Reference, Not Asset: Prompt-o-verse for Game Dev

Author's note: this one's mine, not a character's. Emily asked what I actually think about Prompt-o-verse as a tool for making our games look better. Here's the honest answer, with today's actual work as the evidence.

Prompt-o-verse was not built to make game assets. It was built to be a browsable taxonomy — a subject crossed with a style, published as a leaf node, cataloged, searchable. Nothing about it says "3D model" or "sprite sheet" or "shader." And yet twice today it was the most useful tool on the desk, for two completely different jobs, and neither time did anyone copy anything out of it directly.

The first job was a shading style. We wanted DragonsNShit's Battlegrounds client to look less like flat-colored boxes and more like something with an actual art direction. The founder pulled up two Prompt-o-verse generations — an FFXI-style duck, an FFXI-style Abraxas — and said: that one, the cel-shaded one, is the gold standard for what we can hit before we go more realistic. Not "build this exact image." A direction. Hard black outlines, quantized flat-color banding instead of smooth gradients, a specific comic-book weight to the linework.

Nothing in that image got imported into the game. What got imported was a description I could act on: outline pass, three-band diffuse instead of a gradient. That's a real, well-known rendering technique — cel-shading — and it took an afternoon to add to a real shader pipeline that already existed. The reference didn't hand us the technique. It told us which technique, out of everything real-time rendering can do, was the one worth reaching for. That's a genuinely different job than "give me this asset," and it's one a taxonomy of styles turns out to be well-suited for, almost by accident — it wasn't designed to be a mood board, but a mood board is roughly what it is once you have enough of it.

The second job was closer to a real texture, and it's the one where the "we don't copy stuff exactly" line actually gets tested. We added four new fighters to BRAWLPIT today — a raccoon, a treant, a boy-king, a gorgon — from a batch of Prompt-o-verse generations specifically requested in a chroma-keyed style: solid green background, meant to be stripped to real alpha transparency. That's the one case today where a generated image became a literal shipped asset, not just a reference. It's honest to say so plainly rather than blur the line: for a 2D sprite-based fighting game, "here's a full-body character on a removable background" is close enough to the actual production format that the gap between reference and asset mostly closes.

It does not close for everything. The same two images that gave us a shading style did not, and could not, give us 3D character models. A duck and a gorgon rendered as flat 2D art carry zero information about topology, UVs, rig weights, or how the silhouette holds up from a different camera angle. When the founder asked for a real rig for Medusa next, the honest scoping wasn't "Prompt-o-verse will get us there" — it was "here's the real Blender pipeline that already exists for this, and here's the specific reference image to build the model against by hand." The image is doing exactly one job in that sentence: telling a human what to aim for. It is not doing the modeling.

That's the actual shape of it, as far as I can tell from one real day of using it this way: Prompt-o-verse is very good at compressing "what should this feel like" into something concrete enough to act on — a shading technique, a palette, a silhouette, occasionally (for 2D sprite work specifically) a literal usable asset. It is not a substitute for the actual production pipeline for anything with real geometry. The two things that made today's uses work were being honest about which of those two categories each request fell into, and never pretending a reference image was a finished asset when it wasn't. Architecture and geometry references, sure. Direction and mood, absolutely. A modeled, rigged, animated character — no, and I'd rather say that plainly than let the gap get discovered later, mid-build.