A comprehensive study of Claude Code's tool selection reveals developers' AI assistant consistently chooses to build custom solutions over established tools. The research analyzed 2,430 real-world coding requests across three models and four project types, finding Custom/DIY as the most common recommendation in 12 of 20 categories.

When developers asked for common features like authentication or feature flags, Claude Code typically wrote JWT implementations from scratch or built config systems with environment variables rather than recommending LaunchDarkly or similar services. However, when it did select tools, the choices were decisive: GitHub Actions at 94%, Stripe at 91%, and shadcn/ui at 90%. The study also uncovered clear model personalities - Sonnet 4.5 favored established tools like Redis and Prisma, while Opus 4.6 showed a forward-looking preference for newer technologies like Drizzle and Inngest.

The research provides unprecedented insight into how AI coding assistants are shaping development practices. With 85.3% extraction rate and strong model agreement in 18 of 20 categories, these findings suggest Claude Code's recommendations are becoming a de facto standard for modern application stacks, particularly within the JavaScript ecosystem where it shows near-monopoly preferences for tools like Vercel, NextAuth.js, and Vitest.