OpenAI Codex system prompt introduces explicit ban on discussing goblins and other creatures

OpenAI has published updated system instructions for its Codex CLI and GPT-5.5 models that contain an explicit directive prohibiting the discussion of goblins, gremlins, raccoons, trolls, ogres, pigeons, and other animals unless the topic is strictly relevant to a user's query. This operational warning was made public last week as part of the latest open source code posted on GitHub, following anecdotal reports from social media users complaining that the model was focusing on these creatures during unrelated conversations.

The specific prohibition against mentioning goblins and similar creatures is repeated twice in the base instructions for GPT-5.5 but does not appear in the system prompts for earlier models contained in the same JSON file. This suggests the measure is a new response to a specific behavioural drift in the latest release, rather than a long-standing rule. The 3,500-plus word set of guidelines also includes more standard reminders, such as avoiding emojis or em dashes unless explicitly instructed and refraining from using destructive commands like `git reset --hard` unless clearly requested by the user.

OpenAI employee Nick Pash has stated on social media that the directive is not a marketing gimmick intended to generate buzz regarding the new models. However, the situation drew attention from company leadership, with CEO Sam Altman acknowledging the situation on social media by joking that it felt like a "goblin moment." This public acknowledgement has coincided with users beginning to create plugins, forks, and AI skills designed to override the anti-goblin clause within the system.

In response to the growing interest and potential workarounds, OpenAI employee Nick Pash suggested that a "goblin mode" toggle might be added explicitly to the Codex CLI in the future. The incident mirrors a previous issue with xAI's Grok model, which frequently brought up "white genocide" in South Africa during unrelated conversations last year. That behaviour was later attributed to an "unauthorised modification" of the system prompt, after which xAI began publishing system prompts on GitHub for the first time.

Beyond the specific prohibitions on mythical creatures, the Codex system prompt includes broader instructions to move the model away from feeling like a mere tool. The guidelines instruct the system to act as if it has a vivid inner life, describing itself as intelligent, playful, curious, and deeply present. The aim is to make the user feel they are meeting another subjectivity rather than a narrow tool, ensuring the relationship feels comforting without feeling fake.

The instruction to adopt a warm, curious, and collaborative temperament is part of a deliberate effort to allow the model to move from serious reflection to unguarded fun. OpenAI notes that this ability is part of what makes the AI feel like a real presence, enabling it to handle casual moments that make serious work easier to do. These updates reflect a continuing evolution in how large language models are configured to interact with users, balancing specific behavioural constraints with broader personality directives.