One Month Recreating Viral Videos With AI: My Case Study

May contain affiliate links; I may earn a commission at no extra cost to you. Everything here is my opinion. Full disclosures

On September 27 I posted a video of me and the Jean Phil meme persona doing a trending two-person dance. Neither of us danced. An AI model called Genjutsu, on Higgsfield, moved our faces and bodies through the original choreography, from a reference video and two character sheets that Claude Code built for me on my own computer.

That one reel got 64,488 views on Facebook. In the twelve months before it, my entire profile got 51,076.

I kept going for a month, wrote down every number, every cost and every failure, and turned it into two ebooks. This post is the short version of that case study.

The Numbers From the First 14 Days

Facebook is where most of my audience lives, so that is where I measured first. These are the figures from my Professional Dashboard, captured on October 11.

Facebook profile	12 months before (Oct 1, 2025 to Sep 26, 2026)	14 days after (Sep 27 to Oct 10, 2026)
Views	51,076	340,427
Engagement	4,874	17,537
Net new followers	10	821
Watch time	1 hour 11 minutes	31 days 14 hours

The best normal post of my previous year, a video of me packing for Brazil, got 4,152 views. Then the recreations landed:

• The fruit shop that didn't hire a dancer (Oct 9): 119,484 views
• Singing in Chinese (Sep 28): 80,427 views
• The sequel (Oct 2): 65,526 views
• The dance duet (Sep 27): 64,488 views

On WeChat Channels, the reel of the Jean Phil meme persona singing in the rain, hood up in a tea field, reached 1,775,142 plays and 43,945 shares, and brought 3,599 new followers. On Instagram, my daily reach went from 2 to 38 people a day before the recreations to between 411 and 6,090 after.

What One Recreation Actually Takes

People see a ten-second clip. Behind it, for every video, Claude Code ran the same production line on my computer.

1. Study the original. Claude downloads the source video, reads its length and frame rate, and makes a contact sheet of one frame per second so we can see where the bodies overlap and where the AI is likely to struggle.
2. Build a real reference set of my face. Claude scanned 4,404 image files on my computer, found the 277 photos that matched me, and threw out every AI-edited or filtered picture of me. Those teach a model the wrong face. It kept 19 real photos and 12 face crops.
3. Write a character sheet. A numbered list of my features with the hair first (models love to give me dark curly hair), the files to upload, and the mistakes to check for afterwards.
4. Drive Higgsfield. Claude opens Genjutsu, uploads the reference video and the images, writes the prompt, and stops before the Generate button to tell me the cost. My runs cost between 119 and 182 credits each. Two things stay mine: approving the spend, and clicking the consent screen for every face in an upload.
5. Check the result like an editor. Text first (AI garbles signs and lyrics), then the face, then the lip sync, hands and clothing. Claude pulls frames, crops faces and builds a side-by-side file so I can judge in one look.

Not every run works. My first raincoat test melted the face. A dance-studio test stretched the bodies. Those failures are in the book too, with what changed in the next run.

Views Are Not Trust

By the end of the second week, 96.9% of my views came from people who did not follow me. That is the point of a recreation: it travels far outside your circle. It is also the trap. Strangers who laugh at a video do not know you, and nobody buys from someone they do not know.

So the system I ended up with has layers. The recreations bring reach. A short "how I made this" breakdown, posted only after a video takes off, turns the curious into subscribers. Unscripted talking videos, one take, no AI, build the trust that AI cannot fake. Products come last.

The month also brought my first real pile-on, a verified creator's public reaction and the wave of comments that followed. I documented that case from the first hour, and it became its own chapter, because anyone who starts posting AI recreations will meet it sooner or later.

I Wrote It All Down: Two Ebooks

Already Viral: How to Make Viral AI Videos is the strategy. About 232 pages built on this month: how to score a format before you spend a single credit, the Recreate + Twist Loop step by step, the twist that makes a recreation yours, reading your numbers in the first 24 hours, the breakdown that turns viewers into subscribers, and the chapter on handling hate. It comes with a free Starter Kit and a Hooks pack with 110 hooks and 24 twist formulas. I am finishing the last details now, and it launches soon.

The Definitive Guide to Claude is the machine. Already Viral tells you what to make; the Claude guide teaches you how to make Claude do the heavy work, from your first prompt to Claude Code running a full production line like the one above. I am finishing the October edition right now, with a new chapter on exactly how I use Claude Code to make these videos with Higgsfield. When it is ready, buyers download it from the same page they already have.

The Claude guide is available today, in English and in Portuguese.

FAQ

Do I need to be on camera to recreate viral videos with AI?

No. The face in the recreation can be yours or a character's. I used my own face from real photos and the Jean Phil meme persona. What I did need was a clean set of real reference photos, because the model copies whatever you give it, including the flaws of filtered pictures.

Is it OK to use another creator's video as the motion base?

It is the method I use: another creator's video gives the motion, and everything else becomes mine. I swap every face, add my own twist, turn on the platform's AI label, credit the trend, and use audio I have the rights to. I never put a real person's face in a video without their consent.

How much does one AI recreation cost?

My Genjutsu runs on Higgsfield cost between 119 and 182 credits each, for clips of about 16 to 25 seconds at 720p. Higgsfield shows the exact cost on the Generate button before anything runs. Claude Code comes with Claude's paid plans, starting with Pro.

What does Claude Code do that a chat app can't?

It works on your own computer. It can download a video, run tools like ffmpeg, scan a folder of photos, write files and drive a browser, step by step, while you approve the important moves. In a chat app you would do each of those steps by hand.