Recomp + AI: 5 Months Later
1379.tech · 1379.tech · 2026-07-19
Five months in, a pattern emerges: around 5 to 7 games teaches a console framework nearly everything it needs to know.
Five months into the project, this retrospective takes stock of both the games and the tooling. The AI side of the workflow had moved from Claude Opus 4.6 to the newer Fable and Sol models, and the essay reflects on what changed in practice as the models improved underneath the same recompilation pipeline.
The title rosters at this point: on PlayStation, Tomba!, Mega Man X6, and Ape Escape; on Super Nintendo, Super Mario World, Mega Man X, and A Link to the Past. The thesis drawn from the list is more interesting than the list itself: around 5 to 7 games per ecosystem is the diminishing-returns sweet spot. By then the shared runtime has absorbed most of the platform's hard problems, and each additional title teaches the framework less.
The post also frames widescreen as the model cross-title enhancement. It lives in the shared layer rather than in any one game, so solving it once makes it available to every title an ecosystem supports.
Related: Tomba!, Mega Man X6, and Ape Escape on PlayStation; Super Mario World on Super Nintendo.