This is probably something that will be fixed with an OpenAI to "some chess AI"-plugin. Then for the current model it's just enough to understand that it's a chess problem.