Experiment · Interaction design
Two browser-based characters one person can perform live on TikTok.
The point
Let one person run an expressive character on a livestream without character rigging, streaming automation or heavy programming.
AI video
This experiment uses AI-generated video clips instead of programmed animation: idle loops, triggered actions, automatic transitions and costume changes, with seamless switching and effects. More detailed performance, same simple controls.
Pixel art
A game-style scene with a fortune teller, his owl, crystal-ball readings and dialogue. I defined the interaction model, character behaviours, visual direction and controls, and worked through the limits of code-generated pixel art.
Design decisions
A person stays in control.
The operator triggers every reaction instead of handing it to automated audience triggers.
Controls and stage are separate.
A private console for the performer, a clean window for the broadcast.
Work within the constraints.
Browser limits, platform overlays, consistent media and the cost of AI generation all shaped the build.
Refine from working builds.
Each version was tested live, and the problems found decided the next change.
Don’t rebuild what exists.
Broadcasting stays in established streaming tools. The apps only do the performance.
Outcome
Two working browser apps, and two different answers to the same question: AI-generated video or programmed animation.
Claude wrote the code and handled media processing. I set the requirements, made the visual assets, tested every build and decided what to change.