Client
Self-Initiated Project
Role
Creative Director
Year
2026
Tools
Claude Code, Higgsfield (Soul Cinema, Nano Banana Pro, Seedance 2.5), HyperFrames, CapCut
I wanted to know if a character invented in text could maintain its identity across four different models, from the first face to a fourteen-second video that closes with my logomark. The challenge wasn't just narrative: I wanted to create a consistent, end-to-end reel in one hour of work. I put it together in a single session with Claude Code, using skills that each handle a part of the process and with Higgsfield's CLI for generation.
The Character, before the scene
A character-building skill first wrote a text brief: features, skin, hair, and a thin nose ring. With that brief, I did a test run in Soul Cinema and then set the definitive face in Nano Banana Pro, on a neutral gray background with flat lighting, so that no lighting would stick to the character. On that face, I added the costume and finished with a reference sheet in three views: front, back, and face. Reviewing each image before moving forward was part of the job. The Soul Cinema test brought a green halo around the silhouette. The costume slipped a slight shadow under the boots, even though the prompt explicitly forbade it. One of the attempts for the reference sheet directly returned another face, so I discarded it and continued with the one that maintained the identity.
A location without people
In parallel, a location plate skill described the pool: municipal, drained, at sunset, with a cold spotlight on the diving platform. I generated it empty because a location without a character works better as an anchor for the video.
The Scene
A video direction skill combined the two references into a prompt for Seedance 2.5. She enters the pool, looks at the platform, turns to the camera, and says "Am I real?... I don't know anymore." Eight seconds in a single shot, with the exact line in Spanish.
The Extended Scene
To extend the scene, I wanted her to grab the camera and, as the frame lowered, the dry floor to appear covered in water, with her full reflection. Achieving that shot took a couple of iterations: each round adjusted the prompt a little closer to the original idea, until the framing, the water, and the reflection matched what I had in mind.
The Decision
For the ending, I chose for the ripple to be generated by the model and for me to place the logomark in composition. The ripple can vary from one shot to another; a brand mark needs to be exact. I assembled the ending itself in HyperFrames, which creates video from HTML. The outline logomark enters just as the ripple is born in the water, grows from 10% to 100%, and remains alone when the background fades to black. I fine-tuned the timings manually in its visual editor.
The Final Edit
With both parts resolved, the dialogue scene and the ending with the water and the logomark, I assembled the complete piece in CapCut: I joined the two clips, adjusted the color so they responded as a single shot, and added a background audio track.
Conclusion
Each model did what it saw in its reference. Choosing what to show it was the directorial work.
Up next
Next project