Sam Nadler built an applicant tracking system front end in his first Claude Code session, running three agents at once. Plus Envato's AI image tool, tested.
A non-engineer can build the front end of a custom applicant tracking system with Claude Code in a matter of hours. On Episode 9 of Built This Week, Sam Nadler did it in his first time using Claude Code, starting from two PRDs and running up to three agents in parallel on different pages. Jordan Metzner pushed him to keep going: use Claude Code to design the back end too, so the engineering handoff is a spec, not just a prototype.
An applicant tracking system, or ATS, is the platform that shows where every candidate is at each step of the hiring process. Sam decided to build one for his team's exact use case, as a front end he could hand to their developers.
His process had a few steps. First, a very generic PRD written in OpenAI, which he dropped into Claude Code. After some initial challenges and a few rounds of back and forth, he had a first version. Then he wrote a second PRD that described their real use case and the features of their current ATS he liked, and made sure those were represented. Claude Code had that version ready almost right away.
The part that impressed him most was running several agents at once. One agent built out the jobs page to his spec while another worked on the candidates page and a third handled something else.
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.
— Sam Nadler, Episode 9
Watch Sam build an ATS with Claude Code in Episode 9 (starts at 3:14) · Episode 9 notes and transcript
All of it runs on fake data, so no personal information is shown. Sam walked through:
Some buttons were not working yet. Sam says that in about an hour it was built to pretty much how they use an ATS today, which was the goal of his second PRD. You can see more on the Sam's ATS build page.
Sam's plan was to spend a couple more hours making every page and button work, then pass it to engineers. He guessed a first live version in a couple of months. Jordan thought it could be faster. He suggested first using the prototype with the recruiting team to find the paper cuts, then using Claude Code to design the back end, the database structure and the edge or Lambda functions. That way the handoff becomes "here's the front end, here's the back end, here's the database schema," and the conversation is about implementation.
Sam says the iteration cycles are so fast that he can sit with a recruiter for forty-five minutes and implement their feedback within the call. Jordan's tip for those sessions: record the call with Granola, screenshot each page, ask ChatGPT for a list of feedback per page, and push it back to Claude Code. On timelines, Jordan hopes a standalone version could be built in a few days, or a few more weeks if it lives inside one of their larger internal apps.
Jordan's bigger question was how AI could take a custom ATS beyond simply replacing the current one. Sam listed three ideas. One is resume scoring, using the scoring tool they covered in an earlier episode. Another is AI communication features such as writing custom emails. The third is job matching: they screen a lot of candidates, and many who don't land the role in front of them could be great fits for a future one, so pulling them up quickly with AI could speed up the funnel.
Both agreed on the order: get the front end fully aligned with the recruiting team first, then build the back end. As Jordan put it, he would not design any back end until the main core feature set is built out.
Jordan noted that only about eight weeks earlier Sam was building in Bolt, as he did for the daily check-in app in Episode 7. Would he still start there? Sam says it depends: Bolt is so fast that for something small or a landing page it makes sense to try it first, but this was a big prototype.
Jordan's view is that it depends on what you are building. Sometimes Bolt or Lovable is the right fit, sometimes Replit because it has a back end built in, and sometimes Claude Code, either for a front end or a full web app with a back end. Sam's message to anyone not using these tools yet: if he can do it, you can do it.
Jordan has used Envato products, including ThemeForest, for ten or fifteen years, and they subscribe to Envato Elements for stock video, audio and graphics. Envato now has its own AI tool for images and video, and Jordan tested it with a prompt for an Instagram ad promoting new Built This Week episodes every Friday, using its enhance prompt button.
His guess is that it runs the prompt against several models, maybe four or five with variants, and returns the results. His verdict: none of them were good, and all of the text was spelled wrong. He then ran the same prompt through their own in-house image tool across Ideogram, Google, Flux and ByteDance at once. Sam noticed the Google result got the text right. Their conclusion was to keep using Envato for real stock photos, which Jordan says it is still great for.
Jordan pointed to a report from The Information that Meta was reorganizing its AI group again, soon after spending heavily on Scale AI, and noted there had not been a new Meta model in a while. Sam's read is that Zuckerberg is moving quickly and changing the org structure to match the direction set once the new hires arrived.
Jordan's view comes second-hand, from friends inside Facebook and from what he has read: he has heard of cultural issues between newly hired engineers on very large packages and long-time staff. He says that when you know the person next to you may be making $250 million, even if that is over four years, you start to question your own pay. Unlike sports, he says, it is hard to measure who is better in software, so people start to feel there are haves and have nots.
Sam clicks through every page of the ATS on screen, and Jordan runs the Envato and multi-model image tests live. Watch Episode 9 of Built This Week, where the full transcript is also on the page.
