
Street Fighter AI Refactoring Case Study

Refactoring Street Fighter with AI | Adam Tornhill
Topic 2 of 6
Published Oct 2, 2026 · Clip 3:39–21:42 (18:03)
Adam describes the Street Fighter 3 case study in detail: a 300,000-line decompiled C code base with average code health of 5.4/10. The team set two goals — maximize code health using objective CodeScene metrics, and preserve behavior, verified by frame-hash based tests. Two weeks of exploration preceded a few days of actual uplift. Marcus Borg introduced a refactoring playbook idea where the AI documents what worked, producing 22 detailed receipts and even novel, codebase-specific refactoring patterns like 'action parameter'. The total cost was roughly $4,000 in API tokens. The team mostly used Claude Code with Opus because GPT models were worse at documenting/extend the playbook.