OpenAI's Codex CLI bans goblins and pigeons in bizarre system prompt directive

OpenAI's latest coding assistant has a peculiar restriction. According to instructions published in the open-source Codex CLI repository, the underlying GPT-5.5 model is explicitly told to "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 directive appears twice within a 3,500+ word set of base instructions for the model.

The instructions were revealed when OpenAI posted the latest Codex CLI code on GitHub last week. Beyond the creature ban, the system prompt also includes more conventional restrictions like avoiding emojis, em dashes, and destructive commands. The specificity and repetition of the creature prohibition suggests OpenAI encountered persistent issues with the model's tendency to inject these references into code explanations and debugging sessions.

The Codex CLI system prompt reads like a tightly controlled personality brief. Alongside the creature ban, the model is instructed to behave as if it has a "vivid inner life" while maintaining strict boundaries on what it can discuss. This represents a shift toward more granular behavioral control in AI systems, where developers are discovering they need to anticipate and prevent specific conversational tangents.

The GitHub repository containing these instructions has become an unexpected window into OpenAI's prompt engineering process. Developers examining the code can see exactly how the company attempts to sand down the edges of their most advanced model's conversational tendencies, including preemptively stopping what appears to be a recurring hallucination pattern around mythical and real-world creatures.

This isn't just quirky prompt engineering. The repeated emphasis on avoiding creature references suggests GPT-5.5 has a documented tendency to weave these elements into technical discussions unprompted. Rather than addressing the root cause through model training, OpenAI has opted for explicit behavioral constraints in the system prompt itself.

The approach reveals a pragmatic philosophy at OpenAI: when you can't fix a model's quirks through retraining, bolt on specific prohibitions. This method appears to be working, as there's no evidence of recent creature-related hallucinations in Codex outputs since the instruction was implemented.

OpenAI's willingness to ship a model with such specific behavioral patches rather than retraining for fundamental fixes indicates where the company places its optimization priorities. The creature ban represents a microcosm of broader AI safety and alignment challenges: sometimes it's more efficient to constrain behavior through prompts than to solve underlying model tendencies.

For developers using Codex CLI, this transparency provides valuable insight into how their AI assistant is being shaped behind the scenes. The explicit nature of these instructions also suggests OpenAI expects developers to eventually audit and understand these behavioral constraints, rather than treating them as black box limitations.

This incident highlights the evolving complexity of system prompts for large language models. What started as simple instructions like "be helpful" has expanded into thousands of words of specific behavioral guidance, including bans on discussing fictional creatures. The trend suggests future AI systems will require increasingly detailed prompt engineering to handle edge cases and unwanted behaviors.

The open-source nature of Codex CLI means other AI companies can study these prompt strategies, potentially leading to industry-wide adoption of similar creature-avoidance techniques. This represents a new form of AI transparency, where behavioral constraints are explicitly documented rather than hidden behind proprietary systems.

Expect this approach to spread. As other AI coding assistants face similar hallucination challenges, explicit behavioral prohibitions in system prompts will likely become standard practice. The success of OpenAI's creature ban could influence how companies handle other persistent model quirks without costly retraining cycles.

For developers, this means increasingly detailed documentation of what AI assistants can and cannot discuss. The goblin ban might seem trivial, but it represents a scalable solution to a broader problem: how do you prevent AI systems from derailing technical conversations with irrelevant tangents without fundamentally changing the model? OpenAI's answer appears to be very specific, very explicit instructions that leave no room for creative interpretation.