Simon Willison Maps the Hidden Tooling of OpenAI's Codex Environment
A new comprehensive reference exposes the internal skill files and tool interfaces powering professional LLM workflows.
Developer Simon Willison has published a detailed inventory of the tools and skills available within the 'Codex' environment, the system powering professional versions of ChatGPT. The project provides a rare window into the structured framework OpenAI uses to enable agentic behavior in its advanced LLM interfaces.
The Codex Tool Reference catalogs 232 tool interfaces and 44 complete main skill files. According to the reference site, the system distinguishes between two primary components: tools, which are defined as callable endpoints, and skills, which are reusable instruction packages stored as SKILL.md files that guide the AI on how to utilize those tools. The site provides verbatim copies of these SKILL.md source files alongside exposed TypeScript declarations for the tools.
The Architecture of AI Skills
This project arrives as the ecosystem for LLM 'skills' and 'tools' continues to evolve. Simon Willison, known for documenting and reverse-engineering internal AI capabilities, created the reference to make these hidden systems transparent to the developer community. The release sparked significant discussion on Hacker News, where users debated the underlying architecture of LLM skills and the proliferation of AI-generated 'slop' sites.
Why Transparency Matters
By exposing the internal SKILL.md files, the reference reveals exactly how OpenAI structures the instructions that guide an AI to interact with external tools. For developers and researchers, this is critical for understanding prompt engineering at scale. It moves the conversation from trial-and-error prompting to a structural understanding of how agentic AI is bounded and directed by system-level instructions.
What to Watch
As LLMs move toward more autonomous agentic workflows, the distinction between a 'tool' and a 'skill' becomes a blueprint for how future AI assistants will be built. While the reference provides a snapshot of the current Codex environment, it remains to be seen how OpenAI will evolve these declarations as more complex, multi-step tool chains are integrated into the professional ChatGPT experience.