0:00Sam Nadler says he built a front end for an applicant tracking system in record time, his first time using Claude Code. The episode also covers Envato's new AI tools and the week's AI news.
2:41Sam Nadler says he wrote a generic PRD in OpenAI, then a second PRD describing the features he liked, and ran up to three Claude Code agents in parallel on pages like jobs and candidates.
4:56Sam Nadler shows a Kanban-style pipeline board, a filterable candidates page, offer templates, custom feedback forms, analytics and light, dark and party modes, built in about an hour with fake data.
7:17Jordan Metzner says Claude Code could design the back end, database schema and Lambda functions so developers get implementation specs, not just a prototype. Sam Nadler says he can apply recruiter feedback within the call.
10:56Sam Nadler lists RyzScore resume scoring, AI-written custom emails and matching past candidates to future roles. Jordan Metzner says not to design the back end until the core feature set is built.
12:29Sam Nadler says Bolt is fast for small builds like a landing page. Jordan Metzner says Bolt or Lovable, Replit for its built-in back end, or Claude Code can each be best depending on the build.
13:58Jordan Metzner, a longtime Envato Elements subscriber, tests an Instagram ad prompt. He says it seems to run the prompt against several models, but none of the results were good and the text was misspelled.
16:33Jordan Metzner runs the same prompt through their in-house tool across Ideogram, Google, Flux and ByteDance; Sam Nadler notes the Google model got the text right. Jordan says Envato is still great for real stock photos.
18:48Sam Nadler says Zuckerberg is restructuring to move fast after new hires arrived. Jordan Metzner says he has heard of cultural issues when long-time staff know a new colleague may earn $250,000,000.
20:44Jordan Metzner says Grok's public share links were indexed by Google, exposing personal information, a week after the same happened with ChatGPT. He says Grok is not better than Claude for coding.
Build an Applicant Tracking System With Claude Code Agents: the key questions from this episode, answered with quotes and timestamps.
Jordan Metzner [0:00]: Oh, nice. Party mode.
Sam Nadler [0:02]: Yeah. So we'll we'll wait for the beat drop because I made this song custom as well. Sam's ATS. Yeah. Sam's ATS. You're probably familiar with what's called an applicant tracking system or ATS for short. Basically, it's a platform to know where your candidates are for every step of the process.
Jordan Metzner [0:29]: It seems like Envato themselves have gotten into the AI space. On Friday, the information announced that Meta's shaking up its AI org again. We haven't seen a new model from them in a while, you know. Obviously, Elon's been spending hardcore on getting Grok out to the users.
Sam Nadler [0:58]: Hey, everyone, and welcome to another episode of built this week, the podcast where we share what we're building, how we're building it, and what it means for the world of AI and startups. I'm Sam, a cofounder of Ryz Labs, and I'm joined with my friend, cohost, and business partner, Jordan Metzner. What's up, Jordan?
Jordan Metzner [1:17]: Hey, Sam. Super excited to be back. Another new episode, another crazy week in AI. So, yeah, lots to talk about this week, and lots of leveling up, I think. You know? AI has been a crazy tool to power and level everybody up, and I think this week's episode will show just that.
Sam Nadler [1:34]: Yeah. Just to before we get into the docket, just please remember, like and subscribe. Every week, we're covering a new product we've built over the course of a week. So whether you're a Spotify fan, Apple Podcast fan, or YouTube, I personally am a YouTube fan, just, sign up, like, and subscribe, and hear us every week. So the docket today, Jordan, is a tool I built. It's a front end tool for you know, hopefully, I can pass off to the our developing team and have ready within a short period of time. It's a pretty intense tool, actually, but what's amazing is I built it in record time, my first time using Claude Code. So we'll dive into that. We're also gonna cover a tool that we've used all the time here at Ryz Labs. You're gonna give us a quick kinda demo of what we've used it historically for and some new features they seem to have layered on. And then lastly, as always, top AI news for the week.
Jordan Metzner [2:36]: Okay. Cool. Yeah. Let's just jump right into it. Do you want to move to share your screen, I guess?
Sam Nadler [2:41]: I will. Before I share the screen, let me provide a quick bit of context. So at Ryz Labs, one of our main businesses is staffing, technical staffing. And if you've been a part of an organization that's you know, a large organization that's hiring or, you're probably fam or you've been a hiring manager, you're probably familiar with what's called an applicant tracking system or ATS for short. Basically, it's a platform to know where your candidates are for every step of the process. So I said, hey. Why don't I just try and build a ATS, applicant tracking system, for our exact use case? I would say there was, like, a few steps to my process. Number one was just a very generic PRD in OpenAI, threw it in Claude Code pretty quickly. There were some initial, challenges getting started, but within just a few back and forth with Claude Code, I got the initial version. Then I took it to I did a second PRD really explaining kind of our use case of the ATS, what features of our current ATS I really liked, and really made sure those were represented and highlighted in kind of my next PRD. Threw it into Claude Code, pretty much ready right away. And then what was really interesting about Claude Code, and I'll kind of click through the different sections of this ATS, was I was able to open up different agents and have them focus on different parts of this ATS at different times. So for instance, I would have an agent focusing specifically on the jobs page and really building it out to my specifications. While that was working, I had another agent focused on this candidate page.
Sam Nadler [4:32]: Third agent at one time, working on something else. So it's really quite amazing, how fast you can do this. I think, for me, personally, my limit was around three agents, making sure I was, like, engaging appropriately with each agent on what it was building. In terms of speed and able to work in parallel, it's pretty amazing. So let me quickly run you through this just to see what it built. I mean, you have the jobs page, which you can the different roles you have posted, candidates page. This is all fake data, so no personal information shared. And this is all pretty standard for a applicant tracking system, kinda have the pipeline board and can move candidates over or back.
Jordan Metzner [5:17]: Can you describe it for somebody who's listening maybe who, like, what these pages look like and how nice they are?
Sam Nadler [5:22]: Yeah. So they're beautifully designed. This pipeline board is you know, kinda looks like a Kanban board with the cards, the name of the candidate, the title of where they're currently at, has some tags. You can click on there. I don't think these buttons are working, so that's something I need to follow-up on. But you can move them over, into the different statuses of your process. You can edit these statuses. So maybe your process is slightly different than what's represented here. For the candidates page, you have all the candidates and what stage they're in. There's different filters, so you can say, all the source candidates or in about an hour, and I'm not arguing this doesn't need a bit more time, it's pretty robustly built to, like, pretty much exactly how we use an ATS today, and that was my goal with the second PRD. We even have an offers page. You have offer templates you could use, feedback. You can create your feedback form. So different candidates may require different types of feedback. If you have a very technical candidate, you'd want some sort of technical feedback, not maybe nontechnical, not necessary, so you can customize the feedback process.
Sam Nadler [6:36]: Different analytics, theoretically, this is to be built out, but that's, different tasks of your recruiting team, whether it's, scheduling an interview, etcetera. And then, one really fun feature that I think we like to do is I have a light mode, a dark mode, and then getting ready, a party mode.
Jordan Metzner [7:03]: Oh, nice. Party mode.
Sam Nadler [7:05]: Yeah. So we'll we'll wait for the beat drop because I made this song custom as well.
Jordan Metzner [7:13]: Sam's ATS.
Sam Nadler [7:17]: Yeah. Sam's ATS. Alright. So, that's it. I mean, at a and, like, what I would do with this now is I would continue to work on it, get it to a place where it's, like, pretty much exactly what I need, which I think I can do, honestly, within a couple hours. Just go through every page, make sure every button's working, make sure, it's exactly how I'd want it to be displayed, and then pass it off to our engineers. And, I don't think this is a super quick build for them to implement. It's a pretty big system. But probably within a couple months, we could have the first version, live. What do you think about that timeline? Would that be reasonable?
Jordan Metzner [7:58]: Well, I think I mean, maybe faster. Right? Like, some of the things well, I wanna I mean, maybe ask you first, but, like, before we even pass it off to engineers, how valuable is this to use as a prototype to work with the recruiting team to find, where are the paper cuts and how can we make this, like, the best version of itself. But I think, using things like Claude Code, we might be able to, design the back end, design the database structure, design what are the edge functions or Lambda functions that we need here. And so, maybe you only spent a few hours so far in building out the front end. I think if we can coalesce on, like, what are the finalized features, we probably could systematically go through and start building out the back end to the point that, like, your hand off to developers isn't here's a prototype, like, go build this for me. It's like, here's what I wanna build. Here's the front end. Here's the back end. Here's the database schema.
Jordan Metzner [8:54]: Here's the Lambda functions. Like, now you're just more talking about implementation. And maybe it's not perfect, but, it might probably save them, like, some more days and loops and whatnot.
Sam Nadler [9:03]: I think your point of showing this to recruiters who are in this tool their entire day, the iteration cycles are so fast that, just spending forty five minutes with a recruiter, the next day, another forty five minutes, I can have literally have implemented their feedback within the call. So, that's so fast. And then as you mentioned, there's a lot of we can build on the back end as well. I haven't really tried at all back end with Claude Code, but I'm assuming it's better as well.
Jordan Metzner [9:34]: Yeah. Well, I mean, first on the recruiter side, I mean, if you just, like, use something like Granola or something like that to record your call, I would just take their feedback on each page and then screenshot the page, throw it into like ChatGPT and ask it to give you like a list of feedback, for that page and then push it back to Claude Code, like trying to take advantage of best practices. I think another thing, that you can think about is, we talked a little bit about how does AI get implemented here and kind of what value does it add? Basically, rather than just replacing our current ATS, how can we use AI and a customized application to kind of take it to the next level? And then, back to your point of question on timelines, I mean, if this was a vanilla application, I mean, I hope we could build it in, a few days. If this is part of another application structure like our tracker app or something else like that, it might take, probably a few more weeks of development just because, like, that's a much larger structured app and have to build within the infrastructure of that application. But I hope it's not months away anymore, you know. I hope, things are able to go to market in days and weeks and not months and years anymore, you know. So but I think if we just really spend the time to refine each view and make it, like, the most powerful thing for the recruiting team and management and leadership, then it'd probably the value of replacing the current ATS might increase significantly. I don't know what are your thoughts there.
Sam Nadler [10:55]: The I would say, there's a few tools we've already built. One we've covered called RyzScore that immediately scores the resumes of the candidates. That that is, something that we'd wanna build in here. There's probably some AI communication features we could build in, writing custom emails. There's or there's probably this job matching feature. We screen a lot of candidates. And, just based on the process, A lot Of candidates don't end up with their next opportunity, but that doesn't mean they couldn't be great fits for a future opportunity. So, like, having quick access and pulling those candidates up using AI to do that could accelerate our funnel for certain roles. Like I said, the iteration cycles are super fast, and I think, doing a Granola call, having them give me all their feedback, then, throwing it into a new PRD and having Claude Code run with it is, would literally be done in minutes.
Jordan Metzner [11:54]: Yeah. It's awesome. Yeah. I mean, and then you can do it again and do it again.
Sam Nadler [11:58]: And it. Exactly. And then once I get it, 100% aligned with the recruiting team and have all those features, then move to building the back end.
Jordan Metzner [12:09]: Yeah. Yeah. I wouldn't design any back end until kind of, like, you have the, main core feature set built out. But, yeah, great job, Sam. I mean, like, just seeing your evolution on the podcast from, building out, like, things in Bolt to now, using multiple agents inside of Cursor. It's only been, like, eight weeks or something like that. So, yeah, it's incredible to see just your own personal revolution. Would you really start your next prototype in Bolt now that you know how to use Claude Code?
Sam Nadler [12:38]: It depends on the use case, but, Bolt's so fast. But I this was this is a pretty kinda big prototype. So, something small or just super quick or just like a landing page, I could understand it makes sense. Maybe try Bolt first. But, this was super easy too.
Jordan Metzner [12:58]: Yeah. I mean, like I said to you the other day, it kinda depends on, what the utility of the thing you're building is. And, sometimes, Bolt or Lovable is a good solution. Sometimes, Replit is a good solution as it has, like, built in back end. Sometimes just building, with Claude Code and building a front end, or, sometimes building a full web app with a back end, is the best solution. So it kinda just depends, I guess, on your implementation, but even you yourself are able to, like, size up these tools and kind of now you're deciding, like, which one to use for which. Right?
Sam Nadler [13:26]: Yeah. Absolutely. And, hopefully, I'm a push to anyone out there not using these tools. If I can do it, you can do it. And with that, let's move on to, I think we're gonna share our tool of the week, which, I was actually surprised. I mean, I guess it makes sense, but I was surprised that they had layered on some of these new AI features. But why don't I pass the mic to you and just what are we getting into this week?
Jordan Metzner [13:58]: I love Envato. I've been using Envato products and their different websites like ThemeForest for, I don't know, ten or fifteen years. We currently subscribe to what's called Envato Elements, which is a library tool that gives us a stock content of all different kinds, video, audio, graphics. And before AI, this was like our go to for things like music, for videos and things like that. But now with AI, we're we're generating a lot of things. So it seems like Envato themselves have gotten into the AI space, and this is their AI tool. They've got video and image generation and all these different pieces. So let's test it out a little bit here. We normally use Envato for marketing purposes, most like creating an Instagram ads. So, maybe we can do, like, create an Instagram ad for BTW built this week podcast to promote new episodes every Friday. What do you think? Should I change that prompt? Let's try to enhance it.
Sam Nadler [15:06]: I think it's great. Yeah. Enhance prompt.
Jordan Metzner [15:08]: Okay. So I don't know what the enhanced prompt does, but I presume they're using some AI here. Okay. So podcast advertisement, blah, two hosts, cohosts. Okay. Cool. Sounds good. Alright. Let's generate that. Let's see what comes out of this thing. And I haven't tried this yet, so we will see what comes. It looks like it's making a lot of images. Oh, that's pretty good. It's pretty cool.
Sam Nadler [15:31]: Yeah.
Jordan Metzner [15:32]: That's us right here.
Sam Nadler [15:35]: Yeah. Not so bad.
Jordan Metzner [15:36]: I mean, yes and no. Is it? Because like all the text is like, back to this Jumbo. Yeah. Yeah. And I mean, like, maybe this works. That's not really what we look like, but, you know. Yeah. So maybe I say like, no text. No text on screen. Because it seems like they can't do text very well. So, maybe if you were to ask me just looking at this, I would say that what they're doing is they're taking my prompt and they're spitting it against, like, it looks like eight models, but it could be just like four or five models with some variants. And then they're just getting back the response. So, like, this could be Google and this could be one and this could be this and this could be that. And so they're just, they're basically, testing like six different models, running them up, and then bring it back to you. But, if you look at this, none of them are good.
Sam Nadler [16:29]: None of them. Yeah. It's not as great as I thought it'd be.
Jordan Metzner [16:32]: Yeah. Where whereas, like, we could probably use our own, like, image generation tool that we showed on a previous podcast. Okay. I mean, like, we could just try that as an alternative here. So, all of these have text. All the text is spelled wrong. So, that doesn't really work very well. If we try our in house build tool, and let me just try to generate an image with, like, let's just say, like, Google. So I'll put this in here, the same prompt. I'll let but I just put the same one in into the Google model, and let's see kind of what comes out of it. And then we can go back to the Envato in just a second. But yeah. I mean, you can see that this is just a very rudimentary elementary space. See, still has a lot of text on it. And if you can just see, this is the Google model. Right?
Sam Nadler [17:19]: The it looks like the text worked in that model.
Jordan Metzner [17:21]: Yeah. Friday episode release. Listen now. I mean, all the text is good. We're we're not these two people. So that might be a problem. But overall, and we could try this like, if we go back to our image generator, we could try it, in multiple models. Okay. So the idea of this like wide model generator and if it works right is that you basically take one prompt. Yeah. It's looks like it's gonna work. And then what we do is we actually run it against, like, four different prompts simultaneously. And then with all of that, we should be able to get four different responses. So this is, like, Ideogram. This is Google. This is Flux. This is ByteDance. I didn't wanna spend like too much time on our image generation tool. You see, like, we're kinda
Sam Nadler [18:02]: There we go.
Jordan Metzner [18:02]: Yeah. Then Text is much better here. Built is better than this one, but anyway, you get the idea.
Sam Nadler [18:08]: So we'll we'll just use Envato for the kind of stock footage we've historically used it for, which continues to be less and less every day, to be honest.
Jordan Metzner [18:16]: If you need stock images or if you need, stock content, it's still, for stock photos, for example. Yeah. Here's a stock photography of office space. Like, it's still great for that stuff, and it's obviously all real. So, you really get a you get feedback if you need to throw that in there. But anyway, so yeah. Some you know, it's hard. AI is hard. We're on the cutting edge, and, I don't think we've seen, like, a silver bullet here that does one thing really, really well for everybody all the time.
Sam Nadler [18:43]: And in terms of news, what are we covering today?
Jordan Metzner [18:48]: On Friday, The Information announced that Meta's shaking up its AI org again, following up on kind of the AI story here. But I think this is funny because of the fact that they just started to build this thing, and they just spent all this money on Scale AI. But curious what your thoughts are. I've talked to a few friends inside of Facebook, so maybe I can give a little inside baseball there. But love to hear your thoughts as, they're trying to figure it out, and we haven't seen a new model from them in a while. You know?
Sam Nadler [19:13]: Yeah. I think it's just Zuck is moving quickly, and he prob there was probably some org changes that better aligned with kind of once everyone got in the door, what direction they were gonna take and to make the org structure reflect, those new decisions and to move fast. That's kinda how I'm interpreting these changes. But, well, what'd you hear on the inside?
Jordan Metzner [19:37]: I just think there's some cultural issues I've heard. I mean, I've read it as well, like cultural issues against, some of these, new engineers making so much money, and then there's all these other folks who've been at, Facebook and Meta for a long time. Not to say they don't get paid well. I'm sure they get paid really well. But, when you know the guy next to you is making $250,000,000 a year or even if that's over four years, like, you look at your comp and you say, like, well, maybe I'm not earning enough. So I think, on a basketball court or on a soccer field, I think most of the players recognize pretty easily Lionel's better you know, Messi's better than them or LeBron's better than them. But, in the product development world and the building of apps and software engineering, there are some things that you can be done to quantify, who's better than somebody else. But I think it becomes harder and the line gets blurred and people feel like, maybe there's like the haves and have nots or, maybe there's kind of some fiefdom or kingdoms being built. You know?
Sam Nadler [20:37]: Got it. But, yeah, that makes sense. Anything else top of mind in the news today?
Jordan Metzner [20:43]: Yeah. So we can talk a little bit more just on AI as well here. Obviously, Elon's been spending hardcore on getting Grok out to the users. And I don't know if you saw this, but basically the same thing that happened with ChatGPT a few days ago happened with Grok now, but essentially their public share links of the Grok responses got indexed by Google leading to some personal information getting shared. But one, this is funny because it just happened a week ago with ChatGPT. But two Yeah. They don't let nothing gets by them. I mean, literally, took this public this link, they made it public, and immediately Google's indexing it. It just shows, like, how fast and how good their indexing technology is as well. Have you tried Grok? Have you tried Grok 4?
Sam Nadler [21:25]: A little bit, but not nearly as much as the other models. I know you Yeah.
Jordan Metzner [21:34]: It's okay. It's not better than anything else. I mean, it's not better than Claude for coding, or, GPT five, I think, now. At least not in my opinion. But I haven't tried the heavy at length.
Sam Nadler [21:44]: Well, great episode, Jordan. Thanks for walking me through the latest news this week. Thanks for listening. Like and subscribe to Built This Week. We're here every week trying to highlight what we build.
Jordan Metzner [21:55]: Yeah. Just wanna say like and subscribe. Hit that bell button so you can get notifications every time a new episode is out. Check us out on YouTube, Apple, Spotify, or your favorite podcast platform. This is Jordan and Sam for another episode of Built This Week. Thanks, everyone.
