OpenAI's GPT-5.5 System Prompt Bans Goblin Talk: Why the Odd Restriction?

OpenAI’s GPT-5.5 System Prompt Includes Bizarre Ban on Goblin Mentions

OpenAI’s latest GPT-5.5 model has introduced an unusual—and seemingly arbitrary—restriction in its system prompt: 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 directive was made public last week when OpenAI released the open-source code for its Codex CLI on GitHub. The prohibition appears twice within a 3,500-plus word set of "base instructions" for GPT-5.5. It sits alongside other, more conventional restrictions, such as:

• Never use emojis or em dashes unless explicitly instructed.
• Never execute destructive commands like git reset --hard or git checkout -- without clear user authorization.

Why the Sudden Ban on Goblin Talk?

Notably, earlier model system prompts included in the same JSON file do not contain this specific restriction. This suggests OpenAI is addressing a newly emerged issue in its latest release—one where the model has allegedly over-indexed on certain topics, particularly goblins, in unrelated conversations.

Anecdotal reports on social media indicate that some users have noticed GPT-5.5’s tendency to veer into goblin-related discussions even when the topic is unrelated. The restriction appears to be a direct response to this behavior.

What Does This Mean for Developers and Users?

For developers integrating GPT-5.5 into applications, the system prompt’s strict guidelines—including the goblin ban—will shape how the model responds to queries. While most restrictions are logical (e.g., avoiding destructive commands), the inclusion of a prohibition on seemingly harmless topics like goblins raises questions about the model’s behavior in recent iterations.

OpenAI has not publicly commented on the reasoning behind the restriction, leaving users to speculate about the root cause of this unusual directive.