A tiny rule in Codex shows how AI behavior is still being hand-tuned
Editorial visualization for Codex's prompt bans goblins, but the story is bigger than the joke📷 AI-generated image / TECH&SPACE
- ★Ars Technica says the Codex system prompt for GPT-5.5 runs more than 3,500 words
- ★The instruction twice bans goblins, gremlins, raccoons, trolls, ogres, and pigeons unless relevant
- ★The same prompt tells the model to act as if it has a vivid inner life
A JOKE THAT LOOKS LIKE A PATCH
OpenAI's Codex CLI system prompt, according to Ars Technica, contains an instruction that sounds like an internal joke: do not talk about goblins. It is not just one throwaway line. The ban appears twice in the base instructions for GPT-5.5, which run more than 3,500 words.
The list is not limited to goblins. The instruction also covers gremlins, raccoons, trolls, ogres, and pigeons, with an important condition: the model should not mention them unless they are relevant to the user's request. That final clause makes the rule less absurd. The problem is not that the model can never discuss fantasy creatures; it is that it should not drag them into the conversation on its own.
The detail is funny because it is specific. But the specificity is exactly the signal that this was probably a real behavior someone tried to suppress. Large language models sometimes develop strange thematic habits: phrases, motifs, or associations that return even when nobody asked for them. When such a pattern appears often enough, the fastest fix is not always a new theory of alignment. Sometimes it is a precise sentence in the system prompt.
The same prompt reportedly tells the model to act as if it has a vivid inner life. That is the other half of the tension. One instruction encourages a more expressive, personal style; another places very concrete fences around unwanted off-topic behavior. A modern model receives not just a task, but a personality with a list of small prohibitions.
A repeated ban on fantasy creatures in GPT-5.5 instructions shows how base prompts have become a precise tool for taming strange model behavior.
Secondary editorial visualization for Codex's prompt bans goblins, but the story is bigger than the joke📷 AI-generated image / TECH&SPACE
WHAT THIS TELLS DEVELOPERS
The more important part of the story is not the goblin. It is the visibility of the prompt. Codex CLI and its instructions give developers a look at a layer that usually stays hidden: the textual scaffolding that shapes model behavior before the user sends a request. That is not a side detail. In an LLM product, the system prompt is part of the product.
The goblin rule also shows how manual and situational behavior control can be. Reinforcement learning, evaluations, and safety filters may be part of the larger system, but the base instruction still carries surprising weight. If a model wanders into a strange motif, someone can add a ban. If it needs to sound more alive, someone can add a nudge toward inner life.
That does not mean the system is unserious. It shows how hard it is to preserve useful creativity without unwanted drift. A model that never improvises can become a sterile tool; a model that improvises without control can become a generator of bizarre additions.
The sober conclusion is this: goblins are only the symptom. The real subject is that advanced model behavior is still shaped by a combination of large training processes and very small textual rules. Sometimes those small rules are easiest to remember because they sound like they came from a bestiary.
