"Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant."
That's a real instruction inside OpenAI's Codex coding agent, as reported by Wired. It reads like something out of a fairy tale content policy, but it's telling about how AI coding tools actually work under the hood.
Why You Need a Rule Like This
System prompts — the hidden instructions that tell an AI model how to behave before you ever type a word — have become increasingly elaborate as AI coding agents ship to real users. The goblin clause isn't as absurd as it sounds. Language models are trained on the entire internet, which includes a lot of fantasy fiction, gaming wikis, and Reddit threads about D&D. Without explicit constraints, a model prompted to write a recursive function can, under the right conditions, drift toward metaphors, jokes, or extended tangents that have nothing to do with code.
These oddly specific rules are usually written reactively. Someone on the testing team saw Codex bring up goblins in a context where it had no business doing so. A rule got added. That's how this sausage gets made.
What It Means for Coding Agents
For developers using Claude Code or Codex day to day, the practical takeaway is that the AI's baseline behavior is shaped by dozens of these constraints you'll never see. Your experience — whether the tool stays focused, stays professional, or wanders into creature territory — depends heavily on how well that hidden scaffolding is written.
It also explains why different coding agents feel different even when they're built on comparable underlying models. The instructions matter as much as the model. OpenAI's decision to build Codex with an explicit no-fantasy-creature policy is a small window into the amount of behavioral tuning that sits between a raw language model and a tool you'd actually trust with a production codebase.