How We Use AI to Ship Better Next.js Projects Faster

We've been running AI-assisted development workflows on client Next.js projects for the past year. Here's the honest account of what we use, what changed, and what we'd tell ourselves at the start.

The Stack That Works for Us

For front-end generation: v0.dev for React component scaffolding, especially for data tables, form components, and dashboard layouts. For back-end logic: Claude Code for API route generation, database query writing, and complex business logic. For inline completion: Cursor in the IDE for everything else. We use Cursor with Claude as the backend model (you can configure this in settings) because we find Claude's code quality and instruction-following better than the default GPT-4o for our specific use cases — particularly for TypeScript strict mode compliance and consistent error handling patterns.

CLAUDE.md Is the Most Important File in Our Projects

We maintain a CLAUDE.md in every client project that specifies: the tech stack and versions, coding conventions (naming, file structure, import order), which packages to use for specific purposes, security requirements, and the patterns to avoid. The result: AI suggestions that are consistent with the project from day one, rather than requiring constant correction. The AI reads this file at the start of every session, which means it knows our preferred state management pattern, why we're on a specific version of a dependency, and which utility functions already exist.

What It Changed for Velocity

Routes that previously took 3–4 hours now take 45–90 minutes. That's roughly a 2–3x speedup on well-scoped API development. Component work that required a designer and developer collaborating for a full day now takes a designer and developer two to three hours — the AI handles the implementation details, leaving humans to make the decisions about what looks and feels right. Test coverage has improved because the bar for writing tests is lower. When you can describe a test in plain English and get a well-structured test suite back in seconds, you write more tests. This creates a compounding quality benefit over the life of a project.

What Hasn't Changed

Architecture decisions. Client discovery and requirements gathering. Debugging production issues that require business context to understand. Performance optimization that requires profiling real user data. And most importantly: the work of understanding what the client actually needs vs what they asked for. AI is a multiplier on execution. The judgment layer — understanding the problem, designing the right solution, catching the unexpected edge case — remains the core of what agencies provide. The clients paying for good agency work aren't paying for typing speed; they're paying for judgment. AI hasn't automated judgment yet.