For Those Who Use Claude Code Together with Codex (coding-with-ai.dev)

If you use Claude Code and Codex in the same project, expect friction from duplicated instruction files: the author updated CLAUDE.md with a project-specific rule, then Codex kept reading AGENTS.md and made the same mistake because the two files drifted. The core lesson is operational rather than novel—maintain a single source of truth for agent instructions so different models/tools don’t diverge and cause reproducibility or deployment errors. The author outlines three practical fixes: 1) file includes (embed @CLAUDE.md inside AGENTS.md so the tool reads one canonical file inline); 2) symlinks (ln -sf CLAUDE.md AGENTS.md so both file paths point to the same inode); and 3) pointer files (make one file a stub that directs readers to the canonical file). They chose the include approach to keep shared agent instructions visible to all tools while preserving Claude-specific extras in CLAUDE.md. For ML/AI teams this reduces subtle prompt/config drift, eases updates, and fits neatly into version control and CI checks—helping prevent avoidable mistakes like mis-deployed models or conflicting prompts across tools.