Arc Agi 3 Prime Agent — Development skill for Claude Code

Prime Agent on ARC-AGI-3

[Prime Agent](https://github.com/PrimeIntellect-ai/prime-agent) playing the 25 public [ARC-AGI-3](https://three.arcprize.org/) games. Each game gets its own agent session and its own sandbox; the agent reaches the game only through a local broker socket, so it never imports the ARC SDK and never sees a provider credential.

**Scorecard:** https://arcprize.org/scorecards/2af780b4-f2a1-43e9-a794-b23da3cd3f9f

Results

Three independent runs with Claude Opus 5 at `xhigh`. `run-1` is the median by RHAE and is the run reported above.

run	RHAE	levels	games won	est. cost
run-1 (median)	95.24%	178/183	24/25	$1,059
run-2	94.99%	183/183	25/25	$1,288
run-3	95.5%	179/183	24/25	$944

Per-game numbers are in `results/results.json`. Open `results/index.html` in a browser for a sortable view (no build step, no dependencies).

Method

The agent is unmodified Prime Agent. What makes the run is the prompt plus the protocol:

• AGENTS.md — the behavioral guidance every game session follows. Game-agnostic: it never names a game, and it forbids inspecting engine source, other games, or prior runs. Frames are to be analyzed programmatically, not transcribed by eye.
• game-prompt.txt — the initial prompt.
• broker_client.py — the only interface the agent gets: observe(), status(), act().
• game_protocol_guard.py — validates action shape before anything reaches the engine (ACTION6 requires integer x,y in 0..63, and so on).
• arc3_local_broker.py — owns arc_agi.Arcade(OperationMode.OFFLINE) over a local copy of the games and speaks newline-delimited JSON over a Unix socket.

Protocol, matching the reported runs: 500 actions per game, at most 20 actions per agent call, a fresh run per game, and a game-over consuming one counted `RESET` before the agent re-observes.

Requirements

• Python >= 3.12 and [uv](https://docs.ast