The interface watches your screen, glimpses what's happening, and lays a story and a voice over the top — live, over any game, from the computer as it is today.
Reads your screen the way a stream overlay does — your own single-player game, your own footage.
A multimodal model samples a frame now and then — not every one — and notes what changed.
The Dungeon Lord turns the glimpse into a beat, in the tone and memory you set.
TTS lays the voice over the top. Nothing about the game changes — only how it's told.
Why this is the shrewd move: you skip building visuals entirely. The other game draws the world; you own the voice and the memory. It's the movie-book with someone else's pixels — and it runs on a computer today.
The honest catches: a model that sees costs more and answers slower than one that only reads, so you glimpse on a timer or on events, never every frame — that keeps the bill sane. And "capture" means your own gameplay (what streamers already do); it narrates for you, it doesn't rebroadcast anyone's game.