I built this as my entry for the CapCut Seedance 2.5 Video Challenge because I wanted to test how far a single continuous shot could push a multi-world action narrative — something Seedance 2.5's native 30-second generation and multi-reference control finally make possible without stitching separate clips together.

My approach was to treat camera continuity as the hardest constraint on purpose. Instead of writing a normal shot list, I built a "camera continuity master directive" into the prompt: the camera has positional memory across every beat, so wherever it ends one scene (left side, right side, behind the character), the next scene has to pick up from that exact position and keep arcing from there. That's what let me move the character through three completely different environments — a collapsing jungle ruin, a space dogfight, and a hellish cavern — without ever faking a cut.

I iterated the prompt in stages rather than trying to get everything right in one shot. Early drafts had issues that only showed up after generating: the camera would go static during transitions, creatures would spawn illogically out of portals instead of existing in the environment already, and timing between action beats felt off (like a character walking before being grabbed instead of being grabbed instantly). Each of those got fixed by making the prompt language more explicit about sequencing and physical logic, not by relying on the model to infer intent.

I also used seven image references (character sheet, three environments, two vehicles/creatures, and the final reveal room) specifically to keep visual consistency — same character, same pod orientation, same environment geometry — across a story that jumps genres and settings within one unbroken take. The ending twist, revealing the whole chase was a video game session, was designed as the payoff for committing to the one-take format: the final camera pull-back-and-rotate literally reframes everything the viewer just watched.