With AI coding tools such as Codex, there are multiple decisions made during the construction of the prompt that is ultimately submitted to the model. As a result, what actually ends up in the submitted prompt is largely guesswork today.
For example, if an AGENTS.md file is present along with one or more SKILL.md files, it is unclear which of these are used, and how they are combined when constructing the final prompt.
Equally important is the handling of the context window. Is the full context included, or only a subset? If filtered, what criteria determine what is retained or discarded?
While it is understood that the submitted prompt includes a system prompt, there is no expectation that its contents be revealed. However, it would be useful to know where it is inserted—for instance, whether it is always prepended to the prompt or merged in some other way.
What I am looking for with each prompt submission is something like…
We want to keep the one file that will always be read, AGENTS.md, short and concise. We move everything that is not needed at all times to other files and these could be skills. The skill prompts will only be added to the context window when actually needed by the agent.
