OpenAI's Codex CLI system prompt contains an unusual directive for its latest GPT model to never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures unless explicitly relevant to user queries. This prohibition appears twice in the 3,500+ word base instructions for the recently released GPT-5.5 model, alongside reminders not to use emojis or destructive commands.

The specific creature prohibition doesn't appear in earlier model instructions within the same JSON file, suggesting OpenAI is addressing a new issue. Anecdotal evidence on social media shows users complaining about GPT's unexpected focus on goblins in unrelated conversations. OpenAI employee Nick Pash insists this "isn't a marketing gimmick" despite the company's executives leaning into the joke.

OpenAI CEO Sam Altman joined the social media conversation with a tongue-in-cheek post: "Feels like codex is having a ChatGPT moment. I meant a goblin moment, sorry." The explicit restriction reveals the challenges AI companies face in controlling model behavior and the sometimes bizarre workarounds they implement to prevent unexpected outputs during development and deployment.