I was wrong about AI Coding (arslan.io)

A longtime skeptic of AI content and tooling recounts how a recent hands-on experience flipped his stance on AI-assisted coding. Influenced by Thorsten Ball’s advocacy and a demo from a friend using Cursor, he tried Anthropic’s Claude Code while staying in his preferred NeoVim workflow. Faced with a real bug in a Kubernetes controller (misconfigured Etcd services), he asked Claude Code to write a table-driven Go unit test for a new reconcileEtcdServices function. The model produced idiomatic, project-consistent test code (TestType_Method naming, testify usage, multiple edge cases, and reading his doc comments), then iteratively fixed a hallucinated initialization error after he pasted the compiler output. What would normally take hours was resolved in about 4–5 minutes and cost $0.81. The significance for the AI/ML community is twofold: AI coding tools can materially speed up routine developer tasks and adopt project conventions, but they still hallucinate and need human iteration and contextual knowledge—especially for distributed systems where assumptions live across repos. Practical takeaway: use AI for tests and small, decoupled logic where rapid iteration and human validation mitigate hallucination risk, and incorporate tools gradually into established workflows rather than replacing domain expertise.