Realistic AI UGC ads come from a sequenced workflow of real customer research, a phone-style reference image, and a malleable studio app that Claude Code operates and extends for you.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
A marketer or founder running paid social who wants to test many talking-head UGC ad variations without hiring creators for each one.
A Claude Code user who wants to see how an agent can operate and modify a custom internal tool instead of clicking through it by hand.
Someone whose AI avatar videos keep coming out plasticky and who wants to know what reference inputs fix that.
SKIP IF…
You want a no-tools, no-terminal walkthrough; this assumes you are comfortable giving Claude Code commands and API access.
You need a neutral tool comparison; the demo runs on the presenter's own company platform (Graph) throughout.
TL;DR
The full version, fast.
Good AI UGC ads need a sequence, not a single prompt. First research real pain points and desired outcomes from Reddit, social and TikTok so scripts are not generic. Then write scripts, record a voice read, and pick a reference image pulled from a real phone-shot TikTok frame and remixed so the avatar looks filmed on a phone. Feed script, audio and image to Seedance, bulk-generate in parallel at 480p for cost, then add word-timed motion graphics. The presenter runs this through an open-source UGC Studio that Claude Code can operate and extend, such as adding a resolution slider.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Seedance quality claim, the phone-shot look via reference images, and a note about the open-source tool.
00:58 – 02:06
02 · The five-step workflow
Research, scripts, voice read, reference image, Seedance generation, then editing with blur, upscale and motion graphics.
02:06 – 03:30
03 · Researching pain points
Use a tool router with Reddit, social and TikTok data so Claude Code writes from real customer language.
03:30 – 04:30
04 · Launching UGC Studio
Give the open-source skill to Claude Code and have it build the studio on the Graph cloud.
04:30 – 06:40
05 · Building a realistic avatar
Scrape TikToks, take a frame, remix it with an image model, then upload avatar and voice via the agent.
06:40 – 08:35
06 · Scripts and bulk generation
Generate 10 scripts, attach them to a campaign, pick avatars and bulk generate in parallel with a cost estimate.
08:35 – 10:20
07 · Motion graphics
Add word-timed motion graphics to finished videos, several at once.
10:20 – 11:14
08 · Malleable software and CTA
Ask Claude Code for a resolution slider, then the pay-as-you-go pitch and subscribe ask.
Atomic Insights
Lines worth screenshotting.
Asking an agent for ad scripts with no source material produces generic copy, so research customer pain points first and write from that.
AI avatars look plastic when generated from a prompt alone; giving the model a real phone-shot reference image is what restores the soft, natural look.
Pull a frame from a real TikTok in your target setting, then remix it into a new person so you borrow the aesthetic without using the actual creator.
Rendering at 480p keeps bulk test batches cheap; you can raise to 720p or 1080p once a concept wins.
A tool router with bundled data sources replaces separate subscriptions for social scraping and search APIs, billed pay as you go.
When your software is an open-source app an agent can edit, a feature request like a resolution slider is one sentence to Claude Code.
Bulk generation shows the estimated cost before it runs, so you can gate spend per batch.
Word-by-word transcripts let an agent place motion graphics at the exact moments, much like a manual After Effects pass.
Takeaway
Sequence beats prompt for AI UGC ads
WHAT TO LEARN
Believable AI UGC ads come from real customer research and a real reference image, with an agent running the repetitive steps.
Source scripts from actual pain points and desired outcomes scraped from Reddit and social, because an unprompted agent writes generic copy.
Use a frame from a real phone-shot video as the avatar reference, remixed into a new person, to avoid the plastic look.
Match the reference image to who you are selling to, such as late-20s creators in an office in a big city.
Generate test batches at 480p in parallel and check the cost estimate first, then raise resolution only for winners.
Add motion graphics from a word-level transcript so graphics land on the right words, across many videos at once.
Treat tools as editable software: ask the agent to add a feature such as a resolution slider rather than waiting on a vendor.
Glossary
Terms worth knowing.
UGC ad
An ad styled to look like user-generated content, usually a person talking to camera on a phone, used because it feels less produced than a polished spot.
Seedance
A video generation model that takes a script, an audio read and a reference image and outputs a talking avatar video.
Tool router
A single service that gives a coding agent access to many third-party tools and data sources through one account and usage-based billing.
Reference image
A still photo the video model follows for the look of the avatar, such as lighting, softness and setting.
Super Whisper
A dictation app the presenter uses to speak prompts to Claude Code instead of typing them.
“Software is this malleable thing that you can build and add to based off your personal workflows and desires.”
Short thesis on agent-editable software→ newsletter pull-quote↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
Seedance 2 .5 has completely changed the ability of the quality of creative that you can make using AI avatars to do talking head UGC. But you have to build a workflow that goes through a sequence of steps to get the quality that you're actually looking for that will actually make impact when you're using these as ads. And today I'm going to teach you this exact workflow.
And then I'm going to also show you how you can use this open source tool that enables you to make these AI avatar videos at scale. using this exact workflow and you can customize the avatars and I'm gonna show you how to make these avatars look and feel like they're actually shot on the phone so that they don't have that plasticky feel that you get and the way to do this is by providing reference image don't worry I'm gonna go through all this and if you're new here I'm Cody Schneider I'm the co -founder of a company called graft I've been doing digital marketing for the last 10 plus years of my career I do videos like this every week where I teach you how to use AI to do growth and distribution for your company and if you want more content like this go and subscribe to the channel now the link is in the description below all right let's jump into it so first let's talk about the workflow that we have to make to create a good outcome so this workflow is going to combine a couple of things we need to research the pain points and desired outcomes of our target customer so that's going to be step one that we're doing and then step two we're going to go and we're going to write scripts based off of these pain points and desired outcomes we're going to do a script read using a voice model we're going to find a
static image as a reference image for our AI avatar. And then we're going to provide the script, the MP3 of the AI voice read and the reference image to see dance to generate the final image and then the last thing we're going to do is edit the video by upscaling it adding a little bit of blur so that it feels like it was filmed on a phone and then also adding motion graphics to this all right so we have a lot to get through here uh let's start from the beginning why are we researching the pain points and desired outcomes we do this so that we can identify what the ad should even be talking about if i go ask an agent write me a script it's gonna come up with something that's super generic and so instead what we have to do is go and give it source material to write scripts based on.
How you do this is you go and you extract information from Reddit, from social, from TikToks, et cetera. You can use these data sources as a way to collect that information. So if you want to collect this information, you need something called XAI or you need something called Friar Crawl to go and source this information.
But today what I'm going to be using is this company called graph .com that has a tool router that has all of these built into it. So I don't have to go get multiple subscriptions. pay as I go and use this basically.
And this tool router, once it's installed, I can go and I can literally prompt it, use the graph CLI, go and research the pain points and desired outcomes for somebody that would buy tiny cow. And then I just pasted that in and I'm using a transcription software called super whisper. If you're wondering, and we'll say tiny cow come.
All right. And I'm just going to hit go. And this tool router that I'm using, my clog code now has access to everything that's in there.
It can go and do. whatever research that it needs using those tools as it sees fit or as it sees what's necessary so imagine we just went through that process and we now have scripts that we want to generate so let's spin up this platform to be able to do this so we're going to do this skill called ugc studio and all i have to do to launch this skill is again just give that to my cloud code and say build this and the skill is open source so under the hood It needs all of this different tooling to be able to actually host it.
Again, you can go buy all this tooling. You could buy something like a railway. You could get access to seed dance.
You can get access to modal so that it can render the videos quickly. So anybody could use the skill and get all the underlying infrastructure that's necessary to actually do this. But again, just for today, I'm going to be using the graph cloud for this.
So I would say build this on the graph cloud. But if you don't want to do it in this way, all you do is say, hey, hey Opus, investigate this.
What do I need to buy and give you access to to be able to build? Okay, so I hit enter. That will build and launch the UGC Studio that looks like this.
And at this point, there's all of these preloaded avatars that are in the platform. But I can add my own avatar in here whenever I want. I can go click, hit avatar, name it, and I just give it a source image.
Now, this is where it's critical to pull a source image that accurately represents who that you're trying to sell to. So how do you do this?
The best way to do this is, again, we're going to go and we're going to source information using a tool called Scrape Creators. I'm going to say Graph CLI, use Scrape Creators, extract TikTok reels of late 20s creators. I want them in an office setting in a big city.
And it's going to go find TikToks that have people that match that setting. And then I'm going to take a frame from that video. I'm not going to use that actual person.
But what I'm going to do is I'm going to remix that video into a static image like this that we can use as the source image for the AI avatar. And why do we have to do this? Because otherwise it makes these plastic looking people that don't look real.
If I just ask it to generate this person out of the box, I have to give it that source image and have it follow it. Once I found one of these videos, so these images, I would then go and say, use a frame from this video, use chat GPT image 2 .5, and I want you to remix. that person, but keep the same aesthetic on the photo, the softness, et cetera.
We're just trying to turn them into a different hair color, a different variation, something along those lines. And once I have that image reference, at that point, I can go and upload that avatar. Now, what's really interesting is this whole piece of software I don't have to interact with.
I can have my cloud interact with it for me. So I could say the image that I just made, that's the reference image. I want you to upload it to the UGC studio.
library and then I also want to make a voice using the voice from the video that we pulled as the reference voice for this avatar. And when I do this prompt together, it will go and modify this software for me. This is malleable.
customizable software that anybody can go and make changes to. And your whole team can work on top of this. This is why this is the future of how this all works and all happens is basically teams go and they co -work in this way, constantly improving the outputs of the software.
So that's how I would add an avatar and add a voice. And again, why are we doing this? Because we have to do this sequence, right?
We have to go from pain points down to the scripts. First, we had to like build this custom avatar. So think about another section that was right here.
Now we're going to have the script read happen and then all this is going to be sent. So let me show you how easy this is now. So say I uploaded my avatar.
I'm just going to use one of these for this, but say Maria. Actually, let's go and do script. I can add this script to a campaign, but we're just going to do a totally new campaign.
I'm going to say add avatars and I'm going to do Maria in this hoodie. And then I'm going to say add script. Let's go back actually to the scripts that it wrote.
And I'm going to say write scripts based off of these pain points and desired outcomes. I want 10 of them. And I now have these scripts.
You can see these here, but I don't want to have to upload these in to this campaign. So what I can do is I can take that campaign URL and I can just paste it in and say, add these scripts to this campaign. And let's rename this campaign.
We'll call it Tiny Cal. All right. So just to show you different avatars that we can then source from, here's examples of these that we can use.
And if we take these and we have chat GPT image 2 .5, remix them, this is the way that we can get this soft office look of somebody in a cubicle that was filmed on their phone all right so you can see now those 10 scripts were uploaded into that campaign and i can add more avatars say i want to add may let's give aaron in the living room actually let's do aaron in our car do fray outside add those three and if i want to generate these i now just select them or i could select them all and i can go and bulk generate these at scale and this will do this in parallel and before it even does the generation it'll tell me the number that's in of what this will cost.
But this generation, it's about to go through this whole loop. It's about to do this whole thing for me to get this raw video file output that I'm looking for. So let's hit generate.
Let's do it. And again, all of this is just API calls to underlying providers. You do not have to use these providers.
You can go and get your own subscriptions for all of these providers individually and hack this together yourself. But we'll hit generate now. And it's going to start generating those videos.
All right, so that video has been generated. And now say we want to add motion graphics to this. Super easy for us to do that.
We're going to use this thing called Tesseract. I'm going to just say add motion graphics to this. Use graphed word by word transcription to know when and where to put the motion graphics.
And because there's this motion graphics skill that's wrapped up within that skill file, it's now going to be able to add motion graphics to the video just like you would do in After Effects. And you don't have to do this individually one by one. You can do multiple.
of these at the same time so all four of these that are still rendering we could take these and add motion graphics to them simultaneously all right now here's an example of one of these videos that has these motion graphics added to it and you can add and make this style motion graphic to every one of these videos that you create and maybe you're looking at this and you're like oh the quality isn't good enough i don't like this this was rendered at 480p originally so it's cheaper but say i wanted to do it at 1080p i wanted to add that functionality i can just tell cloud code to add that functionality to my application so i'll take this and i'll say on a campaign page i want a slider button that allows me to select 480 720 and 1080p for when i'm doing the video selection for the generation and now this feature that i just asked for it's now going to go and generate and add that feature for me and this is how crazy this is getting now is that software is this malleable thing that you can build and add to based off your personal workflows and desires of what you're trying to get to happen all right so here's another
example where it just made all these motion graphics over the talk track and it knows where to place these because the word by word transcription and again it's just doing a generic style but you can give it any style and tell it to make it like the style that you provide All right, you can see now it's a couple minutes later and I can go and I can select the videos that I want to generate.
And then I can also select the aspect ratio or basically the quality level that I want to do those generations. All right, and if you want to do this exact same thing, just go to graph .com and get started for free. It's all usage based.
You put in $10 and you can go and do these generations. There's no subscription. And this open source studio that I showed you today, it's in the description below.
I'll give you a link to that code. Just go and give it to your client. log code.
You can go and make your own version of this, customize it, do whatever you want with it. And if you like this video and you want more like this, please subscribe to the YouTube channel. It's what keeps the lights on here.
And comment below if there's anything that you want to learn. It could be about making ads. It could be about doing the actual media buying using a coding agent, cold email, AI search, you name it.
But I'll prioritize what videos I make next based on this. All right, we'll see you next time.
The Hook
The bait, then the rug-pull.
Seedance made AI avatars good enough for ads, but only if you stack the right steps in front of it. Cody Schneider shows that sequence live, from pain-point research to bulk renders.
Frameworks
Named ideas worth stealing.
00:58list
AI UGC ad pipeline
Research pain points and desired outcomes
Write scripts from them
Voice read with a voice model
Pick a phone-style reference image
Generate with Seedance
Edit: upscale, blur, motion graphics
The sequence Cody says is required to get UGC-grade quality out of AI avatars.
Steal forany paid social creative testing
CTA Breakdown
How they asked for the click.
VERBAL ASK
10:19product
“Go to graph.com and get started for free. It's all usage based. You put in $10 and there's no subscription.”
Soft product pitch woven into the demo, followed by a subscribe ask and a comment prompt for next topics.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A 29-minute walkthrough of wiring a coding agent to a data pipeline, a data warehouse, and a marketing API gateway so it can write ads, pull cross-platform data, and ship blog posts on its own.
A creative strategist wires Claude Code to Slack, Notion, and an ad-research tool to build a self-updating brand brain that mines real reviews for ad copy and files its own reports.
A 52-minute masterclass on the environment a coding agent needs to run marketing: a data warehouse, write-only ad API access, and a stack of generation tools, then live static ads and a $5 UGC avatar ad.
A HubSpot podcast interview that turns into a live Claude Code build session, wiring a Facebook ads agent, a Google Ads agent, an SEO agent, and a cold-email workflow into one voice-driven marketing stack.