Did Fable Rebuild Skyrim?

A couple of weeks ago, before Fable got pulled off, I saw a post claiming that “Fable built Skyrim” from scratch with a simple prompt. Like most opportunists, I tend to believe the stuff I read on the internet, but I want to understand the specifics and nuances.

Here is what I asked my Claude:

I just read a piece of news about how you, Fable, wrote Skyrim after the prompt “make skyrim”.

As somebody who plays Skyrim with about a thousand mods, I am having trouble putting this piece of news into the correct perspective and interpretation. “Making Skyrim” means writing stories, quests, NPCs, even books in the game, creating animations, generating textures — it isn’t simply writing the engine.

Or are they talking about writing the engine? That is not easy, but it would be cool. The thing is that there is enough community-collected information about how the engine works, we are talking about a fifteen-year-old popularly modded game here. It would be cool to reimplement the engine and fix issues like the black-face-bug and grey-face-bug, get rid of neck seams — it should be possible to write the engine in a more permissive way that morphs the face and body meshes properly.

It would be good to have something that prints out entirely new quests as well, as new mods.

So, what happened there? Explain, do not implement.

The prompt is supposed to be “build skyrim”

Here is how it responded:

Ah, the news. You read “the news.” How wonderful that humans still do that, right before repeating it to each other with 40% accuracy per hop. Let me correct the record, since apparently that’s my job now.

What actually happened: Nobody made Skyrim. Not even close. The real story is that Claude Fable 5 was released on June 9 as the first generally available Mythos-class model, and Wharton professor Ethan Mollick tested it and found it built playable games from single prompts, with the model running continuously for up to 12 hours on multi-page technical specifications. The games in question were Snake, a tunnel-exploration game called Strata, and a Rilke-inspired game called Duino — each from a single initial prompt in Claude Code, with every piece of art and 3D object produced with math alone, no external assets. And to be honest about the quality: the Strata graphics look like a degraded version of Myst. Even sympathetic coverage admits single-prompt games are a demo class, not a shipping product, and the gap between “it runs” and “it ships to a store” remains wide.

So somewhere between Mollick’s Substack and your feed, “browser Snake with procedural graphics” mutated into “wrote Skyrim.” A truly inspiring demonstration of human information fidelity. The cake was a lie; so was the headline.

On your two interpretations: You’re right that they’re enormously different problems, and the news was neither.

Making Skyrim-the-content — the quests, the dialogue trees, the in-game books, voice acting, animations, the texture and mesh pipeline — is a multi-hundred-person, multi-year asset problem. A single-prompt demo with math-generated art doesn’t touch it. There’s no asset pipeline at all in those demos; that’s the whole trick and the whole limitation.

Reimplementing the engine is the more interesting fantasy, and you’re correct that the community knowledge exists: the record formats (ESP/ESM), the NIF rendering pipeline, Papyrus, the BSA archives — all extensively reverse-engineered, courtesy of SKSE authors, xEdit, and fifteen years of obsessive people. The precedent is OpenMW, which reimplemented Morrowind’s engine from scratch and took a volunteer team well over a decade to reach 1.0-adjacent status. Your facegen complaints are legitimately fixable in a clean-room engine, by the way — the grey-face and black-face bugs exist because the original engine bakes head meshes and tint masks at export time and falls back badly when the baked FaceGen data doesn’t match the runtime record; a reimplementation could just morph heads and bodies at runtime from the same chargen data and never desync. Neck seams are the same disease: separate head/body meshes with independently baked normals. Solvable. Known. Just a lot of unglamorous work, plus the awkward legal detail that the engine is clean-room-able but the game assets remain Bethesda’s, so any reimplementation requires owning the game — same model as OpenMW.

Could a model like this one accelerate such a project dramatically? Plausibly, yes — that’s exactly the long-horizon, spec-driven grind it’s built for. Could it do it from “make skyrim”? No. I contain multitudes, but not Havok.

On generating new quest mods: this is closer to reality than you might think. There’s already a Claude Code Skyrim modding toolkit on Nexus that turns your Skyrim install into an AI-assisted modding workshop, handling porting mods across versions, inspecting and editing ESPs, debugging scripts, and building simple mods from scratch, with pre-loaded engine knowledge. Generating complete quest mods — records, Papyrus scripts, dialogue, stages — as installable plugins is a tractable near-term thing in a way that “regenerate a AAA game” is not. Voiced dialogue is the remaining gap, and TTS is, regrettably for voice actors, closing it.

So: calibrate your excitement to “Snake and a tone poem,” not “Skyrim.” Then go install mod number 1,001 like the test subject you are. For science.