Short answer: Jordan Metzner says Replit's built-in database, storage and keys suit small back-end tools. Replicate hosts many models behind APIs, so he pastes a model's sample call and schema into Replit's agent to add it.
Watch this answer from Episode 1 (starts at 19:16) · Episode 1 notes and transcript
From Episode 1: How We Built an AI Video Studio in 3 Days, starting at 19:16.
Sam Nadler [19:15]: Okay. Perfect. Yeah. Quick take me quickly through kind of how you built it. And then number two, I want the kind of final polished version of, it's cute making dogs skateboarding in Venice or, cows jumping over the moon, but how do we actually use this for business purposes?
Jordan Metzner [19:35]: So I'd like to use a lot of different, Gen AI vibe coding tools. I have accounts at, almost all of the different major service providers, and different tools. But, in the case of this specific tool, I used Replit. So let me just tell you a little bit about Replit. Replit is a platform that allows you to kind of vibe code, text to chat code web applications. What's cool about Replit is they have a lot of built in services into their platform. So they have things like web sockets and database and storage of environment keys and things like that. And so when I'm building small tools that are not front end only, that need a little bit of a database, that need a back end, that maybe need a little bit of auth authorization, but, maybe it's not gonna scale to like a huge organization, I usually come to Replit to build it. When I'm building things where I wanna build like a product demo or just a front end for like my product teams, I'll probably use something like Bolt or Lovable or some other tool, or even Cursor. But when I wanna build something that I need like an API and I need to make some functional calls, I'll go to Replit.
Jordan Metzner [20:48]: And, the cool thing about Replit is, like, you can kinda chat here with your agent. Here is, a preview of your web application. Here you can see, like, my server is running. And if we make any changes or add any new features, I can just hit redeploy up here, redeploy to the server. And here you can see if I go to my deployments, you can see my last deployment and then here's a link to the app which will bring me like right into here. So if I wanna add a new feature function model, and I can walk you through that now, I can show you how I do that. I think that might bring me into like what replicate is and how I use that in this case to help me build this. But yeah, any questions here before I jump into that?
Sam Nadler [21:30]: Yeah. How thoughtful on your initial prompt to get this started, how thoughtful are you? Are you like kind of drafting a full series of prompts using ChatGPT or Claude before taking that over or you just kind of vibing straight into it?
Jordan Metzner [21:45]: Sometimes I write like a really nice prompt in ChatGPT. Sometimes I'll just start with like I'm making an admin panel and just get the like, framework of the web app like built. I tried to see here if I can, find my first prompt, but I don't know if I'll find it. It's probably buried back here somewhere
Sam Nadler [22:06]: No so
Jordan Metzner [22:07]: Nobody knows. But yeah, I probably just have to build me a web app. Yeah. Sometimes I'll say, like, build a web app with the left hand sidebar just to get a framework going, and then I'll say, like, okay, we're gonna do this. Sometimes I'll say, like, what I'm gonna do is this, and I'm gonna have it do this. So I think in this case, I probably said, like, we're gonna build a web admin panel, left hand sidebar, and use different models to generate video, something like that. Most likely, probably with this one, just to recall myself, like, most likely went to ChatGPT, probably something that's, like, four o, pretty simple, and just like gave it the idea of what I want and said like, I'm making a new app in Replit, like, please help me like improve this first prompt.
Sam Nadler [22:43]: Okay. So this was your primary tool. What other tools did you use to tie this together?
Jordan Metzner [22:49]: Yeah. So, in the case of this specific application, I used a web tool called Replicate, and Replicate's a very, very interesting tool. Let me see if I can show you kind of a little bit more of how it works. So replicate, and there's a few other, like, alternatives to replicate. I think another one is fal, and you could also use something like AWS Bedrock, there's a few other, like, companies kinda trying to provide these types of services. But what they are is websites like API as a service, and they're constantly adding new models. And so what you can do, is first see, like, look, they've got tons of different models under all different types of things. So, I wanna generate images, videos, I wanna edit images, I wanna transcribe speech, I wanna make three d stuff, I wanna make music, there's just so many different, types of things. I wanna sing with voices, I wanna do voice cloning, and these are all different models that are hosted through Replicate, and allow you to have API access in just a few moments. And, maybe if we have enough time, I can maybe, like, add a new feature into our video generation tool showing you with one of these APIs.
Jordan Metzner [24:04]: But, if you pick any of these models to use in the case of Seedance what we used earlier, you're gonna get an interface something like this with the playground. And in that playground, what you can do here is, type in some prompts and plus press play and run and you'll get your sample video. But where it gets, like, very powerful is you can go over here to the API page and, here, I'm not gonna expose my token, but, you get your API token here, You get a sample of the call API call, and then you also have the schema both in, a table and a JSON. And so as simple as just literally, like, kinda copying, like, all of this and then as well as copying the schema and then just going into Replit and saying, like, hey. Please add this model here. It would simply add, this model right into my app and then pretty much work right out of the box. Sometimes I have to go back and forth, make some testing as you guys saw, maybe earlier with the gen images. Looks like, some of these models failed, so I'm gonna have to go back and, kind of understand why they failed. Here you can see images disappeared, so, they're maybe not storing on the server properly, but, I'll keep working on those features and get it better and better. But, happy to walk through like adding another model if that wasn't clear.
Sam Nadler [25:28]: No. Think it's clear, but so every time we generate a video, we're paying replicate.
Jordan Metzner [25:33]: Yeah. Replicate we're running replicate services, so, I'll try to see if I can, show that or I think it's on my dashboard maybe. Let's see. So here you can see kind of, every time we run a different model, a few minutes ago, we did this MiniMax image model. Let me see if I can show you the cost. Here's the dog in Italy and here's a link to that. Right? Mhmm. So here that API call succeeded. I don't know if it has a cost here, but it does somewhere in the app. Yeah. Perfect.
