A trailer is the one thing an AI video pipeline is genuinely good at: narration-led, cut fast, built from striking disconnected shots. Here is how to make one, and what it can and cannot do.
A single render here is capped at 180 seconds. That is not a limitation to work around — it is the natural length of a trailer, and every constraint the pipeline has happens to be a trailer’s native grammar: one narrator, no dialogue, quick cuts, and shots that do not need to connect to each other.
A 90-second narrated cut with real shot language is a trailer. Calling it anything bigger only sets up disappointment.
A film is written scene by scene, with questions asked before a word is committed. That is a conversation, not a web form with a dropdown. So the app connects to Claude as an MCP connector — one URL, added once, no key to paste:
https://automatedvideoapp.com/mcpAdd it under Settings → Connectors → Add custom connector, sign in when prompted, and Claude gains ten tools: fetch the production guide, see what each option costs, create a video, check its progress, list its scenes, drop a scene, set the end card, re-cut, list your uploaded images, and join finished videos into one longer film.
From then on you ask in plain sentences. Claude reads the production method first, asks what it needs, writes the script to the right length, and renders it — then edits it with you afterwards, for free, because re-cutting reuses clips that already exist.
That is the difference from a trailer generator you fill in: the thing that plans the film is the same thing you are talking to, and it can go back and change one scene without rebuilding the rest. More on automating video with Claude →
Connect the app to Claude, then say what you want:
Claude fetches the production guide first, then asks five short questions — runtime, shape, look, narrator, and whether you want generated imagery or free stock footage. That last one is the only choice that costs anything.
Scene length follows the measured narration, not the number you asked for. Roughly two words per second: a 60-second trailer is about 120 words across 10 scenes. Overwrite the budget and a 60-second request comes back at 97 seconds. Write to it and it lands within a few seconds of target.
For anything that exists in the real world, real footage looks better than generated imagery, not worse — real light, real texture, and no drift between shots. It is also free and unlimited, where AI stills spend your own Pollen.
Choose AI when the film needs a world that does not exist. Choose stock for anything else.
Every scene’s picture is generated independently, so a character will not look the same twice. This is true of every tool on the market, not just this one. The fix is framing:
Two characters in one frame blurs both identities. Write two scenes instead — one speaks, the other reacts. That is how films stage a confrontation anyway.
Ten unrelated clips look like ten unrelated clips. Turning on the cinematic finish applies one colour grade across every scene and a short dip at each cut, so a film assembled from stock footage, AI stills and generated motion reads as one piece. It costs nothing and adds no render time, and it is off for ordinary videos.
If one shot missed, drop it. Removing a scene re-cuts from the clips already rendered — it spends nothing and takes seconds. Removed scenes are kept, so the edit is reversible.
Connect Claude and ask for a trailerNext: the cinematic craft behind it, a story video for children, or go past three minutes with chapters.