What makes an AI launch video go viral: 74 films, measured.
In the first weeks of Claude Opus 5.5, people posted hundreds of motion-graphics and explainer films made with it. We read all 475 public prompts behind them, measured the 74 highlighted films, and studied the strongest frame by frame. The short answer: the best films barely cut. They keep one stage, keep it moving, and let a single object carry the story.
Key findings
- 30 sThe median length. 15 seconds is the second most common.
- ~2Hard cuts per 30 seconds (median), with the first cut at about 3.8 seconds. Everything else changes in place.
- <10%Time with nothing moving. 21 of the 74 films never hold still for even half a second.
- −14LUFS, the median loudness: the level feeds play at. 89% of the films have sound.
- 43%Rendered at 60 fps (51% at 30 fps).
- 84%Are 16:9. 8% are square and 7% vertical.
The numbers
| Measure | The viral films (median) | What it means for your video |
|---|---|---|
| Length | 30 s (15 s is the second peak) | Make a 30-second cut first; cut a 15 from it. |
| Aspect ratio | 84% 16:9 · 8% 1:1 · 7% 9:16 | Design for 16:9, keep the message in a centre column so a vertical cut works. |
| Frame rate | 51% 30 fps · 43% 60 fps | Render the final at 60 fps: motion design reads smoother. |
| Sound | 89% have audio | Score it. Silent product videos are the exception. |
| Loudness | −14 LUFS | Master to −14 LUFS so it isn't quieter than the feed around it. |
| Hard cuts | About 2 per 30 s; the first at ~3.8 s | Let the hook build for 2 seconds or more. Use 2–4 cuts, on the beat. |
| Still time | Under 10% | No hold longer than about a second, except the end card. |
Ten habits the best films share
- A thread. One object travels through the whole film and carries its continuity: a dot that is the full stop of the headline, then a pixel, then the button; one line that links every scene. It replaces most cuts.
- One stage, many states. A card or window on a quiet background changes size, content and colour while the camera reframes, instead of cutting to a new shot.
- Type that builds. Headlines add a line per beat, swap one word in place, cross out a claim and correct it, or roll a number like a counter. The words are the motion.
- Real numbers, big. A price, a count, a distance, set large with a small label for what it counts. 173 of the 475 prompts asked for data or charts.
- Persistent chrome. A thin chapter bar or timecode that never leaves the frame, so the viewer always knows where they are.
- Designer notes. Hand-written annotations with arrows and small monospaced spec labels make the craft visible.
- A quiet, warm background. Mostly warm off-white with one accent colour; developer tools go near-black with one glow. Never busy.
- Real UI, rebuilt. Product screens appear as crisp rebuilt components driven by a visible cursor, not as blurry screenshots.
- Personality in explainers. A character, a callback, honest numbers with their uncertainty, captioned line by line.
- An ending that lands. Logo, one line, a button the cursor clicks and the URL, held for two seconds or more.
What the strongest prompts had in common
Most prompts were short: the typical one is 31 words, and about a third simply ask the model to "go all out", leaving the taste to the model. The prompts behind the most-shared films were longer, not because they asked for more effects but because they locked decisions:
- One central visual rule for the whole film ("one shape, never cut"; "the camera only flies forward").
- The message, word for word, with no extra slogans.
- A strict palette of three or four colours and a type system of two or three roles.
- A storyboard with timestamps and a tempo, so something happens on every beat.
- A banned list: bouncy easing, particle bursts, glows, template looks, dead time.
- A loop of build, render, inspect and refine before anything is delivered.
A checklist for your own launch video
- Write the message first, word for word, in under 15 words.
- Pick one object that will travel through every scene.
- Plan five to seven moments but only two to four hard cuts.
- Put one real number on screen, big.
- Keep something moving at all times; give the end card two seconds.
- Score it to the cuts, master to −14 LUFS, render at 60 fps.
Method
Source: the public awesome-opus5-5-videos collection (MIT licence), which lists 475 creator prompts, including 288 motion-graphics and 62 explainer films. We downloaded the 74 highlighted motion and explainer films and measured each with ffmpeg: hard cuts with scene detection (threshold 0.3), still time with freeze detection (0.5 s minimum), loudness with EBU R128, plus length, aspect ratio, frame rate and audio. We then read frame strips of the 15 strongest films. Figures are medians unless stated. The films belong to their creators: we kept only measurements, deleted the downloads after analysis, and reuse none of their work.
You can measure your own video the same way: TeaserKit includes this check (rhythm.mjs) and runs it on every draft before the final render.
TeaserKit follows all of this for you.
Paste your website link and TeaserKit directs a launch video to these numbers: a thread through the film, two to four cuts, an original score mastered to −14 LUFS, and a 60 fps render. It runs on the ChatGPT or Claude you already pay for, for a one-time price.