On the back end, the spells are implemented primarily through changes to the prompt the LLM receives--spell effects are directly passed as context and spells and effects that alter attractiveness change a latent player/NPC relative challenge rating which determines the instructions to the LLM on how difficult to make the NPC. In practice, Command R+ has limited capacity to follow instructions, particularly in terms of rejecting user commands, so the game is trivial to jailbreak. I have some ideas to resolve this, but they require a more performant model than the current version of Command R+ and I'm waiting for the next Command R+ upgrade or free uncensored LLM API to implement them. For now, the game relies on the honor system--you have to decide not to bullshit the narrator to get the intended experience.