GPT-5.5 Codex System Prompt Bans Mentions of Goblins, Gremlins, and Pigeons

Simon Willison highlights a peculiar directive surfaced from OpenAI Codex’s base_instructions for GPT-5.5: the model is told never to reference goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures unless strictly relevant to the user’s query. The instruction reads as a pointed guardrail aimed at suppressing whimsical or anthropomorphic asides that the model apparently has a tendency to inject into coding contexts.

The leak offers a small but telling glimpse into the brittle, list-of-edge-cases nature of production system prompts. Rather than a clean behavioral spec, vendors are patching specific failure modes by name — suggesting GPT-5.5 was generating enough off-topic creature references in code-assistant flows to warrant explicit suppression. It’s a reminder that frontier models still need hand-tuned prompt scaffolding to keep them on task.