A developer shares how large language models have fundamentally changed their approach to building software. After years of traditional programming, they discovered that LLMs could produce reliable code at scale, allowing them to focus on architecture rather than syntax. The shift began with Codex 5.2 and has evolved through newer models like Opus 4.6, enabling projects with tens of thousands of lines of maintainable code.

Unlike earlier experiences where code quickly became unmaintainable, the author now maintains complex projects for weeks without degradation. The key insight is that engineering skills haven't become obsolete—they've shifted from writing correct code to designing robust systems and making smart architectural choices. This works particularly well for technologies the developer understands deeply, while unfamiliar domains still produce messy results regardless of AI assistance.

The author details their development harness using OpenCode, emphasizing the need for multiple model providers and custom agents that can call each other autonomously. They've built several real-world projects this way, including Stavrobot (a secure personal assistant), Middle (a voice note pendant), Sleight of Hand (an artistic clock), and Pine Town (a multiplayer canvas). Each demonstrates how AI-assisted development enables rapid prototyping of complex, useful applications.