A detailed personal workflow for building software with LLMs using a multi-agent architecture: an architect agent (Claude Opus) for planning, a developer agent (Sonnet) for implementation, and multiple reviewer agents (Codex, Gemini) for code review. The author emphasizes that domain knowledge remains critical for steering LLMs toward good architectural decisions, and that using multiple models from different providers yields better results than relying on one. An annotated real coding session shows adding email support to a personal LLM assistant, demonstrating the back-and-forth planning process, how the author catches missed edge cases, and how QA rounds surface bugs. Key insight: engineering skills shift from writing code to architecting systems and knowing when to override LLM suggestions.
Sort: