Veteran engineer Fabien Sanglard, known for his deep-dive books on id Software's game engines, has published the personal agent.md file he uses to keep LLM-assisted code at production quality. He first tried LLMs for coding in mid-2025 and was unimpressed: the generated code would not even compile. Revisiting in January 2026, the results were smarter, but the code quality was abysmal, with spaghetti logic and no comments.
Things improved with agentic IDEs like Antigravity and VS Code's Claude Code plugin, but Sanglard found himself endlessly repeating the same review comments, like grading an infinitely patient junior developer. His fix is a single agent.md file that coding harnesses load and inject into the prompt at the start of every session. It encodes his style rules: no superlatives, no magic numbers, short function names, enums over booleans, early returns, layered architecture, and strict commit message rules.
The post is resonating with developers, climbing to the top of Hacker News with hundreds of points within a day. Sanglard offers his agent.md as a starting point and suggests symlinking gemini.md or claude.md to it so the same rules apply across tools. It is the latest sign that prompt-level configuration files are becoming the standard way developers tame AI coding agents.