How to Make AI Video Ads That Aren't Slop

There is a lot of AI slop on the feed at the moment, and most of it fails for the same structural reason. The product shows up in the first three seconds, so the viewer is asked to care about something before they have been given any reason to.

The films that work do the opposite. They spend the first half teaching you something you did not know, and the product turns up at the end as the answer. That is not a style choice, it is the whole mechanism, and it is what unlocks a new pocket of audience: people who would scroll past an ad will happily watch a two minute explainer about where something came from.

Build it in this order

Order of operations matters more than any individual prompt. Ours runs like this, and steps two to four are the ones everybody skips.

1. Research the spine. Three or more facts you have actually verified, not recalled. Real history reads as a film. Invented beats read as an ad.
2. Write the voice-over. Before any picture exists.
3. Generate it and measure every line in seconds.
4. Derive shot durations from the audio. Lead in, plus the lines, plus the gaps, plus a tail. Never pick a shot length before you know what it has to carry.
5. Write one prompt per shot, restating the style every time. The model has no memory between shots.
6. Render one shot and look at it before committing to the set.
7. Assemble the silent master, then lay the voice-over over the top.

The dead giveaway of a film built the wrong way round is thirty seconds of silence in the middle where the narration ran out and the pictures kept going.

One shot is a scene, not a picture

A shot is up to fifteen seconds of continuous animation carrying three timecoded beats, and something has to happen in each of them. Something enters, tears away, fills, collapses, is revealed. Hard cuts inside a single shot are allowed and good.

The failure mode is a sparse diagram: one icon centred on a flat background. It reads as a slide, not a film. Scenes need people, places and depth, and light that comes from somewhere inside the scene, a candle or a desk lamp or floodlights.

The other thing that kills a film is sameness. Before writing a single prompt, fill in a diversity matrix: for every shot, declare location, scale, light source, time of day and dominant motion, and do not let two adjacent shots share more than one of those. That single constraint is what stops six variations of the same frame.

Never bake the voice-over into the render

Some models will happily generate narration inside the clip. Do not let them. End every prompt with an explicit ban on speech, and ask for diegetic sound design instead: the paper tears, the fire crackles, the crowd roars.

What you get is a silent but fully scored master, which is an asset rather than a one off. You can re-voice it, re-cut it, run a completely different script over the same pictures, or hand it to an editor. Change the words later and it costs one text to speech call instead of a whole render.

What the models still get wrong

Short words in capitals set reliably. Date stamps, single keywords, that is the ceiling. Anything longer garbles. A full frame hero shot of your packaging will come back repainted with misspelled label copy, so keep the product mid-size with a reference image attached, and typeset any end card yourself afterwards.

Watch brand names that are not spelled the way they sound, too. Ask a model for an end card and it will happily typeset the pronunciation instead of the spelling, which is how you ship a dead URL.

The arc

Unexpected origin, then the strange detail nobody knows, then the mechanism, then why it matters now, then the product. Last.

Write the film you would actually watch. The ad is what happens once you have earned the right to mention the product.

If you want us to run this for your brand instead of running it yourself, book a free creative audit.