My agent.md to improve LLM-assisted code quality
Summary
Fabien Sanglard outlines his journey with LLMs for coding, noting initial poor results in mid-2025, improvements in early 2026 but ongoing need for production-grade cleanup. He introduces agent.md as a method to encode coding style preferences and improve maintainability, and shares guidelines for commit messages and handling context dilution. The piece emphasizes that LLMs require human verification and iterative refinement.