How to run your first AI UGC campaign (step-by-step guide)
In a Nutshell
The video demonstrates how to generate hyperrealistic AI UGC content for $0.20/second instead of paying $20-50 per video by using JSON-based prompts to create authentic-looking characters from trending references, then leveraging Cance 2.5 with audio references to produce talking-head videos that bypass model restrictions via Track's endpoints. The workflow involves analyzing trending hooks from competitors, generating reusable character skills, and pairing demo clips with multiple hook variations for A/B testing at scale. Track serves as the central platform providing cheap access to 3,000+ data/tools including image/video models, with a complete UGC talking head skill available in their GitHub repo.
These notes were generated by AI and may contain inaccuracies.
When you see a video on Instagram or TikTok that appears to be a real user sharing a product they like, it's most likely UGC content created by the product company. Companies typically pay $20 to $50 per video and sometimes even thousands of dollars for just one piece of content.
The basic UGC strategy involves finding people with time on their hands to create new social media accounts dedicated to posting about your app, paying them $20 to $50 per video. The payment is not for their audience since they usually don't have one. The videos are not posted on the company's own account because the goal is for them to appear as normal people who discovered the app organically.
After researching and finding examples of UGC campaign videos, they were extremely low effort but worked effectively. Given the capabilities of models like Cance 2.5, there must be a way to produce hyperrealistic UGC content. After testing various approaches, a workflow was developed that produces hyperrealistic UGC content at $0.20 per second instead of $20-$50 per video. All UGC content is produced by cloud code or codecs directly using free plugins like Track, an open source project for agents to access different data and tooling.
The process begins by finding videos that are working well in your vertical and analyzing them. UGC content consists of two parts: the hook (most important) and the content demo showcasing the product. The same product demo can be paired with 5 to 10 different hooks to test variations. By watching trending content in your vertical, you understand how to design hooks and identify which faces or characters resonate with your audience. The same script with different characters can significantly change the impression.
Track is a tool that full contact uses because agents need access to different data and tools, but most data is locked behind paywalls costing hundreds of dollars regardless of usage frequency. Track requires connecting only one key to access various services, paying only for usage rather than subscriptions. Since it's used for their own company, it has rich data for agents to find what's trending and working on social channels or from competitors.
To set up Track, sign in, choose the agent you're using, and copy the prompt to your agent. After setup, give the prompt: "Help me pull recently trending videos in our verticals AI tooling." Within one minute, it provides a list of videos trending recently on TikTok and Instagram for the specific vertical.
From trending videos, analyze hooks, find characters you like, and create AI characters that capture that vibe but are different people. Among the videos, one character appeared smart and approachable. The goal is to take an existing character you like, capture their vibe, and generate your own version to create accounts based on that character for UGC content.
Generating realistic videos is easy, and making realistic videos based on existing pictures is also straightforward. However, the difficult part is generating videos with a character that has a specific vibe you like. Attempts to create AI characters based on a Google Cloud YouTube channel presenter using GP2 image 2.5 and Gemini 3 produced results that didn't look like real people.
An effective prompt from Harry changed the output quality. Using the right prompt generates characters that capture the original person's vibe and look real, not like typical AI-generated images. The key is restricting the model's exploratory space because default batch models like GV image 2.5 have out-of-box characteristics like large eyes and doll faces. Large models often include quality boosters like "4K AK ultra detail" which create high-clarity HDR pictures that look like AI slop.
The solution is to output JSON-based prompts that force the model to define every single detail they would otherwise omit and revert to defaults. Testing this approach with a famous Chinese tech vlogger produced a new character that looked different but captured the original vibe. The same worked for a Chinese AI podcaster and Jen Yan from Silicon Valley, producing realistic results that didn't look like AI slop.
Harry turned this knowledge into a reusable skill that allows feeding the model any screenshot and character to output a new character that looks different but has the same vibe. The skill link is in the description, and following Harry on Twitter is recommended as he's one of the best builders.
The first trick is using JSON-based prompts that define every detail to constrain the model's explorative space. The second trick is that different models have different performance even with the right prompt. Testing the same prompt across five different models showed Gemini 3 Pro produced the best results. While JP2 image 2.5 should technically be the best, the characters had patterns where eyes didn't look real and faces looked more like doll faces. Gemini 3 Pro results are super realistic.
Track includes all major image and video generation models, allowing cloud code or codecs to generate characters at super cheap prices with no markup. You can prompt Track to generate characters using portrait clone skill with different models like GPT image 2.5, Gemini 3 Pro, or Cream to compare based on original screenshots and choose the best result.
Once characters are ready, feed them to the Cance 2.5 model using the generated image as reference. The model does a good job, but by default refuses to take hyper-realistic images to prevent impersonation. Platforms like Hickview provide endpoints to bypass this restriction but charge huge margins that double or triple costs.
The correct approach is buying the API from by plus directly, which has advanced creation rights. Even the free version has a premium creation benefit package that allows uploading images and creating digital characters to generate videos, though it requires hurdles to get approval. Track has a Cance 2.5 restriction endpoint with an automated process to upload pictures as character assets, bypassing restrictions while offering the cheapest market price with 0% markup.
You can use the Cance 2.5 endpoint on Track just like buying directly from by plus. Track's Cance 2.5 less restriction model generates videos using scripts with the character. Generally it works well with just one image, but the default generated voice doesn't sound right.
The solution is getting audio clips from original videos you like and passing them as audio references. This is the shortest way to get the video rhythm and speech feeling right because many details are hard to describe in text.
The final video example: "3:00 a.m. Last night, my agent was researching hundreds of customers and emailing each of them to book meetings for me. I woke up to 12 replies and four meetings on my calendar. Here's exactly how I set it up."
For the demo content portion, the character doesn't need to be shown. You can record a video, use 11 Labs official audio to clone the voice, output the audio clip, and merge it with the demo video. A phone was used to manually record demos of how people can use plugins to get a codec doing GTM work like finding leads and doing autoreach. A voice was created on fish audio using original audio clips, then put together with product demo clips that have real voiceover and captions.
Once demo clips are right, they can be paired with 5 or 10 different hooks for A/B testing to see which works. Track was given to cloud code to pull recently trending videos in the vertical, with reference accounts like Composio that are performing well. The model analyzed and figured out hooks, grabbing trending videos with lots of views. The model is good at analyzing audio and transcripts to extract hooks.
One video from Composio got 3.3 million impressions: "I email over 5 million people and asked each for $1." This information is useful for analyzing patterns that include stunning claims and showcasing numbers paired with results. After brainstorming options, one hook stood out: "3:00 a.m. last night my agent was researching hundreds of customers and emailing each one of them to book a meeting for me. I woke up to 12 replies and four meetings on my calendar. Here is exactly how I set up."
A UGC talking head video skill was created that takes requests, designs scripts and duration, grabs voice references, generates video prompts, and adds captions and headers to make final videos. The skill is in Track's GitHub repo. Sending the request pulls all resources and uses the Cance 2.5 less restriction model on Track.
After a few minutes, a full UGC clip was generated for only $2.67. The result: "3:00 a.m. last night my agent was researching hundreds of customers and emailing each of them to book meetings for me. I woke up to 12 replies and four meetings on my calendar. Here's exactly how I set it up. First, go to chat GBT. Search for the track plugin. Then just tell it who you're looking for. It researches every lead, writes each one a personalized email, and sends it from my Gmail."
The complete workflow goes from finding trending content and characters to creating AI characters that resonate with audiences to generating UGC content at scale. For more complicated videos than software demo videos, the typical flow involves more detailed character design with different angles, using image generation models to create storyboards, then feeding storyboards plus characters to Cance 2.5 for final videos. For most software videos, one image is sufficient.
Track was built as an open router for data and agent toolings. Connecting to one service gives access to more than 3,000 different premium data and tools across people and company enrichment, buying signals, SEO, AEO, and social. It has free endpoints for managing social and ads accounts through OOS without paying for services like Buffer, Potato, or Uni. It's fully open source, so you can host it yourself or contribute. Track's GitHub repo contains the UGC talking head video skill for free access.
Keep AI Jason in your library
Save the videos and channels worth coming back to, and find them again in one place.
