How to scale your app with an AI UGC Army (Full Workflow)

AI UGC (when done correctly) is one of the greatest hacks for finding and scaling winning videos for organic and paid social.

(FYI I’m giving away 13 prompts throughout this article AND a AI UGC skill file you can use with the top video models for free, all you need to do is comment “SKILL” and follow me so I can send it to you).

And as someone who ran UGC and influencer marketing programs for Cal AI on our way to a $50M run rate, I wish we had a system like this that I’m about to share with you because it would have made me 10x more productive…

…especially as it relates to testing new video ideas on organic that we’d eventually run as paid ads.

This system works great for scaling consumer apps, but the same principles can be applied to tech UGC, ecomm, and any brand or business that is trying to stay competitive in today’s digital world.

I’m literally handing you a step-by-step guide on how to go from zero to having a full AI UGC video pipeline running (nearly) on autopilot so you can:

• Ideate and research faster
• Test faster and fail faster
• Find winning angles to scale faster
• Rinse and repeat

All you need are Claude Code/Codex, Apify, Higgsfield and Postiz.

Here’s a high level view of everything I’m going to go through in this article:

Let’s get into it.

1. Research > Everything else

“Find viral video ideas for my app” is a waste of time and credits.

Claude or codex needs to understand who the app or brand is for, what problem it solves, and what someone can SEE that makes the benefit obvious.

I’ve talked a lot about finding a marketing magic moment, which is the “aha” moment someone gets when they see your product in action and think “damn, that’s cool I need to try it.”

So before you start trying to recreate videos to scale your app, you need to give your AI of choice (mine is Claude for this example) as many layers of context as possible before asking it to make high leverage decisions on your behalf.

Garbage in, garbage out.

So start off by opening a Claude Code project folder and drop in this prompt with your specifics filled in:

This is the very basics of what you should be handing Claude before going into deep research.

Ideally you have other customer data to give to your AI of choice such as:

• Support tickets
• Reviews
• Feedback from your power users
• Customer surveys

AI can almost never have too much context.

2. Find outlier videos in your niche

Don’t try to reinvent the wheel when coming up with content ideas.

You just want examples of people repeatedly earning attention around the problem your app solves.

A large view count is one signal, but an account doing unusually well compared with its normal videos is a stronger indicator you need to pay attention.

If a creator usually gets 10,000 views and one comparable video gets 120,000, I want to understand why.

As a hypothetical 12X outlier, it’s a reason to investigate, not proof the format will sell your app.

For research, I’d use Apify alongside Claude Code.

You can try asking Claude Code/Codex to research and find top performing videos in your niche, but the results can be inconsistent compared to using a tool like Apify.

Apify exposes its scrapers through MCP, so Claude can run a collection job and work with structured results.

(Here’s an Apify MCP setup if you need that info.)

Also, the Clockworks TikTok Scraper supports research across search queries, profiles, and hashtags.

For Instagram, start with relevant creator profiles and Reel URLs, then use Apify’s Instagram Reel Scraper.

Have Claude inspect the current input schema before running either one.

Apify is a key part of this process because although Claude Code’s browser can help discover accounts and check individual posts, it doesn’t give you a dependable, built-in TikTok or Instagram data feed.

Here’s a research prompt to hand claude code/codex to start your outlier video research:

3. Turn the research into a board you can actually use

After the research is finished, have Claude build a local HTML gallery so you can watch the videos it found, compare them, and select the ones worth replicating.

Each card should show the source, available metrics, why it made the shortlist, and a selection button.

4. Reverse engineer the reason someone watched

Once you’ve picked your references, Claude needs the content itself to have any chance of recreating it for AI UGC videos.

A caption and a view count aren’t enough to explain pacing, delivery, or the product reveal.

The easiest way to run this part of the process is to use the Apify Video AI Analyzer

Ask it to break down the hook/opening scene, the unanswered question, the video sequence, the payoff, and anything else you want to know.

Descript is a solid paid option for transcribing videos that’s fast and reliable, and I’m sure there are free versions you can find on the web.

Alternatively, Gemini can view and analyse videos (but only when you give it a video file), so you can also download the TT or IG video and have Gemini handle this for you but this is obviously much more manual and tedious.

Pick you videos then run this analysis prompt:

Once this step is done, now it’s time to adapt those winning formats for your own app.

5. Design a test you can learn from

You can use AI to change the hook, character, setting, format, length, product placement, and CTA.

Changing all of them at once makes the result harder to interpret.

Once you’ve picked the formats you want to test, analyzed them to understand why they were outlier videos in the first place, you have to give each a fair attempt to get similar or promising results with your product integrated into a version of the winning angle.

Then take the strongest candidates and test more specific changes.

Here’s a test-planning prompt to run next:

Organic posting won’t give you a perfectly controlled experiment. Timing, account history, and the audience a platform chooses to show your video all affect the result.

You’re looking for a pattern strong enough to justify the next investment.

6. Build a believable starting frame

This is where a lot of AI UGC starts going wrong.

People spend ten minutes writing “ultra realistic” and almost no time deciding what the shot should look like.

The meta also used to be finding references images by searching “UGC creator” on pinterest, but even now many of the images on Pinterest are AI generated and have too much polish to feel believable.

Kristian Jennings has a useful workaround for this.

His approach starts with a frame from a real UGC creator video, then builds a new character around deliberate choices on framing, light, and setting.

The techniques worth taking from his example:

• Choose the camera angle and body position before generating.
• Look for visible skin detail, avoiding blown-out highlights on the face.
• Give the background a little personality without making it distracting.
• Choose natural hand positions and a frame without text across the face.
• Make the sound believable for the setting: a visible mic or close phone framing can help explain clear dialogue.

He also recommends getting the major image changes into one coherent prompt, then returning to the source reference for fresh attempts if successive edits degrade the result.

Treat those as production techniques to test, they won’t guarantee a perfect generation.

Here’s a reference shot-selection prompt that will help keep consistency much better for videos that have multiple shots that all need to stay cohesive:

Now connect Higgsfield’s MCP to Claude Code. Its connection lets you run all your everything in one chat without having to switch between tabs every 2 minutes.

From my experience, these are the top video models you should be using and what they are best at:

• Seedance 2.5: the most robust and reliable outputs (and the most expensive). Highest quality for AI UGC you can find
• Google Omni: a very capable model, definitely can challenge Seedance on quality sometimes depending on the request and reference images
• Kling 3.0: still a quality option, best used for b-roll footage since it is cheaper compared to Seedance especially

‼️If you want the AI UGC skill file, just drop a follow and comment “SKILL” and I’ll send it right over ‼️

Give Claude this image-generation instruction:

Starter image prompt (example):

For a multi-scene video, create a starting image for each generated shot using the approved character reference. Keep wardrobe, identity, and recurring props consistent.

7. Animate a test take before generating the batch

Seedance 2.5 supports clips up to 30 seconds. That gives you room to work, but you don’t have to put an entire video into one generation.

Another useful lesson from Kristian’s walkthrough is to stabilize the scene direction and performance before swapping in the rest of the dialogue. His demonstration uses a different video model; we’re applying that production principle to Seedance here, namely hings like:

• Specify whether the camera is static or handheld
• Describe the energy
• Write the exact dialogue being stated
• State what should remain consistent

Here’s a Claude test orchestration prompt:

Here’s an example starter animation prompt:

After your test generation, check a few things:

• Does the lip sync seem believable?
• Do the hands stay coherent?
• Does the face change halfway through?

If it repeatedly falls apart, revisit the starting frame or simplify the action before paying for more failed takes.

8. Finish the videos and approve the actual exports

The finished video may still need real app footage, cuts, captions, and a clear ending.

Record the app doing the thing the script promises.

Use that recording in the edit (don’t have the image model invent a beautiful interface your product doesn’t have).

Have Claude assemble the approved clips with a configured editor or a local FFmpeg workflow.

Review the exports at phone size and with the sound on.

I know Claude Code is getting VERY good at editing UGC videos if you’re trying to avoid using an editor, and this post on X the other day proves that.

Video proof:

Feel free to try out this video editor assembly prompt for yourself:

**

9. Use Postiz to get the videos into the market

This is where the workflow needs a publishing system, which is where Postiz comes in.

(Here’s a link to check it out for yourself: https://postiz.com/

Connect the accounts you want to test: TikTok, Instagram, YouTube Shorts, and your Facebook Page for starters.

From Claude Code, the handoff is concrete: retrieve the finished Higgsfield asset, import or upload it into Postiz, then attach the Postiz-hosted media to the draft.

Postiz supports both remote URL imports and local file uploads.

It’s also equipped with AI agents, so you can just tell it you need help creating or scheduling/posting content of any kind and it’s able to help you with whatever content workflows you’re trying to knock out.

Here’s a Postiz content draft prompt to run:

Once the drafts and calendar look right run this:

Postiz’s MCP can prepare drafts and schedules; its CLI also supports promoting a draft into the publishing queue.

Now the batch has a home and you can review it and track results all from one dashboard.

10. Let each platform tell you what to make next

Run the videos across the four channels (or however many you’re focusing on), then judge each platform separately.

A weak TikTok result doesn’t erase a promising Instagram result, and a large Facebook view count doesn’t automatically mean the app is acquiring customers.

Use consistent observation windows and compare against the account’s own history where possible.

The questions I care about are:

• Did the opening earn attention?
• Did people understand the product?
• Did that attention lead to useful action?
• Is the signal repeatable enough to spend real money on it?

Once a concept shows promise, give it to a real creator who already makes content your audience watches.

Use AI as your testing ground to find winning angles, hooks and formats, and remake the top performers with real creators to increase authenticity and trust from the viewer.

Keep producing useful variations while the data shows it’s still performing.

The advantage of this workflow is how much you test and learn before turning on ad spend.

You’ve studied what the niche watches, selected the formats yourself and tested original variations.

Start with a few formats and a batch of content you can actually learn from.

Then earn the right to scale it with paid ads.

-Jake

P.S. If you enjoyed this breakdown, you’ll love my free newsletter. You can join 2,200+ weekly readers of the Operator’s Notebook where I go deep on all things AI and building and scaling consumer apps 👇

https://operatorsnotebook.beehiiv.com/

P.S.S. check out Postiz, the agentic social media scheduler here: https://postiz.com/