This paper presents architecture and design of a Twitch-based audience participation fighting game, where the strengths of game AI characters are dynamically adjusted based on audience cheering and jeering. Limitations underlying existing audience participation games (APGs) are discussed, and followed by solutions as demonstrated in our game. Our study includes AI tests on 500000 one-minute game rounds and a user experiment with 82 participants. We investigated user experience (UX), comparing two different kinds of gameplays (collaborative vs competitive), as well as two different kinds of interaction (cheering vs jeering). A pairwise preference, four-alternative forced choice (4-AFC), version of the Game User Experience Satisfaction Scale (GUESS) is used for UX assessment. At the end, we provided guidelines and hypotheses for future APG research.