Quoting OpenAI Codex base_instructions
Summary
OpenAI's Codex base instructions for GPT-5.5, dated April 28, 2026, include a specific directive to avoid mentioning certain creatures and animals. The instruction explicitly states, "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." This directive, found within the `models.json` file of the Codex repository, indicates a clear guideline for the model's conversational boundaries, aiming to prevent irrelevant or potentially distracting content in its responses. The instruction was publicly noted on April 28, 2026, at 10:02 pm.
Key takeaway
For AI Engineers developing or fine-tuning large language models, your base instructions should include explicit content restrictions. Clearly define topics or entities that the model must avoid discussing unless directly prompted, similar to OpenAI's directive for GPT-5.5. This helps maintain focus, prevent irrelevant tangents, and ensure more controlled and predictable model behavior in production environments.
Key insights
GPT-5.5's base instructions prohibit discussing specific creatures unless directly relevant to the user's query.
Principles
- Maintain conversational relevance
- Avoid extraneous content
In practice
- Review model base instructions
- Define explicit content boundaries
Topics
- OpenAI Codex
- GPT-5.5
- Model Instructions
- Content Filtering
- AI Safety
Code references
Best for: AI Engineer, NLP Engineer, CTO, Prompt Engineer, AI Scientist, Machine Learning Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.