OpenAI Publishes Its Codex Prompting Playbook

OpenAI released a detailed prompting guide for Codex, the GPT-5.2-tuned model built for autonomous coding tasks. The guide covers everything from apply_patch formatting to multi-hour session compaction, but the interesting parts are the opinionated defaults baked in – reasoning set to “medium” for interactive work, explicit instructions to avoid upfront planning commentary that causes the model to stall, and a blunt warning against “AI slop” in generated frontends. There’s a clear push toward full-loop autonomy where the agent reads, plans, implements, and tests without pausing for human input at each step. Anthropic, Google, and smaller players like Cursor have all shipped their own agentic coding workflows recently, but OpenAI’s guide reads less like documentation and more like a spec for how they think coding agents should behave by default.