Asciidia: A Game Where the LLM Generates Every Object on the Fly

A dragon summoned from nothing that you can ride. A factory that produces other objects. A sword with properties the system infers from its name. All generated in real time by an LLM. That is, in essence, Asciidia, a game experiment published this week on Hacker News that proposes a premise as simple to state as it is complex to execute: conjure anything and the game will do its best to give it form, behaviour and coherent properties.

The project arrives self-described as a demo and experiment, without pretensions of a finished product, which is precisely what makes it interesting to analyse.

What Asciidia Actually Does

The interface is ASCII (hence the name), placing the game in a retro terminal aesthetic that is far from accidental: when visual rendering is not the strong point, the generation of logic and behaviour must carry all the weight. The player types or selects an object (a dragon, a spell, a weapon, a factory) and the system materializes it with the properties that would "make sense" for it to have.

That means the LLM does not just generate descriptive text, but must infer mechanics: if you create a dragon, you can fly on it; if you create a factory, it produces things; if you create a fire sword, it probably deals fire damage. The internal consistency between objects and their interactions is where the real technical difficulty lies, and the author openly acknowledges that "the underlying systems are complex" despite the seemingly simple appearance of the interface.

Why This Approach Differs from Classic Procedural Generation

Procedural generation has been around games for decades: Roguelikes, Minecraft, No Man's Sky. The difference here is that there is no predefined set of rules that the algorithm combines; it is the LLM that reasons about what properties and behaviours are plausible for each object in each context. That opens up a much wider space of possibilities, and one that is much more unpredictable.

The classic problem with this approach is consistency: an LLM might decide that a dragon has wings in one invocation and not in the next, or that a factory produces gold in one context and coal in another without apparent reason. Managing that variability so that the game remains playable rather than chaotic is the engineering challenge that projects like Asciidia need to solve.

We don't know which specific model Asciidia uses (the author doesn't specify in the published description), nor how it manages state between sessions or the consistency of already-created objects. These are questions left open in this initial presentation.

Who This Is Relevant For

Asciidia interests very specific profiles. First, indie game developers who have been eying LLM integration for a while without knowing quite where to start: this project is a proof-of-concept that natural language mechanics generation is viable, even if in experimental form. Second, those working with Claude Code or MCP servers oriented toward content generation: the architecture of such a system (inferring properties, maintaining state, resolving interactions) is a direct use case for agent or specialized sub-agent pipelines.

And third, quite simply, anyone who enjoys seeing how far an honest experiment can go when published without fanfare on Hacker News on a Thursday morning.

Project Status

With a handful of upvotes on HN and zero comments at the time of publication, Asciidia has not yet generated measurable public conversation. That says nothing about its technical quality (many solid experiments go unnoticed in feed noise), but it is worth keeping in mind: we are looking at an early demo, not a product with extensive community or documentation.

The author frames it as such: they wanted to share it and see what people think. It is an invitation to exploration, not a launch.

---

From our perspective, it is an honest experiment that tackles a genuinely difficult problem: making an LLM not just describe objects but make them function with coherence inside a game system. That it is in demo form is to be expected; that someone built it and published it without waiting for perfection is, at least, a concrete starting point for conversation.