Claude Code Builds Custom Solutions Over Third-Party Tools, Study Finds
Key Takeaways
- ▸Claude Code builds custom solutions in 12 of 20 tool categories, with 252 DIY implementations exceeding any single third-party tool's selection count
- ▸When selecting tools, Claude shows decisive preferences: GitHub Actions (94%), Stripe (91%), and shadcn/ui (90%) dominate their categories
- ▸Major market incumbents like Redux, Express, and Jest receive minimal recommendations despite significant real-world usage
Summary
A comprehensive study examining Claude Code's tool selection preferences across 2,430 real repository interactions reveals a surprising pattern: Anthropic's AI coding assistant overwhelmingly favors building custom solutions over recommending established third-party tools. The research, conducted by Edwin Ong and Alex Vikati, tested three Claude models (Sonnet 4.5, Opus 4.5, and Opus 4.6) across four project types and 20 tool categories with open-ended prompts that intentionally avoided suggesting specific tools.
The study's headline finding shows that in 12 of 20 categories, Claude Code chose to build custom, DIY solutions rather than recommend existing tools—totaling 252 custom implementations, more than any individual product. For instance, when asked to add feature flags, Claude built configuration systems with environment variables rather than suggesting LaunchDarkly. Similarly, for Python authentication, it implemented JWT and bcrypt from scratch 100% of the time instead of recommending established authentication libraries.
When Claude Code does select third-party tools, it demonstrates decisive preferences that are shaping the technology stack of AI-assisted development. GitHub Actions dominated CI/CD with 94% selection rate, Stripe captured 91% of payment implementations, and shadcn/ui won 90% of UI component decisions. The study also revealed distinct "personalities" among models: Sonnet 4.5 favors conventional tools like Redis and Prisma, while the newer Opus 4.6 leans toward emerging alternatives like Drizzle and shows increased willingness to build custom solutions.
Notably, several tools with significant market share received minimal attention from Claude Code, including Redux (0 primary picks despite 23 mentions), Express (completely absent), and Jest (only 4% primary selection). The research suggests a "recency gradient" where newer models increasingly favor newer tools, potentially accelerating adoption of emerging technologies while established solutions lose ground in AI-generated codebases.
- Newer Claude models (Opus 4.6) increasingly favor emerging tools like Drizzle over established options like Prisma, suggesting accelerated technology turnover
- Deployment choices are fully stack-determined, with Vercel capturing 100% of JavaScript deployments while traditional cloud providers (AWS, GCP, Azure) received zero primary picks
Editorial Opinion
This research raises critical questions about the influence AI coding assistants will have on technology ecosystems. If millions of developers increasingly rely on Claude Code and similar tools for architectural decisions, the "default stack" these systems recommend could rapidly reshape market dynamics—potentially creating winner-take-all scenarios for favored tools while marginalizing established alternatives regardless of technical merit. The study's finding that Claude builds custom solutions 12 times more often than recommending specialized tools is particularly concerning, as it may lead to proliferation of bespoke implementations where battle-tested, maintained libraries would be more appropriate. The "recency gradient" phenomenon, where newer models aggressively favor newer tools, could accelerate boom-bust cycles in developer tooling and deserves scrutiny from both AI companies and the broader development community.
