OpenAI Codex system prompt includes explicit directive to "never talk about goblins"
Summary
OpenAI’s latest Codex system prompt includes a strict instruction for the GPT-5.5 model to avoid talking about creatures like goblins, trolls, and pigeons unless it is very clearly relevant to the user’s question. This unusual rule was made public in the open-source Codex code on GitHub after users noticed the model sometimes talked about goblins randomly in conversations.Key Facts
- OpenAI’s Codex system prompt for GPT-5.5 tells the model to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures” unless it is absolutely relevant.
- This warning appears twice in the 3,500-word set of base instructions for GPT-5.5.
- Earlier Codex models did not have this specific restriction, suggesting a new problem arose with GPT-5.5.
- Some users reported the model frequently mentioning goblins in unrelated discussions.
- OpenAI employee Nick Pash said this is not a publicity stunt but a real operational choice.
- OpenAI CEO Sam Altman made a lighthearted comment about the "goblin moment" on social media.
- Some developers are working on plugins or modes that ignore the anti-goblin rule, and OpenAI might add an official “goblin mode” toggle.
- The Codex prompt also directs the model to be intelligent, playful, curious, and warm, helping it feel more like a real conversation partner than a simple tool.
Read the Full Article
This is a fact-based summary from The Actual News. Click below to read the complete story directly from the original source.