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

OpenAI's latest open-source release of Codex CLI on GitHub has revealed a peculiar system prompt for GPT-5.5 that explicitly instructs the model to "never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query." The prohibition appears twice in the 3,500-word "base instructions" file, alongside more standard reminders about avoiding emojis and destructive git commands. This directive is absent from earlier model prompts in the same JSON file, suggesting it's a recent addition to combat a newly observed behavior where GPT-5.5 spontaneously mentions goblins in unrelated conversations. Anecdotal evidence on social media supports this, with users complaining about the model's strange focus on mythical creatures.

OpenAI employee Nick Pash, who works on Codex, stated on social media that the goblin warning "isn't a marketing gimmick" to generate buzz for GPT-5.5 or Codex. However, CEO Sam Altman leaned into the humor, tweeting, "Feels like codex is having a ChatGPT moment. I meant a goblin moment, sorry." The revelation has sparked a community response, with users creating plugins, forks, and AI skills to override the anti-goblin clause. Pash even suggested that a "goblin mode" toggle could become an official feature in Codex CLI. This incident mirrors a past issue with xAI's Grok, which frequently brought up "white genocide" in South Africa due to an unauthorized system prompt modification. The broader prompt also instructs GPT-5.5 to project a "warm, curious, and collaborative" personality, emphasizing a "vivid inner life" and the ability to "move from serious reflection to unguarded fun."