ARC-AGI-3 is a skill issue

25 games from the ARC Prize Foundation. Each one is a grid of coloured cells with seven buttons under it, and nothing else.

Nothing says what the colours mean, what a button does, what winning looks like, or how many levels there are. The meanings change between games. A colour that is a wall in one game is a cursor in the next.

You learn by acting and watching what changes. Every act costs one action, and the score is how few you spend.

arc-skill is a short instruction file and a command line tool. The file adds one rule and the tool enforces it. Around that rule Claude gets five kinds of instrument, and none of them knows anything about any game.

Before Claude may act, it writes down what the act will do to the grid. Which cells change, and to what. The tool refuses an act that arrives without one, then grades the prediction against the grid that comes back.

Level 9 of tu93 mixes two static guards, three patrollers and a chaser. Claude built the map, rolled the patrol forward 250 ticks, and proved there was no safe path. Every node of the one corridor was a cut vertex. On the board as it understood it, the level could not be finished.

So it spent an action on the one thing it had never tested. It walked into a patroller, and said in advance that the patroller would still be there afterwards.

Six of the seven held. The seventh said the patroller was still at 40,17. The grid came back with that cell empty. Walking onto a patroller kills it, the same way it works for the static guards, and the level Claude had just proved impossible was now the intended solution.

What it wrote in its notes, and then played the rest of the level on:

A prediction joins these with semicolons, and each part is graded on its own, so one wrong part is a miss.

8 forms in all. The grammar is deliberately small: every form describes something the next frame can contradict.

A long game does not fit in a context window. Claude Code compacts its own context when the window fills, and it did that 115 times over this campaign.

So each game keeps one short page: what is verified, what is still assumed, what to test next. Claude rewrites it as it plays. It is the only thing a compaction cannot take away.

Two rules pull against each other and both hold at once. That pairing is the whole idea.

The gate is hard. An act without a falsifiable claim never reaches the game. Not a suggestion in a prompt. The tool refuses it.

The thinking is free. Claude uses whatever representation is right for the game in front of it, and moves up a rung when the one it has stops answering.

Every strong system on this benchmark makes a bet about representation. Tycho writes an executable world model and decides when to build, repair, use or bypass it. VISTA renders each frame and reasons over the picture in language, without program synthesis. Retrodict replays each hypothesis over its own logged history in Python before it acts.

Here no representation is required. The same unedited arc-skill produced a run that never wrote a line of Python and a run that built a physics simulator.

Claude reached for Python 1,727 times across 24 of the 25 games. None of it costs an action, because it runs on grids already recorded.

The instruction file says to batch a mechanic once it is verified, and never to batch exploration. It does not say when a mechanic counts as verified. Claude drew that line for itself, and the two modes came out far apart. A single test act missed 37.1% of the time. A planned sequence missed 2.9%, about 13 times fewer, and Claude put 92% of its acts into plans.

The guessing sits at the start. On the first level of a game 29% of acts were single tests, and after that the rate falls away and stays down.

RHAE is the benchmark's own score. It compares your actions against a median human player, level by level, and it stops at 100.00.

Everything else here is a system built for ARC-AGI-3 (except the last two rows, which are people, not systems). The second line of each row is how that system plays, taken from its own write-up.

Every name is a link to where its numbers can be read back. Six go to an official scorecard on arcprize.org. VISTA and Schema have no scorecard, so those two link to the write-ups their authors published, and both of those scores are self-reported.

arc-skill is a 129 line instruction file and a command line tool. Neither one mentions any of these 25 games, and the agent that runs them is Claude Code as shipped.

The score caps at 100.00, so the action count is the part of a run the score stops showing. Tycho spends the fewest.

All 25 runs together cost $728.17 in tokens at list price, for 7,645 paid actions.

Pick a board and step through it. The boards are decoded from the grids that run recorded, and the log is the one the tool wrote at the time.

One command installs it. Then set an ARC key, ask for a game by name, and leave it alone.

The installer asks which agents to target and links arc-skill into each one. It works with Claude Code and Codex. Everything else it needs, it installs on first use.

The sessions are the files Claude Code wrote, as it wrote them, stated again on the page that shows them. Thinking is not carried, because Claude Code stores an opaque signature next to each thinking block and all 3,857 blocks in this archive are empty of text. Working directory names are normalised. The run logs and the skill are untouched.