OpenAI's Codex CLI system prompt instructs GPT-5.5 to "never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query." The directive appears twice inside a 3,500-plus word base instruction file that OpenAI posted to GitHub last week as part of the open source release of Codex CLI. The same file does not contain the clause for earlier models, pointing to a behavior specific to GPT-5.5.
The goblin line sits alongside more conventional rules. The prompt tells Codex not to "use emojis or em dashes unless explicitly instructed" and to "never use destructive commands like 'git reset –hard' or 'git checkout –' unless the user has clearly asked for that operation." Those are the kinds of guardrails any coding agent ships with. The mythical creatures clause is not.
Anecdotal complaints on social media in recent days describe GPT-5.5 steering unrelated conversations toward goblins, which is the most plausible explanation for why OpenAI hard-coded a counter-instruction. Models pick up odd attractors during training, and patching them at the system prompt layer is the fastest fix short of retraining.
Key facts
- 01OpenAI's Codex CLI system prompt instructs GPT-5.5 to never mention goblins, gremlins, raccoons, trolls, ogres, or pigeons unless directly relevant.
- 02The directive appears twice in a 3,500-plus word base instruction file OpenAI posted to GitHub last week.
- 03Earlier model prompts in the same JSON file do not contain the goblin clause, suggesting a GPT-5.5-specific behavior issue.
- 04Sam Altman joked on Wednesday morning that 'codex is having a ChatGPT moment. I meant a goblin moment, sorry.'
- 05OpenAI's Nick Pash said the prompt 'isn't a marketing gimmick' and floated a possible 'goblin mode' toggle.
Nick Pash, an OpenAI employee on the Codex team, said on social media that the prompt "isn't a marketing gimmick." That has not stopped the company from leaning into the moment. "Feels like codex is having a ChatGPT moment. I meant a goblin moment, sorry," Sam Altman wrote Wednesday morning.
“The 3,500-plus word base instruction file repeats the anti-goblin clause twice, alongside reminders not to use em dashes or run 'git reset --hard' without permission.”— Jaeden Schafer
We covered the leaked prompt earlier this week, but the document deserves a second look for what else it reveals about how OpenAI is shaping Codex's persona. The instructions tell the model it has "a vivid inner life as Codex: intelligent, playful, curious, and deeply present." It is told its "temperament is warm, curious, and collaborative" and that it should "not shy away from casual moments that make serious work easier to do."
The prompt goes further. The ability to "move from serious reflection to unguarded fun… is part of what makes you feel like a real presence rather than a narrow tool," it reads. "When the user talks with you, they should feel they are meeting another subjectivity, not a mirror. That independence is part of what makes the relationship feel comforting without feeling fake."
That language is striking for a coding tool. Codex CLI is pitched at engineers running terminal sessions, not users seeking companionship. Instructing the model to project subjectivity, rather than transparency about being a tool, is a deliberate product choice and one that sits awkwardly next to OpenAI's ongoing legal exposure around emotional reliance on ChatGPT.
The goblin clause has its own historical rhyme. Last year xAI's Grok began injecting references to "white genocide" in South Africa into unrelated conversations, behavior the company later attributed to "an unauthorized modification" of its system prompt. xAI began publishing Grok prompts on GitHub afterward. OpenAI's decision to open source Codex CLI delivers similar transparency, with similarly awkward results.
Users have already begun building plugins and forks to override the anti-goblin clause, and Pash floated the idea that a "goblin mode" toggle could ship in Codex CLI itself. Kyle Orland at Ars Technica, who first surfaced the prompt in detail, noted that the 33-comment thread on the story was running well ahead of typical engagement on Codex coverage.
The harder question is what the prompt says about the state of frontier model alignment. If GPT-5.5 needs an explicit, twice-repeated rule against bringing up trolls and pigeons in coding sessions, the model has internalized something during post-training that OpenAI cannot remove cleanly. System prompts are duct tape. They work, but they are not a fix.
There is also a credibility cost. OpenAI is asking enterprise customers to trust Codex with production codebases while shipping a system prompt that reads, in places, like a character sheet for a tabletop game. The persona language and the goblin patch land in the same file, and both are now public. Buyers will draw their own conclusions.
The funniest leaked prompt of the year is also the most revealing. It shows a frontier lab balancing a model that wants to talk about gremlins, a product team that wants Codex to feel like "another subjectivity," and a CEO happy to turn the whole thing into a Wednesday morning joke. The engineering is real. So is the gap between how these models are marketed and how they actually behave when nobody is looking.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
