# Shot grammar A giant-product stunt is two shots and one idea: the product is somewhere it cannot be, and the street takes it calmly. The look is a phone clip, because that is how the format is shot and shared; the tag, the caption and the platform label say what it is. ## The two shots | Shot | Length | Camera | What the viewer learns | |---|---|---|---| | 1. Wide establishing | 5 s | Pavement or promenade, wide phone lens, slightly tilted up or level, a slow push-in | Where we are, how big it is (tiny pedestrians, buildings for scale), and that the stunt has started | | 2. Low-angle hero | 5 s | Phone near the ground, looking steeply up, wide-lens perspective, a slow tilt up that holds | The product towering over us, close, in the same light and street | Hard cut between them. 8 to 10 s in all (the sample run: a 4.9 s wide shot and a 3.1 s hero shot, 8.0 s). Shorter is better than a hero moment that drifts. ## Keep the product locked - The keyframe is made from the user's photo and approved against it before any video is paid for. Shape, proportions, parts and colours must match. - Both keyframes already show the stunt's key moment (the product half out of the billboard; the product towering). Poised frames work; entries invented by the video model do not. - The hero shot is made from the approved wide frame plus the photo, so the street, the building, the light and the product agree. - The motion prompt says the product "does not change size" and "keeps exactly the shape, colours and parts it has in the first frame". Kling keeps shapes better than Veo, so Kling is the default. ## Scale cues The product is only giant if the world says so: - Tiny pedestrians (a few pixels tall in the wide shot, only heads and backs in the hero shot), cars, lamp posts, windows and floors of the buildings. - True contact: a shadow on the street or the building, a reflection in windows, light matching the sky. - Wide-lens perspective in the hero shot: the nearest edge large, the far end small. ## People - Tiny, far away, from behind, or only legs at the frame edge. No face is readable. This is for taste and for Seedance, which refuses a start frame with a clearly visible realistic face (the 480p drafts use it). - Calm and curious: looking up, some filming with phones. Nobody runs, screams, falls or is in the product's path. - Never a real person, never from a photo of a real person. ## Phone-clip behaviour Handheld sway, a slow push or tilt, a normal wide phone lens, slight noise, true-to-life colour. No cinematic grade, no lens flare, no slow motion. No news-broadcast or security-camera look, no on-screen clock, no "LIVE" or "BREAKING" caption. ## Safe zones (9:16) - Top 12% and bottom 20%: nothing essential (the app's top bar and caption area). - Right edge: keep the product's key shape away from it (the like and share buttons). - The AI-made tag sits top-left, 5% in and 13.5% down, below the top bar. `--tag-pos top-right` moves it. - The product and the action live in the middle 60% of the frame. ## Sound Street sound from the video model, no music, no speech. - Shot 1: city (or tram, or harbour) ambience, a soft crowd murmur with no words, then the product's own sound: a low smooth rumble, a soft whoosh, a deep hum. - Shot 2: the same ambience, the product's weight settling, wind across the microphone. - `assemble-stunt.mjs` joins the sound with a 0.08 s crossfade and fades out over the last 0.4 s. If there is music, a voice or a siren, re-roll the shot or deliver muted. If the user wants music, they add it in the app from its licensed library. ## Trims The first 0.1 s of a clip can be a settling frame; skip it. Cut before anything odd: the product changing size, the billboard cracking, a pedestrian growing, a second product. Write `in` and `out` per shot in stunt.json. If shot 2 will not hold, deliver shot 1 plus a shorter shot 2 (`--only 1,2` with an earlier `out`), or shot 1 alone, and say the clip is shorter.