Project story / Marketing Engine · 19 Jul 2026
Trying to make AI app videos without the AI slop
Building an app and explaining why it exists are two different jobs. I enjoy the first one far more. Once LightScout and Rest+Rise became real products, though, I could not keep treating the second as something I would eventually get around to.
This project is my attempt to find a repeatable way to make short videos about the apps. The obvious attraction is generative AI: it can turn an idea into footage quickly and remove much of the production overhead. The obvious problem is that it can also produce exactly the kind of glossy, synthetic filler that now gets labelled AI slop.
The constraint is truth
A marketing video still needs to say something specific. If the same clip could advertise a meditation app, a bank and a protein powder, it is not helping anyone understand the product. The experiment is therefore less about generating more footage and more about finding a process that keeps the product idea intact.
For LightScout, that means showing the moment a traveller actually has the problem: arriving somewhere unfamiliar, looking for a composition and trying to work out where the light will fall. It does not mean asking a video model to recreate the app interface. The result would be inaccurate, would age immediately and would make a real product look fake. The app belongs in the narration, captions and final call to action; the visual story should stay grounded in what the photographer is doing.
Rest+Rise needs a different visual language. It is about the transition into and out of a day: putting the phone down, preparing a room, moving, breathing or writing. Again, the useful material is the recognisable human behaviour, not a perfect-looking person performing a vague wellness montage.
A small content system, not a content machine
I am experimenting with a workflow that separates the durable parts from the disposable ones. The durable parts are the audience, the product truth, the opening idea and the reason to care. The disposable parts are individual scenes, cuts, captions and formats. Those can be regenerated and tested without changing what the video is actually trying to say.
That distinction matters because automation creates a temptation to optimise for volume. I am not trying to fill a calendar with interchangeable posts. I would rather make fewer pieces that demonstrate a real use case, have a point of view and look as though someone made a choice.
The test is simple: if I removed the app name, would the idea still clearly belong to this product?
What I am learning
AI is already useful for exploring directions, creating shots that would be impractical for a one-person project and rapidly making variants. It is much less useful at deciding what is worth saying. Without a clear brief it averages everything out, both visually and verbally.
The work is still experimental. There is no grand growth result to report and the project folder is deliberately not pretending to be a finished platform. What I am building is a better judgement loop: start from a real product problem, generate only what supports it, edit aggressively and keep the parts that feel specific.
If it works, the outcome will not be an autonomous marketing engine. It will be a practical way for me to show the products I am already building without the marketing becoming another full-time project — and without adding more synthetic noise to everyone else's feed.