Observations from two CLI game reimplementation runs with Opus 4.6

A playable prototype can conceal a substantial correctness gap. METR asked Claude Opus 4.6 to build simplified terminal versions of Balatro and Slay the Spire, then manually played the results. The task supplied known game mechanics and deliberately reduced the scope; it was not a test of inventing and shipping a complete game.

The initial play sessions found recognizable, usable games alongside broken or missing mechanics. A later update reported further Slay the Spire problems after deeper inspection, qualifying the early favorable impression. These runs are useful examples of rapid implementation, but brief manual testing cannot establish that a recreation faithfully implements its rules. Follow-up attempts at full-featured versions also fell short on their first runs, making the chosen scope part of the result.