AI-powered chart pattern detection for maimai — all planned to be built through agentic coding with Claude Code.
Being built through agentic coding with Claude Code. The AI researches, codes, iterates, and ships — while the human validates by reviewing gameplay captures and testing on a simulator.
maimai is a SEGA arcade rhythm game with a circular screen surrounded by 8 touch-sensitive buttons. Players tap, hold, and slide along the screen in time with music. Unlike lane-based rhythm games, maimai's 360-degree layout enables unique spatial patterns — including techniques where both hands must operate independently, creating some of the hardest coordination challenges in any rhythm game.
This project applies AI to understand and detect these patterns.
Chart Pattern Detector — 9 patterns
Structural detectors for maimai chart patterns: Umiyuri (海底譚), 拍滑, slide reading, 轉圈/掃鍵, 縦連, トリル, 乱打/散打. No ML — pure structural rules derived from understanding game mechanics. Includes a NiceGUI web dashboard with multi-timeline visualization and per-pattern leaderboards.
See detector/ for the full story.
The detectors emerged through agentic test-driven development over 13+ hours:
-
Research — Started from zero maimai knowledge. Parallel web research agents covered mechanics, simai format, community terminology (JP/CN/EN), and existing AI work.
-
Umiyuri deep-dive — The Umiyuri detector went through 20+ iterations of the AI proposing rules → verifying against the DB → human reviewing gameplay captures → reporting false positives → AI fixing. This produced the positional chain rule and the understanding of slide timing mechanics that unlocked all subsequent detectors.
-
Rapid expansion — With the foundational understanding of slide timing (1-beat delay, star position mechanics, hand independence), the 拍滑, slide reading, and all tap-based detectors were built quickly. The 拍滑 detector worked on the first try. The 乱打 detector matched 10/10 community-cited songs immediately.
-
Community validation — Every detector was cross-referenced against community sources (JP: Gamerch wiki, kioblog; CN: Bahamut, Bilibili; EN: Reddit). Rotation detection matched 14/14 community-cited songs.
- 一筆畫 detector — one-stroke connected slide patterns
- 魔法陣 detector — center-crossing rotational slides
- Audio-to-chart pipeline — beat detection → pattern slotting → simai output
- Player recommendations — "play X to improve your Umiyuri" / "play Y for easy rating"
- Chart generation — data-driven, learning from the corpus
MIT