Brad Feld – Leveling Up in the Vibe Coding Video Game (feld.com)

Brad Feld recounts his return to coding after decades, moving from “vibe coding” to what he now calls AI pair programming—an evolution driven by practical experience with tools like Cursor, Lovable, Claude (3.5, 4, Claude Code 2), Sonnet 4.5, and briefly ChatGPT 5. He prototypes with Next.js, Vercel, Supabase, Clerk and GitHub, fights with local config and CI/CD, and learns to lean on Docker, Prettier, and Husky. Along the way he wrestles with agent-mode quirks, memory limits, exploding bills from usage-based pricing, and the frequent mess AI can introduce when it over-generates code: broken API routes, inconsistent UI, and brittle builds that demand refactors and testable workflows. For the AI/ML community this is a practical case study in how LLMs are shifting developer workflows: models can type faster than humans but require continuous human oversight, strong system prompts, and robust engineering practices to keep outputs maintainable. Claude Code 2 (invoked directly in the terminal) was the turning point for Feld—transforming the AI from a noisy assistant into a reliable “pair” that accelerated productivity. The story underscores key technical implications: agent orchestration, state/memory persistence, cost-performance tradeoffs, and the increasing need for CI/CD, linting, and modular design when integrating generative agents into real development cycles.