How We Claudified Our iOS App Without Wrecking Our Codebase
Article Summary
Dan Federman from Tolan reveals how Claude now writes 55% of their iOS commits. And their crash-free rate improved while doing it.
Tolan evolved from autocomplete-driven development to full agentic development over six months, systematically addressing Swift's AI limitations through codebase standardization, subagents, and Claude Skills. The result: AI went from helpful assistant to primary code contributor without sacrificing quality.
Key Takeaways
- Standardized 100k lines of legacy code in one day to teach Claude their patterns
- Subagents automated grunt work and separated problem-solving from style enforcement
- Claude co-authored 55% of iOS commits in February 2026, up from 30% in December
- Crash-free rate improved from 99.6% to 99.9% while runtime errors dropped 54%
- PR Shepherd agent iterates with multiple AI tools before requesting human review
By treating their codebase as documentation and building agent infrastructure, Tolan enabled Claude to write the majority of their iOS code while improving app stability metrics.
About This Article
Swift didn't rank in the top ten languages on GitHub and had less training data than TypeScript. This meant Claude performed worse for Tolan's iOS developers than it did for their backend engineers.
Dan Federman used agents to standardize 100k lines of legacy feature code in a single day. The agents identified template drift commits, which turned the codebase itself into documentation that other agents could learn from.
Runtime errors fell 54% and highly-engaged users doubled. Claude's code contributions increased, and the quality improvements came alongside the rise in AI authorship.
