🧩 Claude Opus 5 solved ARC-AGI-3: 100.00 RHAE and all 25 games with one skill
Poobesh Gowtham (@pbshgthm) published arc-skill — a skill for Claude Code that solved all 25 public ARC-AGI-3 games: 183/183 levels and 7645 actions versus a human median of 17135. Without the skill, the same model scores only 30.16%.
🌍 The 70-point gap was closed not by fine-tuning, but by a “predict before act” rule: before each move, the model predicts what will change on the grid, the harness rejects an action without a prediction, and checks it against the actual frame. The result has been verified by the ARC Prize Foundation.
👤 The pattern can be transferred to your own agents: recording a prediction before acting turns wandering into an experiment with an error log. The skill is installed with the command npx skills add pbshgthm/arc-skill, and a breakdown with replays is available at arc-skill.vercel.app.
Source 1: https://arcprize.org/scorecards/24ddb219-987e-464f-9050-6398a29cf5ac
Source 2: https://arc-skill.vercel.app/