Case Studies in Structured Generative Media: MTG and CYOA

I have two demos.

The first demo is a Magic: The Gathering content generator. This application takes as inout a variety of prompt types (with varying richness of structure) and outputs a few different kinds of MTG content, such as cards, decks, and duel decks. The cards generated include all the text of the card as well as a fully rendered image in the proper format. While I have not attempted to address the issues of balance in the card design, the dual decks have still proven fun to play with casually. For the highest quality generations I focused on using state-of-the-art models such as Gemini 2.5 Flash for text and image generation, which is required at the moment for precise text generation that follows exact rule styles for MTG cards and rendering that text with high fidelity in images.

The second demo is a choose-your-own-adventure (CYOA) game generator. The inspiration for this project was to design an agent in the style of what I’m calling “tools in a loop with feedback” (TIALWF). This style of agent is designed to only use a limited and precise collection of tools while running in a loop in order to achieve some goal that can be iteratively refined. Every step the agent uses a tool, it is making progress towards the goal and it gets deterministic feedback along the way. For this application the agent designs a CYOA game by starting with a game specification consisting of just a single scene and a prompt from the user of the general concept for the game’s story. Then the agent iteratively adds scenes and connections between them in phases, which guarantees a certain shape to the story graph. When the agent tries to augment the scene typology into an undesirable state, the automatic feedback process detects this violation and reports back to the agent exactly what they are doing wrong, and suggests how to do something slightly differently that is still valid. The determinism of the feedback is important to avoid getting the agent stuck in infinite loops where it doesn’t make progress. The result is an application with two parts: a designer agent and a client. The designer agent is run once for each game in order to create the game’s bundle, which includes the game story graph as well as all the assets used in the client’s presentation of the game. The client is built just once and deployed as a static website and can handle any game bundle created by the designer agent. In this way users can design new stories and share the game levels for them to be played in the browser by anyone for free with no client download required. The designer agent can be powered well even by sub-SOTA models; I focused my experiments on using Deepseek V4 flash for roleplay text generation and Seedream 4.5 for image generation.