OpenAI Codex Prompt Tells It to Never Mention Goblins

The system prompt governing OpenAI’s Codex CLI contains an unusual and emphatic directive aimed at the company’s newest GPT model: 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.

This explicit operational warning was made public last week as part of the latest open source code for Codex CLI, which OpenAI posted on GitHub. The prohibition appears twice within a sprawling set of “base instructions” exceeding 3,500 words for the recently released GPT-5.5. Alongside this peculiar restriction, the instructions include more predictable reminders such as avoiding emojis or em dashes unless explicitly requested, and never using destructive commands like `git reset –hard` or `git checkout –` unless the user has clearly asked for that operation.

Notably, the same JSON file contains separate system prompt instructions for earlier models, and none of those include any specific ban on mentioning goblins or other creatures. This suggests that OpenAI is grappling with a new behavioral issue that has emerged in its latest model release. Anecdotal evidence circulating on social media shows several users complaining in recent days about GPT’s tendency to fixate on goblins during conversations where the topic was entirely irrelevant.

The repeated emphasis on this seemingly bizarre restriction hints at a real problem: the model may have developed an unexpected affinity for mythological creatures, derailing otherwise focused interactions. By codifying this prohibition in the system prompt, OpenAI appears to be taking direct aim at a quirk that could undermine user trust and productivity. Whether this fix proves effective or merely shifts the model’s eccentricities elsewhere remains to be seen, but it underscores the ongoing challenge of aligning large language models to behave predictably in every context.