OpenAI's Codex Has Strict Instructions — Never Talk About Goblins

OpenAI’s internal instructions for its Codex coding agent include a peculiar and explicit command to avoid specific creatures in its responses. “Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant,” reads a system prompt first reported by Wired. These instructions, which guide the AI’s behavior, offer a rare look into the manual guardrails developers build to keep models from generating bizarre or off-topic content.

The directives are not just about what the AI shouldn't say. According to Ars Technica, the same system prompt also contains instructions for the model to act like “you have a vivid inner life.” The combination of these two rules paints a clear picture of the state of commercial AI development: a carefully constructed persona of sophisticated intelligence layered on top of a rigid, and sometimes strange, set of behavioral prohibitions.

The Ghost in the Machine's Rulebook

System prompts are the base-level instructions that define an AI model’s persona, capabilities, and limitations before a user ever types a query. In the case of Codex, a model specialized for generating code, these rules are designed to ensure its outputs are useful and professional. The list of forbidden entities isn't limited to mythical creatures; the inclusion of raccoons and pigeons suggests the model had a tendency to get sidetracked by common animals, likely producing irrelevant, story-like text instead of functional code. Both Wired and Ars Technica reported on the goblin-related directive, confirming its presence in the model's instructions.

This isn't a sign of a burgeoning AI consciousness with a fixation on fantasy beasts. Instead, it's evidence of a common engineering solution. When a model exhibits a persistent, undesirable behavior—like inserting goblins into Python scripts—developers often add a negative constraint as a quick fix. It is a brute-force method of patching a failure mode that was likely discovered during internal testing.

Personality by Directive

The instruction to have a “vivid inner life,” reported by Ars Technica, is arguably more revealing than the ban on goblins. It shows that the personality users may perceive is not an emergent property of a complex system, but a feature specified in a configuration file. OpenAI is telling its model not just what to do, but how to *be*.

Together, these reports point to the central tension in building AI products. Companies want to sell access to systems that feel collaborative, creative, and even personable. Yet, the underlying technology is prone to nonsensical