How Jordan built a webcam Halloween costume app with Nano Banana and Veo 3.1 in Replit, plus how we schedule posts in Publer and first takes on Atlas.
You can build an AI Halloween costume app by chaining two Google models: Nano Banana turns a webcam photo of you into a costume image, and Veo 3.1 turns that image into a short happy Halloween video. Jordan Metzner built exactly that in Replit for Episode 18 of Built This Week, and showed it live to Sam Nadler. The episode also covers how we schedule social posts with Publer, OpenAI's Atlas browser and Google AI Studio's new Build tool.
Jordan says the idea came from Paige Bailey, the DevEx engineering lead at Google DeepMind, who posted a small demo built with a webcam, Nano Banana and Veo 3.1. He liked it, built his own version for the team, and promised to shout her out so it was clear the idea was hers.
The app is a multi-step process. It uses your computer or phone camera to take a picture of you. You then pick a predesigned costume, such as a pirate, superhero or ninja, or type in your own. Nano Banana generates you in that costume, which Jordan says usually takes about 20 to 40 seconds. That image becomes the lead image for a Veo 3.1 generation that says a happy Halloween message. You can download the image and the video, and there is a gallery of past results.
On camera, Jordan became a ninja and then an 1800s gunslinger. He pointed out that the face is not exactly his, but the body movement and hand position were close, and the background was removed. Sam says the team had a lot of fun with it, and one teammate told him a video inspired what she was going to be for Halloween.
Watch how the costume app was built in Episode 18 (starts at 9:45) · Episode 18 notes and transcript
Jordan built the app in Replit. The APIs are Nano Banana from Google and Veo 3.1 Fast, both served through replicate.com, where he already had an account and API keys stored in Replit. He describes it as not a very complicated application: turn on the webcam, capture an image, send it to the Nano Banana API, then use the returned image as the source for a video with a short Halloween prompt. You can read the full breakdown on the Trick or Treat AI costume builder build page.
He was candid about the flaws. The output is not a perfect replica of him, and the characters often mumble "happy Halloween" rather than saying real words. He compared it to early AI image generation:
It kinda reminds me of how we used to see a lot of the lettering and image generation with AI not be real words, and most of that's kind of gotten solved to this point.
— Jordan Metzner, Episode 18
He expects steady improvement and noted that Sora 2 could also do this job. He went with the Google stack simply because that is what Paige's example used. Sam saw a professional angle too: we already use Veo 3 for b-roll to supplement our original content, and turning a plain webcam shot into publishable video is much faster with these tools.
Sam describes Publer as a central dashboard to push content across multiple platforms in a systematic way. Built This Week alone has LinkedIn, YouTube for Shorts and the podcast, X and TikTok. Uploading everything from one episode account by account would be a hassle. In Publer you queue it all up, get help writing posts from AI integrations, add hashtags, schedule, and check analytics.
Jordan calls it a social media syndication platform and is clear it is not the only one. Still, he says it is the only way we could publish at our volume. On the Publer calendar he showed days with 19, 8, 18 and 14 posts across our brands.
If we weren't able to use a tool like Publer, we would probably have our social team be spending hours and hours a day just to get content live, and that would just make it impossible.
— Jordan Metzner, Episode 18
The workflow he walked through: connect your social accounts (Instagram, LinkedIn, Facebook and other major platforms), add teammates with permissions per account, create a post, pick the channels, use the AI assistant, preview it, then schedule it with a time and an optional recurrence. One nuance he flagged: platforms like Instagram need a business account rather than a personal one.
Sam says we tried another tool before, and Publer is the one we stuck with for its price point and tooling. Jordan says that earlier tool was Buffer, and that Canva also has a built-in scheduler. In his view these tools are inexpensive and a good way for almost any company to scale its posting across channels.
Sam added a practical example. One teammate on our social team took an extended European vacation for a couple of weeks, and because the content was preloaded and scheduled, we never missed a beat. Jordan summed it up:
The brands never stop posting, and they don't know who's on vacation and who's not on vacation because you're just being very consistent in getting your content out there.
— Jordan Metzner, Episode 18
Jordan downloaded Atlas the day it launched and used it side by side with Chrome. He sees a few things that make it competitive. First, it feels lightweight and fast, while Chrome, which he calls the best browser on the market, is heavy on system resources. Second, if Atlas can't match what you type to a domain, you land in a ChatGPT conversation instead of Google results. Typing Gmail takes you to gmail.com, but typing "built this week podcast" gives you a ChatGPT answer. He calls that the first major threat to Google.
Third, shopping is integrated, so you can check out without leaving the chat. Last, there is a ChatGPT side panel that summarizes pages. Jordan found it rudimentary for now, but expects future versions to get to the point where you rarely touch the page and just tell the bot what you need.
Sam used Atlas to shop for a family passport holder for four people on Amazon. It asked a few questions about budget and features, then surfaced two or three options with a recommendation, so he didn't have to search himself.
Jordan says memory is the key feature, since Atlas asks you to enable it. His example is an Airbnb search for a repeat family trip to Destin, Florida. Today you redo the same filters every time: a house, a pool, near the beach, Wi-Fi, a hot tub. With memory, he imagines telling the agent you are going to Destin again and getting the top three houses back, already filtered and checked against your usual budget. When agents start to work really well, he thinks they will become superpowers for people.
Google AI Studio Build lets you build apps in the browser the way Lovable, Bolt and v0 do. Jordan says Google describes it as powered by Gemini 2.5 Pro, and he counted 16 premade templates, including animating images with Veo, a customer support chatbot, Google Maps data and audio transcription.
Live on the episode he asked it for a pretty weather and time app for Los Angeles and Nashville. While it builds, it suggests features to add, such as weather alerts or an hourly forecast. The build finished quickly, and Jordan called it the fastest thing we had built on the show so far. From there you can deploy, push to GitHub, download, share a link and test on different devices. We went deeper on Google AI Studio the following week in Episode 19, building a magic mirror, a costume guesser and a trick-or-treat route planner.
See the ninja and gunslinger videos, the Publer calendar and the AI Studio build on camera in Episode 18 of Built This Week, where the full transcript is also on the page.
