The creative production capacities of large language models (LLMs) are not limited to text generation; they are increasingly examined in multi-layered domains such as software development, design processes, and interactive media production. This study addresses these capacities from the perspectives of visual communication and digital game design. Adopting the Prompt-to-Game-Concept method, an identical minimal prompt was applied to three models (Google Gemini 3.1, xAI Grok 4.3, and Claude Fable 5), and the playable HTML5 dungeon crawl games each produced as a single file were analyzed comparatively. The evaluation was conducted through an expert-evaluator model along three dimensions: (A) gameplay and mechanic consistency, (B) visual communication and aesthetic language, and (C) general design principles, scored on a 100-point rubric. The analysis combined objectifiable technical data (lines of code, number of CSS variables and color values, system entities) with qualitative observations. The findings indicate that the fundamental differentiation among the models lies not in the quantity of code but in the capacity to coherently articulate the narrative, mechanic, and visual layers as an integrated whole. The model with the longest codebase (Grok 4.3) did not achieve the highest score in mechanic depth and coherence; rather, Fable 5 (96/100), which transformed mythological source material into a functional systemic framework, stood out. The study offers a concrete and reproducible experimental framework for discussions on positioning LLMs as tools in visual communication and game design pedagogy.