An agent skill for interactively playing ARC-AGI-3 benchmark games. The AI agent reads rendered grid images, reasons about game mechanics, and discovers rules through trial-and-error — just like a human player would.
Recommended agents: Claude Code, Codex CLI, and Kimi CLI — their native models support vision, which is required for reading rendered grid images.
-
Install the skill:
npx skills add https://github.com/ParadoxZW/ARC.skill
-
Install Python dependencies:
pip install arc_agi arcengine numpy Pillow
-
In your agent, say: "play ARC-AGI-3 game wa30"
No API key needed — anonymous access provides ~25 games.
The game API holds a stateful connection that can't be shared across processes. A background runner maintains this connection and communicates with the agent through files — rendered grid images, a JSON status file, and an action log. The agent observes, reasons, and acts one step at a time.
As the agent plays, it builds per-game knowledge files that persist across sessions, enabling it to learn from previous attempts at the same game.
| Path | Description |
|---|---|
SKILL.md |
Skill instructions, gameplay strategy, and knowledge system protocol |
scripts/ |
Helper scripts (play.sh, act.sh, game_runner.py, etc.) |
prompts/ |
Sub-agent prompts for knowledge curation and per-platform instructions |
knowledge/ |
Dynamic per-game knowledge accumulated during play (gitignored) |
Apache 2.0
