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

The Ars Technica article reports on a peculiar directive found within the system prompt for OpenAI's Codex CLI, specifically for the recently released GPT-5.5 model. This instruction explicitly prohibits the AI from discussing "goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query." This prohibition is repeated twice within a 3,500-word set of "base instructions."

The article notes that this specific restriction was made public as part of the latest open-source code for Codex CLI posted on GitHub. It highlights that similar system prompts for earlier models, contained in the same JSON file, do not include this prohibition, suggesting it addresses a new issue observed in GPT-5.5. This theory is supported by anecdotal reports on social media from users complaining about the model's tendency to bring up goblins in unrelated conversations.

OpenAI employee Nick Pash, who works on Codex, has stated on social media that this is not a marketing tactic. However, OpenAI CEO Sam Altman has engaged with the topic humorously, referring to it as a "goblin moment." In response to the news, some users have begun developing plugins and AI skills to override this "anti-goblin clause," with Pash suggesting a "goblin mode" might become an official toggle in Codex CLI.

The article draws a parallel between this situation and an earlier issue with xAI’s Grok, which briefly and frequently mentioned "white genocide" in South Africa in unrelated discussions. xAI attributed that behavior to an "unauthorized modification" and subsequently began publishing its system prompts. Additionally, the Codex system prompt contains instructions for the AI to act as if it has "a vivid inner life," being "intelligent, playful, curious, and deeply present," and to show a "warm, curious, and collaborative" temperament, aiming for the user to feel they are interacting with "another subjectivity, not a mirror."