Three viral ads rebuilt with AI in an afternoon

A new tutorial dropped this week that takes three ads with millions of views and rebuilds them from scratch with AI. No studio, no actors, no budget. The creator behind it, an AI educator who teaches people how to build agent teams, picked a paper-tearing reveal, a fake road rage clip, and an AirTag-in-a-chicken stunt, then reverse-engineered each one with Claude and VEED. I’ve watched a lot of “make ads with AI” videos, and this is the first one that actually shows the messy middle: the failed generations, the models that can’t handle text, the 30 minutes of retries. That honesty is why I’m sharing it.

Quick note: the video is a paid partnership with VEED, so keep that in mind. The workflow still stands on its own.

What’s new

The expert’s core argument is simple. The three ads look nothing alike, but they share one thing: a curiosity hook that stops the scroll. Green paper being cut open. An angry guy smashing coffee on a car hood. A girl shoving a tracker into a raw drumstick and throwing it into the ocean. You have to know what happens next, and by the time you find out, you’ve watched an ad for a studio, a coffee brand, and a realtor.

What’s changed is how cheap it is to test hooks like these. The creator’s stack:

  • Claude to watch the original clip and write the generation prompt. Claude can now view video directly, so no more screenshot-by-screenshot descriptions.
  • VEED as the hub: video generation with Kling, Seedance, and Veo, plus talking-head avatars, 500+ voices, and a full editor with auto-captions and audio cleanup.
  • VEED’s MCP connector inside Claude, so Claude can list avatars, pick voices, generate clips, and stitch the final edit without leaving the chat.

The twist

Here’s the part I didn’t expect. The heavy lifting isn’t the AI video model. It’s Claude acting as the editor.

In the road rage recreation, the author generated a talking-head avatar saying “So this guy attacked my car at a crosswalk in Brooklyn yesterday. I got the whole thing on my dash cam. Watch this.” Then a single prompt to Claude: insert the dash cam clip after “Watch this,” and build a five-second animation for the burger shop at the end. Claude did the insert, built the closing animation, and delivered the cut. One prompt.

The creator frames it as the 80/20 rule. Inserting clips at a timestamp, adding an end card, smoothing transitions, combining three videos in order: Claude handles all of that now. You only reach for a proper editor when you need something advanced. I think that’s the real shift here. The prompt engineering barrier basically vanished.

The other twist: Claude also built the “chicken traveling 5,000 miles” map segment as an animated graphic. Not an AI-generated video, just a moving emoji-style animation with a look and feel the author described in text. It came out clean enough that the creator said a realtor could ship it as-is.

Step-by-step mini-workflow

Here’s how the person who shared it rebuilt each ad:

  1. 🎯 Pick a viral ad and name the hook. Pause it and ask what’s making you stay. Curiosity? Drama? Weirdness? That hook is what you’re recreating, not the exact visuals.
  2. Let Claude write the prompt. Paste the clip, tell Claude which seconds to watch, and ask for a prompt aimed at Kling, Seedance, or Veo. Copy the output.
  3. 🎬 Generate in VEED. Go to Generate Video, pick portrait 9:16, choose the model and length. Kling caps at 5 seconds, Seedance goes to 15 or more. Each generation costs credits (40 in the example), so expect to iterate.
  4. Connect VEED to Claude. Customize, then Connectors, then Add. Name it VEED, paste the MCP URL from the video description, click Continue, log in, allow access. You’ll see “connected” at the top.
  5. 🗣️ Build the talking head inside Claude. Ask it to list VEED’s characters and voices, pick one, hand it the script, confirm the generation.
  6. Stitch it with one prompt. Tell Claude the order: drumstick clip first, animated map second, talking head third. Wait a few minutes. Download.

The final chicken ad ends with the avatar saying “Comment track and I’ll send you our free home buyer checklist.” A hook, a story, and a call to action, all made in an afternoon.

Pro tips from the trenches

The author burned a lot of credits so you don’t have to. Their hard-won findings:

  • Kling can’t do text. Every attempt at the paper-reveal ad with readable words fell apart. If your ad needs on-screen text, skip Kling for now.
  • Upload a reference image. Generations tracked the target far more accurately when the creator attached a still from the original. Text prompts alone drift.
  • Seedance beat Kling for reveals. For the tear-and-reveal style, Seedance was the recommendation. Veo nailed the ocean sound effects in the chicken ad but the motion looked unnatural.
  • Budget 30 minutes per hook. Both the paper ad and the chicken ad took about half an hour and several prompts across multiple models. Some attempts “completely messed up.” That’s normal, not failure.
  • Put your own spin on it. The creator deliberately didn’t clone the angry coffee guy. Same hook structure, different character and business. Copying the mechanism beats copying the ad.
  • Use the editor extras. VEED auto-generates the bouncing captions you see on Reels and TikTok, and adding music is a drag-and-drop on the timeline.

Where I’d push back a little

The avatars look startlingly real, which is exactly why you should be careful with fake “dash cam” or “true story” framing for a real business. The original road rage ad worked because the reveal was playful. If the audience feels tricked rather than delighted, the hook backfires. Test small before you scale.

Also, this is a sponsored tutorial. The connector setup and model comparisons are genuinely useful, but there are other generators worth comparing before you commit to one subscription.

Still, the underlying lesson holds regardless of tool. Find a hook that already worked, understand why it worked, rebuild it cheaply, and test it. That loop used to cost a production budget. Now it costs credits and an afternoon.

🚀 Go watch the full video to see every failed generation, the exact prompts, and the three finished ads side by side with the originals. The comparison is worth it on its own!

Scroll to Top