ChatGPT agent mode: The “little helper” that transformed recruiting, crafted user personas, and solved parking nightmares | Michal Peled (Honeybook)

0:00Michal Peled:We're going to start with something that we haven't actually seen on How I AI yet, which is agent mode in ChatGPT. My use case was with our hiring team. Part of their workflow is to browse through many LinkedIn profiles and search for relevant candidates. It takes a lot of time. Let's talk about the prompt. I'd love for you to go through how you thought about structuring it to make it effective with the agent. I want a little helper. I'm a recruiter. I want someone who is like me. So I started by telling you, you're an ID recruiter. And then I described what I wanted to do. I love that you called it your little helper because don't we all want an AI little helper?

0:42Michal Peled:Welcome back to How I AI. I'm Claire Vem, product leader and AI obsessive here on a mission to help you build better with these new tools. Today, I have Michal Pled from HoneyBook. They're a technical operations engineer who's building tons of internal tools and automations to make their team's life easier and reduce friction. Mihal is going to show us some advanced features of ChatGPT, including agent mode, replicate not one, but five of their personas as AI identities, and save me a lot of time on my commute using ChatGPT. I'm really excited about this episode. Let's get to it. This episode is brought to you by Brex.

1:22Michal Peled:If you're listening to this show, you already know AI is changing how we work in real, practical ways. Brex is bringing that same power to finance. Brex is the intelligent finance platform built for founders. With autonomous agents running in the background, your finance stack basically runs itself. Cards are issues, expenses are filed, and fraud is stopped in real time without you having to think about it. Add Brex's banking solution with a high-yield treasury account, and you've got a system that helps you spend smarter, move faster, and scale with confidence. One in three startups in the U.S.

2:03Michal Peled:already runs on Brex. You can too at brex.com slash howiai. Michal, thank you so much for joining How I AI. I'm excited to see what you have to share. Thank you so much for having me. We're going to start with something that we haven't actually seen on How I AI yet, which is agent mode in chat GBT. And so I'm wondering if you can just go ahead and dive into what was the problem that you were trying to solve? And why was this agent mode, this agentic browsing, the solution to the problem you're having? Our problem was, you know, same as our customers are having. You have to do your job. You have a job that you really love doing and you have your proficiencies and expertise.

2:55However, you spend a lot of your time doing the mundane, thoughtless, manual repeating work in order to get the information that you need. So my use case was with our hiring team. And as a recruiter, when you get a job description that you need to recruit to find candidates for, part of their workflow is to browse through many LinkedIn profiles and search for relevant candidates that may be relevant for the job descriptions. and it takes a lot of time. It can be hours of browsing through profiles and going through all of the characteristics that they're looking for. So I wanted to take that load off of them and ChatGPT agent mode came just in time.

3:54We all talk about what agent is and what agents do and how we can use them. In ChatGPT, it's very simple to understand. So you just open a chat with ChatGPT, but then you add an instruction and turn it into an agent mode very simply from the toolbar. And once it goes into agent mode, it means that it can take the prompt or you can actually use specific prompts to tell it not just to search for information online, but also to perform actions for you. And why did I need it in this case? Because I needed to log into LinkedIn. I don't want it to just search for profiles on LinkedIn, just profiles that are publicly accessible.

4:48That's not the information that I need. So I needed it to log in into LinkedIn and I needed it to perform search and I needed it to go through the profiles and look for the restrictions that I want to give it. And those restrictions were provided by the actual hiring team that they actually use it as requirements for potential candidates that they find.

5:15Michal Peled:Yeah, let's talk about the prompt really quickly, because I think this prompt is interesting as I'm reading it. And I'd love for you to go through how you thought about structuring it to make it effective with the agent. Of course. So I usually start my prompts, begin my prompts with telling the GPT its role. And so here I told it, you are an IT recruiter. I want a little helper, right? I'm a recruiter. I want someone who is like me that will assist me with my job. So I started by telling it you're an ID recruiter. And then I described what I wanted to do, what the task is. Log in to LinkedIn using my account.

5:57If not already logged in, let me take control and log in. It is something that is possible. Find up to five LinkedIn profiles where the current title and job description match the attached job description. and here is the part where I just uploaded a job description. In this case, it's for an engineering role, okay? So I have the job description and I have the, you are an IT recruiter, this is your job. This is the task and I provide like a full description of the task. Actually, actually describing what an actual IT recruiter would do. And then I added restrictions or special instructions. It doesn't matter how you call them.

6:49But these are important because these don't just search for something that matches the description. Do it the way that we do it. And when our hiring team goes into a search, they have specific criteria that they go for. So I collected these and I added it as a list, as a restriction. I could call it instructions. It would have been the same. So candidates must be from Israel because the job is being filled up in Israel or currently working at an Israeli company. And they must be active in LinkedIn within the last three months, because that's something that our hiring team is looking for. And the current job role must be close enough to the open role in title and seniority.

7:43And also something that is special, the candidates must either work in their current workspace more than a year or they can be unemployed, but no more than a year and have worked in their last workplace for over a year. These are all things that I didn't invent them. They were taken specifically from our hiring process.

8:11Michal Peled:What I love about this is exactly what you said, which is, first, I love that you called it your little helper because don't we all want an AI little helper? That is my goal. Maybe I'll reran my product to just little helper. Um, but what I like about this is, you know, it, when you're building a tool like this or a prompt like this, the simplest way to get to a good outcome is simply interview somebody and say, step by step, just tell me what you do. Tell me what you do. And if you can codify what a person's step-by-step workflow is, and you can put that into a pretty simple prompt, which here it's only a paragraph and three or four bullet points, you can replicate and automate that at scale.

8:56Michal Peled:and typically this is not the highest order thinking you want your recruiter or sourcer to do you don't want them just to build a list and be looking is this person here a year or not that is an input to what you hope is a great recruiting process great outreach all that kind of stuff so one I think it's just really great to interview your colleagues and say how do you do your job and what parts do you hate and let me automate them the second piece that I think is really interesting here is you're actually pretty specific about a couple outcomes. You're specific about the number of candidates that you want.

9:31Michal Peled:So I think that's really helpful. You're specific about a kind of threshold of matching your criteria. So you say 70%. And I often find these LLMs are very, you know, what is 70 % matching? Like it's not, it's not pure science, but they're pretty good at following a general threshold. So I think that is really interesting. And then the last piece I want to call out, which we'll maybe see in the demo, is agent, while it can be agentic and independent, can also be a co-pilot and collaborator. And so you actually instruct the agent when you're going to take over and when they're going to take over.

10:11Michal Peled:And so I think that is also really interesting things for folks to know, is you don't just have to like press the button and walk away and let the agent run. You can press the button and say, hey, wait, when you get to this point let me take the next step and then you can go on from there so that's really interesting exactly or if you encounter a problem with this and that stop and ask for my assistant um and that's exactly the agent that's exactly the agent mode thinking about it as a little helper will really help you come up with good prompts for it okay i think you gave us our show title will be Chad, GBT agent mode, your little helper.

10:54Michal Peled:So let's see how it runs. Exactly. So once you start running it, and this is something that is mind-blowing for anyone who tries it for the first time, even the ones who are very proficient with using AI tools, suddenly you get a computer opening up. in your, you know, you see your little helper actually doing things in the computer. So it started by reading my job description and then you can see it goes, it will try to go to LinkedIn. It will probably be already logged in because I logged in beforehand. And the thing I like the most you see it's logged in and then you can see like the arrow goes and clicks on things and searches on things and go through the list and the thing I like the most is is that during all of this time you can see the thoughts of the agent now I will go to the feed page loaded again I plan to click on.

12:04First, I need to make sure like you can see inside the brain of your agent while it is thinking. And so and all of this is live. So I will let it run here, but we can go and see results.

12:22Michal Peled:Yeah, you know, I want to call out a couple things because I know that often on How I AI, we have highly technical use cases for highly technical folks. But we also have a lot of people that are actually quite new to using generative AI tools and have probably been pretty familiar with ChatGBT and the direct experience. But I know if I showed my mom this or even some of my friends that maybe don't work in tech and said, hey, did you know that ChatGBT can open up a magic computer and navigate it and narrate its thoughts and look for things for you? They'd be pretty, pretty surprised. And I think, you know, as you watch this, what I hope our listeners and viewers are taking away is you don't just have to rely on text prompts and chats when you're using these tools.

13:12Michal Peled:Now that the next kind of like evolution of these LLMs, especially the more like consumer focused general purpose ones like ChatGPT have evolved, you actually have a lot more tools. And so I just want to call out some of those folks out there. I'm thinking a little bit of my parents and some people in an older generation who are like, how do I get from here to here? I need help searching for flights or I need to do a certain kind of research on a niche site. Having sort of this expert computer operator on hand, I actually think is going to make information more accessible to folks, but it's also going to make UX and websites more accessible to folks that don't have the time to figure out how do I use the best filters on LinkedIn or those sorts of things.

14:02Michal Peled:And so I just, I want to make sure that people that have not experienced agent mode, and I know we're all on the edge. So maybe all of us have just, you know, take a minute to appreciate the kind of use cases that this opens up. Also, it's just fascinating to watch. Yeah. AI operator. Of course, first time I tried it, I just sat and watched the thoughts of the agent while it was thinking. Going to our more technical audience, though, a couple of things that I want to call out is one, props to the OpenAI ChatGPT team. What a great user experience design here. Like, it could feel very strange to watch an agent browse a website.

14:45Michal Peled:It could either be boring or weird. And I think this user experience of like being able to see where the cursor is, showing the reasoning and thoughts, watching it navigate is actually pretty entertaining. That's a hard thing to pull off for a consumer product. And so for anybody designing AI products, it's worth thinking about some of these interaction patterns here. And then again, I just think about how long this would take someone to do. we're watching it because I'm trying to narrate some of the features but you know you could you could walk away you could go to a meeting while this happens or you could do something something else if you have chat gpt on your phone you will get notification and a notification when it's done to tell you that it's done and and here are the results for you we said little helper And the last thing I'll call out, I love, I really do love this use case, is I do this specific thing all the time.

15:44As a hiring manager, executive leader in an organization, I was constantly looking for like, who's a senior director of DevOps and Platform Infra,

15:56Michal Peled:who is either in San Francisco or works for a San Francisco-based company, who has experience in dev tools, that one change I would maybe make on your prompt is like, is one or two connections away so I can actually message them or get a back channel reference on that? I did this all the time. I had my hiring managers do this all the time. So even if you're not in recruiting and you're just somebody who does hiring, I think this specific workflow is really, really useful. Yeah, definitely. That's an excellent use case that you just mentioned. Well, let's actually, let's look at the output. We'll let this run, but I know you have an example output for us.

16:36Yeah, yeah, it runs. I have an output that actually worked for 10 minutes, just that. And within these 10 minutes, I got a list of five candidates as I requested for. and you can see the match score. Well, having a match score or rank for results is something that I really love doing. It's not a must, but if you give a specific, if you provide specific requirements and ask for a match score, it is easier to understand what results have more quality for you. I mean, otherwise you could just get a table of like, these are the five results, but is someone of them better or maybe a better match for what I need than the others?

17:30I won't be able to see it unless I instruct the GPT to provide me with some score. So it's not an exact science, but it does give you some kind, a way to compare between the results. So I will say someone who got a 90 % match is probably as like probably will be a better match than the 78 % match. And I will have to go deeper and understand why. Yeah.

18:04Michal Peled:I'll call out a couple of things here that I think are interesting for people to look like. I know we were talking before the show, you actually made this anonymized data just so we weren't showing people's profiles or showing how person A versus person B fit a specific job you're hiring for. But I will say anonymizing candidate profiles is actually a pretty standard practice in a lot of recruiting flows just to make sure you're not biased, this school, that school, this person, that person, this name, that name. And so I actually kind of like this flow where you're really just comparing the qualifications against your stated objections or objectives.

18:43Michal Peled:And so I think that is a really interesting kind of meta flow that you're showing here. The second thing that I wanted to say as I was reflecting back on agent mode is it's almost exactly like a recruiter or sourcer would navigate LinkedIn, except for one thing. when I log into LinkedIn, I don't go straight to the job to be done. I don't go straight to that search bar and search for like VPs of engineering. No, no. I get distracted by the notification. I start reading the feed. I'm responding to comments. I go through my inbox. And so I think like, why is it 10 minutes? It's 10 minutes because it's like pretty hyper, hyper focused and efficient, but it's also 10 minutes because you're not getting so distracted with all the other things in the application and you can really just get the agent to focus on the task at hand.

19:35Michal Peled:So maybe it's a way for us to all break. Sorry, LinkedIn growth PMs. I apologize. But it's a good way for us to still get the value of these platforms without getting our time sucked into the less value generating aspects of them, maybe. Yeah, that is correct. So tell me a little bit about how this was received by the team. I'd love to know the kind of outcomes here? Yeah, well, I will be real and say that like the first result that I got, I was very skeptic about. So I just took the table and I sent it to our hiring manager. And I told her, this is the job description. This is what AI found for me in LinkedIn.

20:23If you can go through the results and let me give me feedback. Are there good results? Are we familiar with these candidates? Did we try to reach out to them? Or you're looking at them and say, oh no, that's a terrible fit. I don't know why this person is even in this table. And so she went over them. You can see that the table has linked to the direct, like the LinkedIn profile per candidate. So she scanned those five profile and she came back to me and she said, well, you know what? Out of these five, four of them were never found by us manually and they really fit the description. So we would want to approach and, you know, try to get them to come for an interview.

21:17And the fifth one was actually one that we caught manually and is already coming for an interview. And so to me, it was a great sign for quality. I mean, it's not just a list of names. Those are actual, real quality candidates that we can work with. And so now they want the agent to run on a lot more job description, many more job descriptions that we have, provide them with more than five candidates i wanted just five to see if if it if it's worth something um yeah but now it's going to be a real part of of their hiring process freeing their their time to do other things that they love and appreciate a lot more yeah and i

22:10Michal Peled:can't emphasize what you're saying enough because so many people push back on ai saying yeah you may get speed and you may get efficiency, but you're not getting quality. And my experience has has been the opposite of that. You get speed and you get quality. And again, it's those it's that last mile, those edge cases, those ones that are like just a little hard to find, a little hard to research, where I think AI can increase the quality of that last bit. And so it was it's amazing to see that this work for your recruiting has given me so many ideas, not, you know, not just no recruiting use cases, but just in general, one people finding use cases.

22:49Michal Peled:I was thinking about how you could find great candidates or customers on like X or LinkedIn. And then the other thing that I think is really great here is just showing we don't get a lot of GNA functions. We don't get a lot of people functions getting love in how I AI. I feel like all the noise is about like product design, engineering support. And so I just love seeing the recruiters get some love here because you're the people that bring in great talent and fun colleagues to work with. So thank you for showing this. This podcast is supported by Google. Hey, everyone. Shrestha here from Google DeepMind.

23:26The Gemini 2.5 family of models is now generally available. 2.5 Pro, our most advanced model, is great for reasoning over complex tasks. 2.5 Flash finds the sweet spot between performance and price. And 2.5 Flashlight is ideal for low latency, high volume tasks. Start building in Google AI Studio at ai.dev.

23:50Michal Peled:Let's zip to your second use case, which I think is really, we're going from finding real people to creating fake people. So I'm excited about this next workflow. That's an excellent description of it. So let me ask you that. Imagine you're a business owner and imagine being able to talk to thousands of your potential customers all at once and gather their insights on your planned ad campaigns, planned features, product experience, all from your phone or tablet, 24-hour day with one click of a button. actually talk to them. I thought it's mind-blowing. And so it started with HoneyBook invested in comprehensive customer research with a third-party provider.

24:45We interviewed hundreds of our target small business owners, and they created five detailed buyer personas. But the research was trapped in documents, hundreds of pages of insights that teams rarely referenced because it was too time consuming to extract actionable information when making product or marketing decisions. So the end goal was we have five personas, we want to talk to them, let's create a chat GPT that is that person, that actual person. And so I started with And here that there are some technical takeaways, but here I want to put the spotlight on the thinking process. Because it's very easy to go to ChatGPT and everyone with a subscription tier of, it's not the plus, a plus subscription tier and above can create their own custom GPT.

25:50and so you go to create a custom JetGPT. It's quite simple process. You add a name to your GPT, a description, but the most important parts are the instructions you're providing it with and the files that you can upload as a knowledge for that chat that you're talking to. So I needed five like them. But first I thought, okay, this is all I know about custom chat GPTs. I can basically take all of the documentation from the interview and just upload all of the files, text files, presentations, whatever, into this custom chat bot, provided with some instructions on how to answer what the research was about, what to say, what not to say, and ask it questions about the research itself.

26:43But that's not what I wanted. So I was like, okay, if I'm taking just the files per persona. So I'll concentrate in one persona, take the files related to that persona, upload it to the chat, inspect the chat what that persona is, how to read the files, what's included in the files. Then I'll be able to maybe ask about that persona. And I will probably get answers like, that persona would have done this, or that persona would prefer that, but it's not like talking to the persona. I'm still talking about the persona and not with it. And so I realized I'm not going to rely on uploaded files for the chat GPT.

27:34I need the instructions to be the main and most important part of what consists of that persona. And the instructions here will not be what's in a file or talking about someone else. There will be exact instructions that go like, you are that person. And this is your belief system. And this is how you run your business. And this is how you deal with media. And this is how you deal with technology. And everything has to be super, super, super tight. So that chat GPT once running live will actually become that persona that can answer based on what it knows about itself, not files that are attached to it.

28:26And so I needed to bridge that gap. I have all of the research there. I need to make a person out of it or five of them, actually.

28:38Michal Peled:okay so okay but you heard it here first this is our first how i ai where we are manufacturing not one not two not three not four but five people so show us the process yeah so i thought about it and then i decided to go to another tool that i really like this is notebook lm uh it's a google it's a google's tool and notebook lm the thing i like about it the most and the reason i picked that tool in order to construct the instructions or the prompt per persona is actually there are several reasons. But one of them is notebook.lm allows you to upload your own sources and can answer only based on these sources and not things that it goes and finds online or things or knows or filling in the gap.