One clip instead of forty prompts. Proxy watches gameplay of existing games, measures how the camera and the player move, reads the interactions, and writes the brief your AI coding tool builds from.
Building a game with an AI means explaining feel in words: how fast the player moves, how the camera turns, what happens when you hit something. People burn forty prompts on it and still get something off. A clip says it in ten seconds.
Proxy takes the clip. A vision model reads the frames, a separate measurement tracks the motion on the pixels, and the result comes back as a brief a coding AI can build from: evidence with timestamps, mechanics with hypotheses, and an honest list of what the footage cannot tell you.
Four clips of existing games. Turn the layers on: the motion field is measured on the video pixels, the regions come from the vision model at the frames it read, and the blueprint on the right is what it returned. Click any evidence line to jump to that moment.
We gave one generated brief, and nothing else, to an AI coding model in a fresh context. One prompt. What it built is below, unedited. Play it, then compare it with the clip it came from.
Frames
The clip is cut into ordered frames with their timestamps, up to 12 frames and 30 seconds.
ffmpeg / canvasVision
A vision model reads them in order: actions, camera versus character motion, interactions, state changes.
engine/analysis.mjsMotion
Separately, block matching on 10 fps frames measures camera pan, forward motion and energy. No model involved.
tools/flow.mjsBlueprint
Evidence ledger, mechanics with hypotheses, unknowns, and a build prompt for your engine.
technical-brief.jsPick a gameplay clip of 30 seconds or less. Eight frames are cut in your browser and sent to the analysis server running on this machine. Nothing is stored.
The footage and the analysis
Four real clips of existing games. Each report came from a real call to a vision model on frames of that clip and is saved as returned. The motion field is computed from the video pixels.
The brief
Assembled from the evidence ledger. Every mechanic carries what was observed, a hypothesis to validate and an implementation suggestion. Unobserved behaviour is labelled as a design choice.
Source code or exact numbers
A clip shows what happened, not the rules behind it. Speeds, damage, hitboxes and timings stay in the unknowns list. The model can misread a frame, which is why every claim cites a timestamp.
Sources: OpenArena and Xonotic gameplay by Drummyfish (GPLv3) and MineClone2 0.84 by VoxeLibre (CC BY 3.0), all via Wikimedia Commons. The CS2 clip is third-party footage of Counter-Strike 2 (Valve), shown only to demonstrate the analysis.
The coin of the Proxy project. Nothing on this page promises returns or revenue, and any utility is documented here before it ships.