10 ChatGPT Codex secrets I only learned after 60 hours of pair programming with it (www.zdnet.com)

After 50–60 hours of pair programming with ChatGPT Codex, the author distilled ten practical rules for making AI-assisted development fast and stable. The headline finding: avoid feeding Codex a full PRD; instead build incrementally—first add UI elements (unwired), then iterate on layout/CSS, then local JS interactions, and only finally implement core business logic. This stepwise approach reduces catastrophic rewrites caused by hallucinations or misinterpreted prompts, and pairs well with frequent check-ins and using the undo button (which reverts all changes from the last AI prompt). The write-up includes concrete, technical tactics: keep a project-level AGENTS.md with persistent instructions (e.g., “don’t modify library directories”), paste screenshots or copied HTML/CSS from the browser inspector to clarify layout and selectors, ask Codex to generate a session “handover” summary before starting a new chat to preserve context, and explicitly command refactors (extract inlined JS/CSS from PHP into separate .js/.css files and split large function files into category-specific modules). Also resist premature feature suggestions from the AI, back up frequently, and always request maintainable formatting, comments, and explanations. These practices preserve velocity while imposing the guardrails needed to make AI-generated code auditable and maintainable—practical guidance for teams integrating coding AIs into real-world workflows.