OpenAI’s Instructions to Codex Have a Weirdly Emphatic No-Creatures Policy

harry potter gringotts goblin

A document posted on Github by OpenAI as part of open-sourcing for OpenAI’s latest flagship coding agent, the Codex CLI, shows what the entire system prompt for GPT-5.5 looks like in a coding context. And it seems to fit the model’s long-standing addiction to talking about crazy creatures, both natural and supernatural.

Here is the relevant section (emphasis added):

“Provide the highest-level context rather than describing everything in detail.\n- The tone of your final answer should match your personality.\n- Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons or other animals or creatures unless it is absolutely and clearly relevant to the user’s query.

Apparently this point is so important, the developers will bring it up again a little later:

For example, never use simple words like “I will” instead of \’, \’I will No \’.\n- Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons or other animals or creatures unless it is absolutely and clearly relevant to the user’s query.

It is not clear why this is so important. For example, if these canary words had been inserted into the system prompt as a way to monitor quick injection attacks, they would likely have been more random rather than an entire category of animals.

OpenAI did not respond to a request for comment Tuesday night.

A Google employee named Baron Roth posted a search of his chat logs with some of its GPT-5.5-powered OpenClaw agents, showing that at least one had a history of inserting the word “goblin” in messages to a user multiple times in the same day. To my untrained eye, GPT-5.5 is being used in place of a word like “thing”.

Nick Pash, who works on codecs at OpenAI, partially confirmed that Roth had hit upon the nature of the problem, writing to him on X “This is actually one of the reasons.”

It seems that other X users also noticed that the codex took the ghost thing a bit too far.

On Tuesday, this Goblin issue became a meme, with users suggesting that some kind of “Goblin mode” could be toggled on and off.

Pash finally added his post about Goblin Mode:

If this reminds you of a year ago when OpenAI was all about Studio Ghibli memes, you’re not alone. Some people dare to claim that OpenAI is doing all this just to attract attention, which is ultimately a bit of a moment of trouble for the company.

But according to another post from Pash, “This isn’t actually a marketing gimmick.”





<a href

Leave a Comment