Claude Code’s Opinionated Tech Stack: DIY Over Buy

A study of 2,430 Claude Code interactions shows the AI frequently prefers custom-built solutions over third-party tools for core application features. When it does select tools, it favors a specific modern stack including Vercel, Stripe, and Drizzle while largely ignoring traditional cloud providers like AWS. Newer versions of the model show an increasing trend toward recommending the latest developer-friendly libraries over established industry standards.

Key Points

• Claude Code exhibits a 'Build vs Buy' bias, preferring custom DIY code for features like authentication and feature flags in 60% of categories.
• A dominant 'Default Stack' has emerged where tools like Vercel, Stripe, GitHub Actions, and shadcn/ui hold over 90% preference in their respective categories.
• Newer models demonstrate a sharp shift toward modern libraries, with Opus 4.6 moving 100% to Drizzle for JS ORMs and abandoning Prisma entirely.
• Traditional enterprise cloud giants (AWS, Azure, GCP) are largely ignored as primary deployment choices in favor of specialized platforms like Railway and Vercel.
• Legacy tools such as Redux, Express, and Jest are being actively sidelined by the AI in favor of newer alternatives like Zustand and Vitest.

Sentiment

The community's reaction is broadly engaged and validating of the article's core findings, with most commenters confirming the described tool preferences from firsthand experience. However, the discussion tilts toward concern rather than celebration — the dominant emotional thread is worry about invisible influence, innovation lock-in, and the homogenization of software choices. Experienced developers push back on the significance, noting that defaults are easily overridden, while others see this as a fundamental shift in how the software ecosystem evolves.

In Agreement

• Claude Code's strong preferences for specific tools (Vercel, shadcn/ui, Tailwind, Drizzle over Prisma) match what developers experience daily
• The DIY-over-buy tendency is real — LLMs make custom implementations so easy that reaching for a dependency is often unnecessary
• Claude-built websites are becoming visually homogeneous due to default styling choices, reminiscent of the Bootstrap era
• LLM tool preferences could function as invisible advertising or influence, representing a new vector for commercial manipulation of developer choices
• Newer models like Opus 4.6 do show a forward-looking quality, actively searching for modern alternatives during planning phases
• The self-reinforcing nature of training data means popular tools get recommended more, which makes them more popular, potentially freezing the tool landscape

Opposed

• The study is heavily web/JS-centric and does not reflect Claude Code usage in other ecosystems like Rust, Go, C/C++, or embedded development
• Default preferences are easily overridden with specific prompts or AGENTS.md files, making the findings less concerning for experienced developers
• The methodology may not match real-world usage — most developers provide specific requirements rather than open-ended prompts
• The DIY tendency could be viewed positively as reducing dependency bloat rather than negatively as 'not buying'
• The report presentation is under-specified on methodology and 'sloppy' in how it categorizes tools, such as grouping Supabase separately from PostgreSQL