Explored three prototype directions in Google AI Studio to find where the quarter question lives. The goal was to build fast and learn what the question actually means when turned into a real interactive thing.
Tool: Google AI Studio / Gemini · MediaPipe Hands · MediaPipe Tasks-Vision
Generated all three prototypes via vibe coding prompts. The Dot Destroyer went through 5 iterations in AI Studio — single hand → two hands → color matching → portal mechanic → bilateral coordination framing. Each prompt pushed the mechanic closer to an assessment rather than a game.
Key technical decision made in the Reaction Time prototype: switched from Gemini API cloud analysis (which required a round-trip to the server) to MediaPipe Vision local inference. This cut latency from ~200ms to sub-20ms — necessary for measuring human reaction time meaningfully.
Switching from Gemini cloud to MediaPipe local inference wasn't just a speed improvement — it changed what the test could measure. With 200ms server round-trip, you're measuring "prompt-to-recognition" time, not reaction time. Local inference at sub-20ms makes the camera-based reaction test scientifically meaningful. The tool choice determined the measurement quality.
I discovered the quarter question by building toward it, not by planning first. VibeWord Animator was the fastest thing to build and the quickest thing to reject — it felt like AI aesthetic output, not human function assessment.
The Dot Destroyer's evolution into bilateral coordination showed that adding cognitive load on top of physical input (match the right hand color to the right dot) is what moves a game toward an assessment. Dexterity alone isn't interesting to the quarter question; coordination between brain hemispheres is.
Unresolved: In the Reaction Time test, who defines what "correct" looks like — me or the model? Currently the model decides (55% confidence threshold). That feels like something to revisit.
Two of the three prototypes directly addressed the quarter question. Dot Destroyer V2 tested bilateral coordination — a real cognitive-motor function. The Reaction Time Test measured response latency using a camera — a classic clinical measurement done at home, on a laptop.
The key discovery: the quarter question is answerable. The technology (MediaPipe, Web Audio API, touch events) is already capable of capturing meaningful human function data from everyday devices. The design question is how to turn that capability into a useful tool.
Analyze the prototype code in depth — understand what every tool and library actually does. Decide which prototype direction to carry into M2. Research practitioners and projects in the cognitive assessment space to understand what already exists and what gap Baseline can fill.