0:00Sam Nadler says the episode demos internal tools built this week and tests Claude Cowork in real time. Jordan Metzner says more of their tools are now built by Sam, going deeper into the code base.
1:37Sam Nadler says the tool takes a transcript from an AI note taker like Granola, acts as a loose filter on whether to move a candidate forward, and coaches recruiters against a rubric of about six areas.
3:51Sam Nadler says you paste a transcript and optional job description, the recruiter records their own call first, then the AI gives candidate scores, screen feedback such as weak follow-up probing, and suggested questions.
6:03Sam Nadler says every screen becomes searchable, for example candidates who mentioned MercadoLibre or Python, and a manager dashboard gives each recruiter coaching feedback over the past thirty days.
7:35Sam Nadler says he built it in Claude Code in the terminal with Supabase, deploying via Amplify, AWS and GitHub, in about six hours without writing a line of code. Opus 4.5 powers every analysis.
10:45Sam Nadler says Cowork is a wrapper around Claude Code in Anthropic's Mac desktop app. You give it a folder of docs such as a PRD, and it proposes trade-offs in a UI, here for a Hollywood personal trainer app.
13:31Jordan Metzner says the interface does not really matter, but he prefers the terminal because he can run many sessions at once. Sam Nadler says running several terminals is like managing a team of little Claudes.
16:07Sam Nadler says Anthropic closed a funding round above $10 billion. Jordan Metzner says they are shifting spend from Cursor and ChatGPT to Anthropic because its models are top notch and it focuses on enterprise, not ads.
18:28Jordan Metzner says everyone is becoming an agent manager. Giving staff Claude Cowork or Claude Code has raised efficiency, such as faster invoices for finance and proposals for sales, without any reduction in workforce so far.
Build an App Without Code: Claude Code and Claude Cowork: the key questions from this episode, answered with quotes and timestamps.
Jordan Metzner [0:00]: At the end of the day, that's all that really matters. I mean, if, like, tomorrow Google's new, Gemini tools become top notch, we probably would move over to there.
[0:10] Theme song: Built this week, breaking it down. Built this week, we show you how. A fresh idea, a clever tweak you locked in true. Built this week.
Sam Nadler [0:20]: Hey, everyone, and welcome to 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 Nadler, cofounder here at Ryz Labs. And each week, I'm joined by my friend, business partner, cohost, Jordan Metzner. What's up, Jordan? How are doing today?
Jordan Metzner [0:40]: Yo, Sam. Happy to be back. It feels like we've aged a bunch. A lot of new AI products, it's been a crazy few weeks, least for me. But happy to be back, and fun new episode.
Sam Nadler [0:49]: Yeah. Fun new episode 29. If you haven't already, please hit that like button. Subscribe. We have new episodes out every week where we showcase what we're building, how we're building it, and how we're using AI. And this week, we're gonna demo some internal tools we built. We're also gonna take a quick look at Claude Cowork. We're gonna try and use it in real time, see if it works. And then we're gonna cover some AI news. That sound like a decent plan?
Jordan Metzner [1:17]: Yeah. I mean, I think as you've seen the progression of this podcast, more and more tools are being built by you, and more tools are being built using less, like, build for me tools and kinda getting deeper and deeper into the code base. But maybe at the same time while also getting further away from it. I don't know. You tell me. So why don't we jump into what you built this week, and then we can talk about it a little bit.
Sam Nadler [1:37]: Yeah. So let me just provide a minute of context. There was a couple reasons where I why I endeavored to build this product. Number one, and it's a tool that we use here at Ryz Labs in order to analyze the recruiting screens our recruiting team performs on a daily basis. Every day, they're talking to dozens and dozens of candidates across different stacks, and we have a pretty large recruiting team. And what this tool does, and I'll demo for it, we all now use different, like, note taking AI tools. It helps summarize notes. It helps, provide key takeaways from the conversation. But most of the time, it stays kinda with the user. Well, we built a tool that will take the transcript of one of those kind of AI powered granolas or what whatever you're using and will analyze a recruiter screen, and it will then, judge the performance of the screen of not only just the candidate. Is this a candidate who we should move forward? It's actually a pretty kind of loose filter. We don't wanna prevent candidates from moving forward if they could be a good fit. It's the very first step in the process, but it analyzes if this is a candidate that we should move forward or not. But what it really does that's going to be really helpful, it helps provide coaching to the recruiters.
Sam Nadler [3:06]: You know? Did they miss an opportunity to drill deeper into a candidate's, previous experience? Did they cover next steps? Did they engage in a conversation about compensation? And there's we have a rubric of about six, different areas that we hope our recruiters cover in a very fast phone call. It's twenty five minutes. So it's a lot to cover, and it's hard to check off all the boxes. But, hopefully, with this tool, we can really help coach our different recruiters in the areas. We try a lot to make the screen experience consistent, but this now will give us some tools to be able to do that. So let me walk you quickly through this tool. Hoping you can see it. It's got a great name called ScreenEval. This is the home page, new evaluation. So you literally take a transcript. I've got one right here for a back end engineer. Drop it in here. You can paste the job description, which will analyze the transcript against the job description, but it's not even necessary. Once you paste it in, it fills out these fields for you, who the candidate name, the role, the recruiter, the interview date, and then the theoretically, the recruiter will then provide their assessment. So I want them to provide their assessment before getting the take of the AI, and in fact, I never want them to change their assessment. We're really using it you know, there are little things that the recruiter can pick up on, potential red flags that may not be, captured by the AI.
Sam Nadler [4:41]: However, it will hopefully give them, like, at least a little bit reinforcement if they made a bad decision, or the summary will give them more context on how they feel. So either way, it takes two seconds. You submit to see AI evaluation and bada bing. So this candidate, your call was to move forward. AI also says To move forward. It gave a little bit of information about the candidate, their scores, communication experience, enthusiasm. And what I think is really helpful, it also gives me a screen performance. So solid screen overall. Rapport building was pretty good. Probably lacked a little bit in follow-up depth probing questions. We did discuss compensation. We did discuss timeline and next steps. Pacing was good. It gives me a potential other questions to ask. Walk me through your last technical decision. Yada. So long story short, it hopefully, you run this through, you feel confident about what your AI evaluation says. Even if it's not the same thing, we want the recruiters to move forward with their decision. However, it gives me a little coaching on maybe some of these areas I should focus on a little bit in my next screen. And then that what's really interesting is now that we have all the data of these screens, we can now search these transcripts.
Sam Nadler [6:02]: So all our recruiters have access to this search functionality. And let's say, we're specifically looking for candidates from MercadoLibre. MercadoLibre is the largest ecommerce here in Latin America. It is a place where we like to source candidates. They're usually a strong talent pool. But sometimes maybe it could be hard. Our ATS isn't isn't that great. Anyway, long story short is you can search here, and you can pull up all the candidates who mentioned MercadoLibre. You have some other filters here, or maybe you're specifically looking for Python candidates, and you're just pulling them up here. It's a really, really great feature that we can now search not only based on their, resume, their, inbound application, but also their transcripts. And we also do have some managerial analytics here, and we got, some this is the dashboard for the analytics. What's really interesting is this coaching feature. So I'm the only one only recruiter on this demo version, but if you click on my name, you'll get Sam's performance. So, this will look back over a period of thirty days and provides our management team some specific feedback to give to me when I'm conducting screens. I you know, it seems like I really lack in probing those follow-up questions when talking about their experience, And I could sell the opportunity of working with Ryz Labs a little better according to this evaluation.
Sam Nadler [7:34]: So, generally speaking, this is not live yet. It's, a day or so from going live and getting tested by our recruiters. There will probably be additional fixes and features after that. But this was built in Claude Code, Supabase, and that's pretty much it.
Jordan Metzner [7:51]: Oh, man. It's amazing. Honestly, it really is amazing. Obviously, the thoughts of just automating putting the notes in here so that this all gets automated as well. Jumping back into it's incredible. And if you could just go back to where it gave you the assessment, do you have, like, yeah, that one? Yeah. So, I mean, just this key evidence on, like, why the candidate seems good, seems amazing. I mean, this is an incredible demo, Strand. Cool. Yeah. And then, obviously, here you're seeing other candidates who are similar to that candidate. I mean, maybe there's a little bug there, but, for the most part, that looks pretty accurate. And, yeah, I think, as I'm thinking of this out loud, probably the next step is just to provide this feedback in real time while the interview is taking place. So let's, maybe replace Granola with our own custom built solution and, essentially provide real time feedback to recruiters to help them provide an even better interview experience to candidates.
Jordan Metzner [8:40]: So this is really, really awesome. Tell us a little bit, like, what code what tools you used to build it, not, like, kind of how what it's posted on.
Sam Nadler [8:48]: Yeah. I mean, it was done via Claude Code in terminal. I think I started, which we will demo here in a second, Cowork. And I really do like Cowork, but I think, like, pretty quickly, it just makes sense to keep it in the terminal. And then, I had Supabase. Supabase, I find, is is very easy to understand. That was literally it. I mean, then once I got to the point of, like, about to deploy it, which is where I'm at now, you know, our stack is Amplify, AWS, and GitHub and but, really, once it's all connected, you it's pretty simple to get it live on a Ryz Labs subdomain. So, it does take I would say, I started this Monday afternoon. Today's Wednesday. And probably a total, like, working between meetings, somewhere in between six hours to complete. So, like, really pretty reasonable. Unbelievable. And we're not done yet. Right? Like, there's gonna be feedback from the recruiters. They're gonna catch bugs. I've pushed it, like, a similar product on Monday to a different team, and, I've had three or four bug fixes from, live testing in, a live environment. But it's pretty incredible we could build this or I could build this with who I haven't written one line of code in my life, full stack web app development in six hours that has some pretty cool features.
Jordan Metzner [10:11]: That's honestly, it's incredible. And so, yeah, I guess, like, we get the feedback from the interview that we didn't get previously. Recruiters get feedback on how they think the candidate did versus how well the system believes the candidate did. And what models are you using to power that, like, LLM, like, AI analysis?
Sam Nadler [10:28]: Opus 4.5 for everything.
Jordan Metzner [10:31]: Okay. So even that LLM, like, the score is coming from Opus? Yes. Wow. Okay. Full on Anthropic Stack. Alright. Sounds good. Full on. Yeah. I mean, I guess, like, how could somebody build something like this themselves? Like, how long you know, how easy is it?
Sam Nadler [10:45]: It's really easy. In fact, let's transition because I've got a product that is yet to be built, I wanna demo how Cowork works if I can.
Jordan Metzner [10:54]: Yeah. Let's go.
Sam Nadler [10:56]: Okay. So I have an idea that I have kind of queued up here. So Cowork, what is Cowork? Cowork's gotten a lot of, I think, hype in the past couple days. It's basically just, like, wrapper around Claude Code. I think it's a entry point for almost anyone to begin building their own, products, but it's super simple to use. It's via, as of right now, I believe, the Anthropic Mac desktop app. You can see chat up here, Cowork, and then directly Claude Code, but it's pretty much the interface of chat that you're used to. And, basically, I did a tiny bit of work to get started. Basically, I wanna build a Hollywood fitness personal trainer, like a personal trainer who's, like, really focused on, like, aesthetic body transformation, not necessarily overall health, but to make me look as, sexy and buff as possible. And that's, like, the mentality. So, like, really progressive overload and really focused on just building muscle for, like, my premiere as a Hollywood star. So that's, like, the con to the PRD. I, gave all I gave Cowork access to the docs. So that's what's, I think a little bit different of Cowork.
Sam Nadler [12:12]: You provide access to a group, a folder so it has the context, and I just wrote build this. And it says, I'm happy to help you build. What do you want me to build? I said, PRD and doc. Check the PRD. I've reviewed it. This is a mobile web app, aggressive upper body hypertrophy. It's a mouthful. Hollywood transformation trainer. How do you wanna build it? So this is, like, what's kinda nice about Cowork. It proposes these trade offs for you in a nice UI. The terminal does the same thing, but I'm gonna say React and Supabase. Let's see what else. Should I integrate a real LLM API for the training responses or similar API with prebuilt logic? Let's go Real, and then let's go. I mean, we may edit a little bit of this out because it does take some time, but, I mean, it's pretty simple on here's it's working on the plan. And oh, progress. So setting up all these so this is what's kinda nice about the Cowork feature. You kinda get an understanding in a bit more polished design about what it's doing and where it's at in the process.
Sam Nadler [13:15]: And then, it literally but from my understanding, Cowork is literally just Claude Code. So I know you're a Claude Code terminal guy and probably haven't used, Cowork that much, but any thoughts on kind of the its approach to nontechnical users?
Jordan Metzner [13:30]: No. I mean, I've used Cowork. I think it's a great tool. I love that you can just dump a bunch of documents in. Honestly, I don't really care what interface you use. It's gotten better and better. I even see these, like, recommendations. There's, a new feature that didn't have it like a few days ago. But honestly, doesn't really matter. I think like you can I think what does matter is using like the best tools at the time, and it seems that like these Anthropic tools are the best? And like if your preferred interface is the web or your preferred interface is the terminal or it is this app, doesn't really matter. But at the end of the day, it gets you building. And I mean, they're pretty efficient building tools. I think the reason I like the terminal a little bit better is because, like, can you duplicate this? Can you have another window open at the same time? Can you do like more things than once? Whereas like Claude Code, you can literally have as many of these as you wanted running simultaneously. So I think like, it just have a little different form of freedom, but it doesn't really matter. Like, if you as long as it's like a gateway drug to getting like more stuff built, and I mean, look how efficient it is. I think I told you I have a skill, and I can share it with you, but you can use my skill, and you can get all the design features, and like layouts that like I like to use in my projects, and you can copy them.
Jordan Metzner [14:41]: And then, I think like we're going to like, we're gonna have kinda like enterprise skills that we can use like across the board where everybody's sharing these same skills and gets to, have products that work hand in hand together because they don't have any questions about like, what they need to do.
Sam Nadler [14:56]: Yeah. Totally. So, I mean, you make a great point. I don't think you can open multiple windows at the same time. Maybe you can, but that is the great thing about operating Claude Code out of the terminal. You can literally have, five, four, five terminals open working on different projects. It is, some time in between. So you're just kinda switching between each terminal window, providing context, testing, going to the next one, providing a bit more, context, testing, actions, boom. It's literally just kinda managing a team of little Claudes out there, and, like, the pace when you're doing it like that is way faster than kinda waiting for this to go through the progress and the plan.
Jordan Metzner [15:36]: Yeah. Totally. But honestly, it doesn't really matter. It's all progressing so quickly and, like, what's today's interface may not be tomorrow's interface. I'm just like, you gotta be willing to kind of really willing to change the way you work and, like, workflows consistently to, like, keep up with the tool set, you know. I mean, this type of like chatting versus with agents didn't exist. Right? And so, yeah, it's just a new fun new way to build. Okay. Let's just finish up the last news story real quick. I mean, considering what we're talking about all day, you know.
Sam Nadler [16:06]: Yeah. Let's, let's maybe come back. Let's see if it builds. But in the meantime, Anthropic closes the latest funding round above 10,000,000,000, still open, it sounds like. I think Anthropic is just on a you know, it was, like, a month and half ago Google was on a tear. The last sixty days, thirty days, whatever, Anthropic's been on a tear. Seems you know, it's great products being built here, but, like, the capital is flowing.
Jordan Metzner [16:35]: Yeah. I mean, as you see, like, I mean, from our side, we're shifting more and more of our spend over to Anthropic, probably less of it in from Cursor, and less of it to ChatGPT. I think, like, when he comes out publicly and says, like, we're not focused on ads, we're focused on enterprise, that makes me feel good, like, as an entrepreneur, a CEO, because I know that, like, they're gonna continue to build tools to support me rather than, like, tools to market to me. And their, I mean, their models are top notch. And at the end of the day, that's all that really matters. I mean, if, like, tomorrow Google's new, Gemini tools become top notch, we probably move over to there. Right? And so I think there is this like lag and this curve of whatnot, but, as kind of early adopters as we are, we're we're constantly looking for like what's the best way to build. And, right now, it seems obvious to me that, the Anthropic models are the best way, no doubt.
Sam Nadler [17:28]: Totally. And also, like, I think yesterday, they came out with the premium tier just dropped 50%. It seems like, it's getting cheaper. Everything's getting cheaper. So
Jordan Metzner [17:38]: Yeah. Think they have a max plan at a $100 now. They still have a $200 max plan. I think the spend I mean, like, it might be getting cheaper, but we're spending more money using it. Right? It's kind of like a gateway drug in that regard. Yeah. But, yeah, the tools are so good and they're so fast and they're top of the market. And I know ChatGPT's working on a new like 5.3 model or something to that extent, and Google's working on Gemini 3.5. But, as long as Anthropic coding models continue to be the best, it seems like coding is everything because, these Coworker tools are just wrappers around their coder. Right? And so they're still, like, kind of the best tools on the market. So anyway, we'll see what happens, but it's fun to be a developer in this time as you are now. And it's awesome to build things like you built today, like the evaluation tool.
Sam Nadler [18:28]: Yeah. I mean, one last comment, and then we'll wrap up. I mean, how do you think I mean, you manage a larger organization alongside, me. How do you think about, like, now that this it's almost open a gateway to everyone to be able to build their own tools? Like, how do you think about this from organizational perspective?
Jordan Metzner [18:45]: Yeah. I mean, I think we've talked about this internally a little bit, but everyone is kinda becoming agent managers, and, like, we're trying to give everybody, Claude Cowork or Claude Code as like tools so they can build tools for their own jobs, and, start to build kind of automation processes that help them do their jobs better. And I don't think we've seen it like in our use case of like any type of like reduction in workforce, but more of like an increase in efficiency to help teams like do better. So, whether it's like helping the finance team create invoices faster, whether it's for, our sales team to be able to create like RFPs or, proposals faster or better or more exact or other things like that. Like, we've continued to see, like, increased efficiencies. I think we'll see increased spend in using Claude, but hopefully, we'll be able to build a lot more internal tools like the ones you showed today that, like, just makes the team overall better, more efficient, and kind of, the next level of that. And as we like automate your evaluation tool, as we start to build this kind of big data model and then use AI on top, like, what greater insights can we draw across that entire pipeline? So anyway, awesome episode. Great one. Great talking to you, Sam.
Jordan Metzner [19:52]: Check us out online at builtthisweek.com. You can check us out on YouTube at builtthis week. And thanks, Sam. That was a fun episode. See you soon.
Sam Nadler [20:01]: Thanks, everyone. Like and subscribe.
