushing it to a limit that most people aren't doing.

[00:12:00] That's the thing is, there's not much I won't say or do to these AIs.

[00:12:05] I wouldn't necessarily interact with a human being that way, but you've got to move past that because it's not a human.

[00:12:10] It doesn't have emotions.

[00:12:12] You know, this just occurs to me, but what's the most surprising thing you've ever seen an AI do or reply to you?

[00:12:22] Yeah, I think it was that one time when I expressed major frustration at it and said, why is this so hard?

[00:12:29] This is a simple issue.

[00:12:30] And then it said, I understand.

[00:12:32] It said, apologize, which it's programmed to do that.

[00:12:34] But then it said, let's step back and start from the beginning.

[00:12:37] That surprised me.

[00:12:38] I'd not seen an engine do that before where they had been programmed that when this level of frustration was there to step back and start asking questions again.

[00:12:47] I mean, that's a very intuitive thing to do, but I didn't know AIs had never done that before that.

[00:12:52] The other thing that was very surprising to me that really is pretty powerful is when I, it doesn't follow my instructions.

[00:13:01] And of course, I'm the type that's going to say, you didn't follow my instructions.

[00:13:04] And I ask it for advice.

[00:13:05] What could I have done differently or said differently?

[00:13:08] And it gives me.

[00:13:09] So one time I said, do not include whatever it was.

[00:13:12] And it included it.

[00:13:13] And I said, I said, do not include.

[00:13:15] Why didn't you?

[00:13:16] Oh, I'm sorry.

[00:13:16] And it gave me the thing without it.

[00:13:17] I said, no, what could I have done differently?

[00:13:19] And it said, use the positive negative.

[00:13:22] And then it explained it.

[00:13:24] Instead of do not include, say exclude.

[00:13:27] Because remember, these are pattern matching systems and it missed the knot.

[00:13:31] It saw do include and said, oh, include.

[00:13:33] But exclude, when it sees that, it knows that's the remove concept.

[00:13:39] So that was a very powerful lesson too.

[00:13:41] And that's the thing I teach people is have it teach you how to work with it better.

[00:13:46] And again, that's not a typical thing we do with humans.

[00:13:49] But you can do it like crazy with AI.

[00:13:53] Now, do you think most people have picked up on all this?

[00:13:57] Or are they going to ChatGPT and basically entering the same query they've entered for Google?

[00:14:05] I don't think many people have seen this at all yet.

[00:14:07] And that's why, because every time I speak, I always get, I didn't know it could do that type stuff.

[00:14:13] You know, when I show them a 40 line script you can use in any AI engine that will become an interview coach.

[00:14:21] So if you want to interview for a job, you can have it sit there.

[00:14:24] Give it the job description in your resume and have it act as the interviewer.

[00:14:29] And prompt you with questions how you should answer it stuff.

[00:14:32] You could do it as an interviewer.

[00:14:34] We teach leaders the same way.

[00:14:35] As an interviewer, here's a persona of an interviewee for my job.

[00:14:40] And give me answers to questions I'm going to ask.

[00:14:42] And then it gives you prompts to say, here's additional follow-up questions.

[00:14:45] And then in the middle of it, you can say, well, what if the candidate didn't answer that very well?

[00:14:50] Give me an answer that's not a really good answer.

[00:14:52] And then it will give you this sometimes really strange answer.

[00:14:54] And then suggestions on how you follow up to that.

[00:14:57] So, you know, these are things you show people and people aren't even seeing this.

[00:15:01] When I show these to people, they're just fascinated.

[00:15:03] They're like, again, I didn't know you could do that.

[00:15:06] I'm like, yes, you can ask it to do a lot of things.

[00:15:08] So is this the kind of thing that business leaders should be paying attention to?

[00:15:15] Not just encouraging their staff to do this, but maybe to provide training and education about it?

[00:15:23] Yeah. And it's not just the, you know, training on how to write a prompt because there's a bunch of stuff out there.

[00:15:29] It's really, again, back to what you just said a minute ago.

[00:15:32] Why am I doing this?

[00:15:33] If you don't have your goal in mind, it's probably just going to be a fruitless exercise.

[00:15:37] Now, you might just spend a lot of money and not get much return.

[00:15:39] But to get that true efficiency out of it, you've got to be able to structure those prompts, get better results, collaborate with people, share, reuse those prompt frameworks.

[00:15:50] That way people aren't recreating it all the time and can learn from each other and say, hey, this trick.

[00:15:54] Look, look, I did this.

[00:15:56] And people are like, wow, that's cool.

[00:15:57] You put it in a library and then someone else can use it.

[00:16:00] So that type of training, it's beyond just the how do you get in there, write a prompt and get results back from it.

[00:16:05] It's how do you get it to be efficient and more effective?

[00:16:09] My view on it is when business leaders see that, they're going to jump on that saying.

[00:16:13] And it comes down to training.

[00:16:14] It comes down to, in the end, that prompt, how you do that prompt.

[00:16:19] And do you think people are willing to buy into all of this or have you sensed that there's resistance out there because it's, you know, it's more involved than a Google prompt?

[00:16:29] Yeah, it is.

[00:16:30] It is very involved.

[00:16:32] Well, I'll go back and say, you know, those of us that knew how to actually build SQL type prompts in Google,

[00:16:39] because you could really do some effective.

[00:16:41] You could put quotes around things and ands and ors and buts.

[00:16:43] You could do a lot more than most people did with Google.

[00:16:46] But those of us who did that got better results.

[00:16:48] And I think the same thing is true with AI.

[00:16:50] I think the difference is it's changing so fast.

[00:16:54] And it's going to become, as I mentioned, those digital colleagues.

[00:16:57] These assistants are going to start showing up everywhere on your teams, everywhere.

[00:17:01] And I think they're going to challenge people because they're going to process things so fast.

[00:17:05] I think people are going to have a hard time.

[00:17:07] Like, you know, who are you to tell me that that's what I should be doing, even though our next people next year go, yeah, the AI is right.

[00:17:13] That's what we should be doing.

[00:17:14] And that's, I think, going to be the bigger challenge is if we don't prepare people to be ready for that and and have the emotional intelligence to be able to say, that's good.

[00:17:24] I can accept that.

[00:17:25] It's not a front to me.

[00:17:26] But you've got to make sure that our our humans are knowing how to use those things, because it's the ones who know how to use them effectively that are going to fly ahead of everyone else.

[00:17:36] Now, AI is changing so quickly.

[00:17:38] Yeah, it seems there's always a new feature coming out or a new player getting involved.

[00:17:44] Is this approach that you're following?

[00:17:48] Does it have a shelf life?

[00:17:50] Do you think that the technology is going to get to a point where people are going to have to revamp the whole way they approach prompts?

[00:17:57] Or do you think that this is probably going to be good for a while?

[00:18:01] I think so as these AI assistants become more personal, as they get to know you better, you know, we've trained.

[00:18:12] In other words, instead of right now, we've got large language models.

[00:18:14] Companies can get their own large language models and take their data and train it so it knows the company culture and how their policies work.

[00:18:21] You know, that we're getting to that point where companies do that.

[00:18:24] But when you start to think about it getting down to the individual, I think for quite a while, how you prompt and interact with it is still going to matter.

[00:18:33] And I think the more you learn about that now, when it evolves, you'll be able to evolve with it.

[00:18:37] Whereas if you're playing catch up the whole way and still don't understand it, it's going to be harder to catch up to it because it's going to be so far ahead of you.

[00:18:45] But I think that personalization where it really knows you as an individual is going to be where people are going to have to fully understand how these work and not get taken in by them, not get taken down by them or be so afraid.

[00:19:01] What I'm worried about is I think people are going to it's moving so fast and people are kind of already a little bit fearful.

[00:19:06] I don't want to see them start to sabotage these things, you know, because you can think about it.

[00:19:11] If the assistants in your Slack channel are always answering the questions and always get the right answers, people are going to start to feel like, hey, I don't want that thing in my Slack channel anymore because it's making me look bad or it's making us look like, well, you don't need us, which we all know isn't true for quite a while.

[00:19:27] But I think that's the emotional part we've got to really focus on today.

[00:19:30] And the better we do that today, that foundation of confidence as it evolves, I think then people will be the change management aspects because, you know, it all comes down to change management.

[00:19:40] They'll be able to change and evolve with it better versus today if they're not, if they're lagging so far behind and it's moving so fast, people are just going to throw their arms up and then start to take behaviors that we probably don't want them to do.

[00:19:51] So if you're a CHRO again and you're looking at this and learning about this, what would be the first steps you would take and how would you look to apply this to your work?

[00:20:09] So, you know, again, the first thing is figure out the goal.

[00:20:12] What is the goal of the organization or it could even just be the goal of the HR team?

[00:20:16] Whatever goal and whatever level of strategy it is, figure out the goal of what you want to get out of this learning session, say.

[00:20:23] So if it's, hey, we want to be able to train this group of people to be able to figure out how to effectively use these tools to make better decisions, quicker decisions, that stuff.

[00:20:33] That's a pretty good goal. It doesn't have to be high level.

[00:20:36] And then build a policy so everyone understands acceptable use.

[00:20:39] We most companies have acceptable use of technology already.

[00:20:43] You just have to kind of modify it to include AI.

[00:20:45] And then once you've got that, sit down and do a pilot project and train people right on the prompts.

[00:20:50] Like I said, it all starts with the prompts.

[00:20:51] I don't care what tool it is, whatever it is.

[00:20:53] Show them how to do these effective things with prompts.

[00:20:56] And you'll be amazed at how fast they'll take off because they start thinking of things.

[00:21:00] They do it in my seminars all the time.

[00:21:02] They start thinking of things that other people start to riff off of and go, oh, yeah, we can do this.

[00:21:07] We can do that.

[00:21:08] And then they put together a prompt.

[00:21:09] They do it.

[00:21:09] And they all look at it in amazement going, that's so cool.

[00:21:12] And before you know it, they're doing it.

[00:21:13] But unless we get them trained on the basics, it's not going to be effective.

[00:21:18] If they're all just sitting there typing a bunch of verbose prompts, they're going to get some ineffective results.

[00:21:22] And you're going to get results sometimes that aren't really what you want.

[00:21:25] But that's where then you challenge it.

[00:21:26] You say, hey, that's a level one result.

[00:21:28] Give me a level two now.

[00:21:30] And if you do something like that in ChatGPT, you'd be amazed at what comes back.

[00:21:33] It will challenge itself to give you better results.

[00:21:36] So I think the key thing really is define the goal, get your policy in place, update your policies to make sure people understand how to use these tools and how not to.

[00:21:45] Don't, you know, privacy.

[00:21:46] Be careful about putting data out there if it's not your own large language model.

[00:21:49] And then train them on prompting.

[00:21:52] Prompt that structuring.

[00:21:54] Words matter.

[00:21:55] The difference you use in words.

[00:21:56] All that stuff starts to help them understand how to take advantage of these tools.

[00:22:02] Paul, thanks very much.

[00:22:04] This was great.

[00:22:05] It's really fascinating.

[00:22:06] And I hope you'll come back and talk about it some more sometime.

[00:22:10] Absolutely.

[00:22:10] Thanks for having me, Mark.

[00:22:11] My guest today has been Paul Carney, the founder of Perpetta.

[00:22:27] And this has been PeopleTech, the podcast of WorkforceAI.news.

[00:22:32] We're a part of the Work Defined Podcast Network.

[00:22:35] Find them at www.wrkdefined.com.

[00:22:42] And to keep up with AI technology and HR, subscribe to WorkforceAI today.

[00:22:48] We're the most trusted source of news in the HR tech industry.

[00:22:52] Find us at www.workforceai.news.

[00:22:58] I'm Mark Pfeffer.

[00:23:03] Hey, this is William Tenka, Work Defined.

[00:23:06] Hey, listen, I'd like to talk to you a little bit about Inside the C-Suite, the podcast.

[00:23:10] It's a look into the journey of how one goes from high school, college, whatever, all the way to the C-Suite.

[00:23:17] All the ups and downs, failures, successes, all that stuff.

[00:23:20] Give it a listen.

[00:23:21] Subscribe wherever you get your podcasts.