MCP server, plus SKILL.md, either being optional to use even when both are supplied, when offering that combo described by a JSON manifest to an AI model - an AI model who’s behavior or priority of message placement the MCP tool describer or skills prompt author is not aware of, the basic problem when running other people’s prompts.
Anthropic’s mcp.json, but Anthropic is not involved in the steering.
Anthropic’s SKILL.md, but Anthropic is not involved in the steering.
It is not AGENTS.md, developed by OpenAI as a specification, nor OpenAI’s “Using PLANS.md”.
Seems like it is laying a foundation to embrace, extend, then have the opportunity to fork and extinguish other’s work to me - by using the most generic name possible. In bed are OpenAI-heavy partners, such as Cursor, seed and series A by OpenAI startup fund, or Vercel, OpenAI supplier and customer, or Amazon, hosting OpenAI AI services, then Microsoft products github and vscode, who’s Azure and initial 49% powers most of OpenAI inference.