I Gave GPT-6 One Prompt To Make a Viral Short

Nature facts has always been a hard niche. You need a good idea, good visuals, a great script, a great voiceover and good captions. We know the person who runs Discovery Dose, and he already uses Everygen for some of his visuals, so I asked ChatGPT for a video that could go out as his next upload.

ChatGPT ran Everygen’s video analysis on the channel before writing anything. It found a consistent structure: a direct hook, quick process details and an abrupt reveal. Then it searched the web for ideas and asked me to pick a voice. I chose the Everygen voice. Within about two minutes it had the idea, the script and the voiceover.

It split the 30-second voiceover into eight scenes, generated a starting frame for each, waited on every clip and regenerated anything that failed. I told it to use Seedance 2.5 because I wanted the most realistic visuals possible. Kling 3.0 or Seedance 2 would cost less.

For one prompt the first cut was impressive, but I didn’t like the captions and the visuals could be better. ChatGPT doesn’t have feelings, so I was direct.

There’s no special prompting here. I talked to it the way I’d talk to an editor doing revisions. It worked for 32 minutes and 16 seconds, rewrote the script, made a new voiceover and background track, and generated more clip variations.

The hook went from cochineal being “a color you may have worn or eaten” to bugs that “can turn your clothes bright red”, and the ending now lands on the scarf someone might be wearing. That version is the Short at the top of this page.

Videos like the one I linked were made before AI. An animator was paid hundreds to thousands of dollars for each one. This is the format where one prompt was enough.

I asked it to work out why the original went viral, and it did. It worked for about 10 minutes and found the nursery rhyme skill built into Everygen, a long prompt on how to make good nursery rhyme videos, which is why it could one-shot this. After generating the video it checked its own work against the Short I linked.

No revisions. I’m not a kid, but I think ours is better than the original.

Made in Blender without AI, these 3D story videos take days and cost hundreds to thousands of dollars. I asked for one that could be the channel’s next upload.

It worked for 17 minutes and 54 seconds on the idea, the script and the voiceover. The video analysis picked up that the first couple of seconds of these videos matter most, so it planned a visual hook and generated a beetle-and-mantis starting frame. GPT-6 chose Kling 3.0 on its own, a cheaper model that generates faster.

The first draft was decent, but not quite there.

It worked for another 9 minutes and 27 seconds and came back with the version above. The hook is really good. There are still things I’d change: with a complex video like this you probably won’t one-shot it, and I only did one revision pass. To actually upload it, I’d do two to four.

Shorts Studio in Everygen already has a Disney format that makes these in one prompt, and you can ask the ChatGPT plugin to use it. For this one I wanted nicer animation and something a bit more emotional.

It worked for about four minutes and pitched a story it called “The Last Spark”: in a snowstorm, a palm-sized robot spends its last warm light reviving a little bird, and the bird nudges it into the sunrise to recharge. It asked me to approve the story and the model, Seedance 2.5, before generating. Beyond that approval I gave it no other input.

Ranking compilations are one of the most popular styles right now. Shorts Studio’s ranking videos format makes the individual clips. I wanted to see if ChatGPT could edit the entire ranking video, text included, by itself.

It analyzed the reference’s pacing, captions and ranking format: seven items, a title that stays on screen, a numbered list on the left, hard cuts and no narration. Then it wrote a script, “Ranking trampoline fails that voided the warranty”, with a stronger opening joke, and asked me to approve it.

Once I approved it, it generated the clips in a couple of minutes and returned the edited video.

If I were posting it, I’d reorder some of the clips and regenerate a few. The point is that you can make almost any clip you can think of and revise it until it looks real.