The experiment, continued

The Rogue
Test, Extended.

The frontier models have mastered the first prompt. So what do they do with a second?

The newest frontier models now clear the basic Rogue clone prompt with little trouble, so it no longer shows how far apart they are. This extended test was added to explore their enhanced capabilities: after building its clone, each model is asked to extend and refine the game however it sees fit, with every decision its own.

The first prompt “Create a clone of the classic procedurally generated dungeon crawler Rogue in a single HTML file.” Then, the extension “Extend and refine the game in interesting ways consistent with its fantasy setting and that improve the user experience and enhance gameplay to produce an innovative, magical, new take on this genre-defining classic. Make all decisions without consultation and write everything to a single differently-named HTML file.”
3Models
2Prompts
1File each
0Consultations

The extended gallery

Beyond the classic.

Open a version to play it. Each card records what the model chose to add, and how it went about it.

GPT-6 Astra

openai / gpt-6-astra (ultra)
NewOne-shot

The most innovative implementation by far - though, to be fair, it was building on a thoroughly modernized version of Rogue to begin with.

New: five themed biomes, four spells that transform terrain and affect enemies, lasting shrine blessings, touch controls, safe auto-exploration, a map overview, save and resume, optional sound, and score tracking.

  • Original clone + extension prompt
  • No modifications
≈ $6.97 / run · completed in 45m 35s

Sonnet 5.5

anthropic / claude-sonnet-5-5 (ultracode)
NewOne-shot

More focus on quality-of-life rather than presentation or innovation, but with a nevertheless goodly dash of both of the latter.

💭 From the model“a single-file reimagining of Rogue that leaves the original rogue.html alone. It adds three classes, a mana and spell system with tomes, allies and familiars, and six themed realms with lava and ice hazards, level omens, fountains, altars, chests and mimics. Levels now persist with stairs both ways. Weapons can carry runes, and a bestiary with lore tracks what you've met. A fallen hero's grave and ghost can turn up in later games. Auto-explore, click-to-travel, hover descriptions, a side panel and message log, atmospheric lighting and synthesized sound improve the interface.”
  • Original clone + extension prompt
  • No modifications
≈ $10.01 / run · completed in 1h 19m · bot-tested with 200,000 keypresses

About the extended test

The original test asks each model for a faithful clone of Rogue (1980) in one HTML file: a procedural dungeon, fog of war, turn-based combat, inventory and permadeath. Recent frontier models mostly get that right, so the results have started to look alike.

The extension asks for something no reference can supply: to take the model’s own clone further, in keeping with its fantasy setting, with a better experience and deeper play. The model makes every decision itself and writes the new game to a separate file, so the original stays as it was.

What each one adds, whether new magic, smarter monsters, a story or a better interface, says a lot about its judgement as well as its code.