A two-layer launch video: code motion over a Seedance plate

Level 1: motion graphics only. Find a launch video whose style you like. He browses X and a launch-video gallery he names (transcribed as 'watchships.com'; we haven't verified it). Give Claude the link and ask it to go through the video frame by frame and confirm it understands the scenes. Then give it your app's URL and ask for the same kind of launch video adapted to your product, telling it to invent any UI elements it's missing. Revisions are cheap because the result is code. He asked for generated profile photos, a two-step outro (feature name, then app name) and sound effects when chat bubbles appear.

Music is where he steps in himself. He doesn't let the model pick the music. He chooses a licensed track (from Epidemic Sound), drops it into the folder, and asks Claude to add it with smooth fades and arrange it to fit the animations. He also built a free 'Motion OS' player that turns on-screen edits into a prompt for Claude, which saves screenshot markups.

Level 2 adds a cinematic layer under the motion graphics, using four steps in strict order. (1) An AI actor and a location image. (2) The actor composited into the location. (3) All the other shots, with angles Claude reads from the reference video. (4) Only then, animate each shot. He says one-shotting everything gave obvious AI slop, while going step by step gave far better results.

Model choices, as he describes them (he notes Higgsfield sponsored the video). Have Claude write the long, detailed image prompts for you, camera type included. Use a character sheet (the face from several angles) built from a reference photo with Seedream 5 Pro through the Higgsfield MCP. Soul 2.0 is good for realistic portraits but takes no reference image. Use Soul Cinema for locations. Animate with Seedance 2.5, drafting at 480p first and finishing at 1080p.

His own verdict: shots with voice-over over b-roll held up, but on-camera talking still looks like AI.