Two steps, one generation each, and the product is a free variable. Step 1 takes the visitor's product photo (@product) and renders one eight-panel storyboard sheet on Seedream, one panel per scene, so the shot list and the product lock are checked for 19 tokens before the video is paid for. Step 2 is one MiniMax H3 15s 9:16 render at 2K in reference mode with the product photo and the storyboard attached. Every prompt describes the product only as 'the exact same product from the reference image' and says 'light accents in the product's own dominant colour', so the same text works for a watch, a bottle, a chocolate bar or a sneaker. Hands: at most two hands of one person, one calm pick-up and one touch, never a fastening, never a third hand. The beats that depend on the product type carry fallbacks inline (display waking, glass or liquid catching light, a lens flaring, or a light sweep across the logo). Rules that held: never ask for on-screen text, keep any screen or label identical everywhere, and if a Seedream job sits at processing with no progress for 10 minutes, re-fire it instead of waiting.