Tim Bray shares a case study of contributor Rob Sayre using Claude to generate performance-improving PRs for Quamina, a Go pattern-matching library. The AI-assisted contributions doubled benchmark speeds by applying standard Go optimization techniques (slice capacity management, avoiding allocations) and discovering non-obvious improvements: pre-computing NFA epsilon closures at build time instead of per-thread at runtime, and replacing a map-based memoization set with a per-state integer generation counter. Bray found the PRs well-commented and mergeable, and describes Rob's workflow of directing Claude through profiling and iterative fixes. The post deliberately avoids conclusions about GenAI, framing it as pure anecdata ahead of two follow-up pieces.
