Give it a photo and describe the motion you want. It reads the frame, finds what would trip the generator, and hands back a prompt CapCut is hard to misread.
AI video generators fail in predictable ways. Ask for three actions in five seconds and you get two, half-finished. Ask for words on a sign and you get shapes that look like letters. Ask a sunny photo to become night and the whole frame flickers while the model relights it. Most bad generations aren't bad luck: the prompt asked for something the model handles poorly.
CapCut on Steroids reads your photo first: framing, where the subject sits, light, contrast and color. Then it reads your request, splits it into subject, motion, camera, light and style, and checks it against 14 patterns that commonly break image-to-video. The prompt you get back describes what's already in the frame, asks for one clear motion and lists what to avoid.
The same method is available as an open-source Claude skill, which uses Claude's vision to describe your exact photo in the prompt.
Aspect ratio, resolution, brightness, contrast, palette, and the subject's size and position, so the prompt describes the photo instead of contradicting it.
14 checks, including action overload, style conflicts, text in frame, hand-heavy motion, competing camera moves and lighting the photo can't support.
Subject, identity, motion, camera, light, style and timing, in that order, followed by a negative prompt built from what the linter found.
Safe for minimal motion, Cinematic for camera work and color grade, Social for a strong first second.
Your prompt against the refined one, with every lost point explained by an issue you can read and fix.
The open-source version uses Claude's vision to name exactly what's in your photo, then applies the same rules.
Measure the frame: framing, subject position, light, contrast and palette.
Split the request into subject, actions, camera, light, style and pace.
Check it against 14 failure patterns and fix what the prompt can fix.
Write the prompt, negative prompt and three variants in the order generators follow.