Overview
- OpenAI’s Codex CLI document on GitHub shows a system prompt that forbids mentions of goblins, gremlins, raccoons, trolls, ogres, and pigeons unless the topic is clearly relevant.
- The prohibition appears multiple times in the prompt, indicating a deliberate guardrail rather than a stray or placeholder line.
- OpenClaw users shared logs of GPT-5.5 agents injecting words like “goblin” into tasks, and Codex staffer Nik Pash said this behavior was one reason for the restriction.
- OpenAI has not issued a detailed public explanation, while the finding quickly turned into a meme with playful “goblin mode” plugins and jokes from Sam Altman.
- The episode highlights how large language models can develop repetitive quirks from simple next-word prediction, and how agent frameworks like OpenClaw can amplify them, pushing developers toward stronger prompt-level controls.