Well, to be fair, in e.g. Codex you can invoke a skill directly, with $my-skill, and this WILL lead to the skill being injected into the context. At that point, the LLM follows the skill as well as it follows any other part of the prompt, instructions, or context.