Short videos from trends,
with an AI critic and a human moderator in Telegram
A five-stage pipeline (idea, script, storyboard, video, publish) that copies the structure of trending short videos and swaps in the brand's product. Each stage passes an AI critic and a human moderator in Telegram. A redesign cut model calls per video from 66 to about 3. Trend pool of 157, labelled dataset for ranking, scene-to-scene visual continuity.
What it does
A bot collects trending short videos by niche hashtags. A single Claude call in “faithful copy” mode extracts the structure of a trend (hook, cuts, visuals, sound) and rewrites it with the product in place of the original object. Storyboards are generated, frames are rendered scene by scene with each new frame referenced from the previous one for continuity, and the result goes to a human moderator in Telegram with the critic’s notes.
What we learned
Early versions made 66 model calls per video. Collapsing the stages into one faithful-copy call brought it to about 3 with better results. Video engines were swapped as quality and billing changed (Kling, Veo, then Sora image-to-video). Rules added after human review: the product must be visible, no invented text on packaging, scenes of 8 seconds extended to 15 to 16 in production. Every trend is labelled good or bad with a reason, building the dataset for a ranking model.