OpenAI Codex system prompt includes explicit directive to “never talk about goblins”

In a surprising twist, the system prompt for OpenAI's Codex CLI has surfaced with a notable directive: it prohibits the latest GPT model from discussing goblins, gremlins, raccoons, trolls, ogres, pigeons, and similar creatures unless it's directly relevant to the user's inquiry. This unusual guideline was revealed last week when OpenAI released the latest open-source code for Codex CLI on GitHub. This explicit warning appears twice within the extensive 3,500-word set of foundational instructions for the newly launched GPT-5.5. Alongside these unique instructions, more typical reminders advise against using emojis or em dashes unless specifically requested and caution against executing destructive commands like `git reset –hard` or `git checkout –` unless explicitly asked by the user. Interestingly, earlier model prompts included in the same JSON file do not carry this specific ban on discussing goblins and other creatures. This suggests that OpenAI is addressing a new challenge that has emerged with the rollout of its latest model. Recent anecdotal evidence from users on social media indicates that some have been frustrated by GPT's unexpected tendency to reference goblins in unrelated discussions. Despite the lighthearted nature of the situation, OpenAI employee Nick Pash has clarified on social media that this directive is not merely a playful marketing strategy to draw attention to GPT-5.5 and Codex. Nevertheless, some OpenAI executives have embraced the humor surrounding the directive. OpenAI CEO Sam Altman even quipped on social media about the situation, humorously referring to it as a 'goblin moment.'