How AI is used
Which parts are written by a model, which are not, and what that means for you.
Every factual statement here is drawn from what the software actually does, and the text has been reviewed. It is still a description of a service rather than legal advice.
What the model writes
For an explainer video: a model reads your brief and settles what the video is about; a second writes the script — the scenes, what the voice says in each and what appears on screen; and one model per scene designs that scene and its animation.
For a how-to video: a model decides which pages to look at while the website is explored, then writes the steps and what the voice says during each. If a step cannot find what it should click while recording, a model picks the matching element from what is on the page.
The voice
Every video is narrated by a synthetic voice, not a person. A speech provider (ElevenLabs) reads the script aloud in the voice the video’s creator chose. The captions show exactly what is said, timed from the recording itself.
What the model does not write
The timing and the assembly. How long each scene lasts, when the voice starts, where each caption appears and how the scenes and the voice are joined into one file are worked out by code from the recorded voice, identically for every video. That is why the picture and the voice stay in step: it is not something a model is asked to get right.
What is checked automatically
Every designed scene is opened in a real browser before it is rendered: that it loads without errors and that its text stays on screen. Stills of every scene are reviewed by a vision model for visible defects, and scenes it flags are redesigned once. For a how-to video, every step is performed on the real website, and any step that could not be performed is reported in the chat.
These checks catch technical and visual failures. They do not, and cannot, judge whether the explanation is any good — that judgement is yours.
Review before use
A generated video should be watched before it reaches anyone. Models produce confident, fluent narration that can still be factually wrong, and a video carries more authority with a viewer than a chat reply does.