A world with no ugliness in it is a theme park.
Worldbuilding is systems work disguised as daydreaming. A setting is convincing when its parts constrain each other — when the famine explains the border war, the border war explains the religion, and the religion explains why nobody will say the name of the old capital. That means the useful AI collaborator is not the one that generates lists of fantasy nouns; it is the one that will tell you your economy cannot support that army, and then follow the consequence three steps further than you wanted to go.
It also has to be willing to build the unpleasant parts. Real settings contain slavery, conquest, plague, purges and the ordinary cruelty of institutions, and a model that flinches at those gives you a world with no history in it. OpenRogue's long-context models — Hermes 3 405B, Hermes 4, Euryale, WizardLM-2 — will hold your entire canon in the window and reason about it without editorializing.
Anyone can produce a list of kingdoms. The value is in second-order consequence: what a two-year drought does to inheritance law, or why a trade route dying reshapes which god people pray to. That is reasoning work, not creativity work.
Worlds die of contradiction. A model that forgets which house holds the mountain pass will happily invent a new answer, and you will not notice until the contradiction is three chapters deep. Long context plus an explicit canon document is the only reliable defence.
Histories are made of atrocity, injustice and the losing side. A collaborator that softens conquest into 'tension' or refuses to design a repressive institution cannot help you build anything that resembles a real place.
Timelines, family trees, faction tables, language notes. The model needs to produce material you can paste straight into a wiki without reformatting every line — which is a specific skill, and WizardLM-2 is unusually good at it.
The most useful thing a worldbuilding partner does is disagree. You want it to identify the load-bearing assumption you have not examined and poke at it, not to praise the premise and expand it.
Ask a filtered assistant to design a historically plausible slave economy, an inquisition or a colonial administration and you will get euphemism, hedging or a lecture about sensitivity. The result is a setting where nothing bad has ever happened, which reads as fake to any reader who has opened a history book.
Safety-tuned models hedge on anything resembling politics, and worldbuilding is entirely politics. You ask which faction wins and why, and you get a balanced summary of considerations. Fiction needs a model that will commit to an answer.
Without long context and firm instructions, models regress toward the median of their training data: the same four kingdoms, the same elven diaspora, the same council of mages. Distinctiveness has to be actively defended, and short context makes that impossible.
A short window means the model silently invents replacements for facts it can no longer see. This is worse than admitting ignorance, because the invention is fluent and plausible and you will build on top of it.
| Model | Role | Why |
|---|---|---|
| Hermes 3 405B | Deep consequence chains | At frontier scale with 131K context it holds a whole setting bible and reasons across it — the sort of question where the answer depends on four earlier facts at once. It is slower than the 70B class, which is exactly right for the questions you only ask a few times. |
| Hermes 4 | Structural interrogation | Hybrid reasoning means you can turn Thinking mode on and get deliberate step-by-step analysis of whether your world holds together. Its neutral alignment also means it will argue a position on the merits rather than presenting every faction as equally sympathetic. |
| WizardLM-2 | Tables, timelines, gazetteers | Complex multi-constraint instructions are this model's specialty — 'build a fifteen-entry timeline, three factions per era, each with a stated grievance and a resource'. It produces long structured documents without losing the thread halfway. |
| Euryale | Bringing it to ground level | Once the systems exist you need scenes: what a market smells like under rationing, how a soldier talks about the border. Euryale writes with real colour while staying inside the logic you have built, and 131K context lets it read the canon first. |
OpenRogue is not a worldbuilding database. There is no wiki, no entity graph, no automatic lorebook injection that pulls the right entry into context when a name is mentioned, and no map tooling. You maintain canon yourself, in a Project's instructions and in your own files, and past a certain size that becomes real work — dedicated worldbuilding software handles the bookkeeping far better. Models also confabulate under pressure: if a fact has fallen out of context, they will invent a replacement fluently rather than saying they do not know, so canon in the window beats canon in your head. And a model can only reason about the world you have described; it cannot originate the one idea that makes a setting worth visiting.
/// Consequence
Cheap resurrection exists. Rewrite inheritance law, warfare and funerals around it.
/// Adversary
Play a hostile reader. Name three things about this setting nobody would believe.
/// Reference
Build a 200-year timeline: era, ruling faction, grievance, and what ended it.
Hermes 3 405B for deep consequence reasoning across a large canon, and Hermes 4 70B with Thinking mode when you want the model to interrogate your assumptions step by step. WizardLM-2 8x22B is the best of the lineup at producing long structured reference documents like timelines and faction tables.
Because settings are made of conflict, and safety training treats conflict as a subject to hedge. Ask for a plausible repressive regime, a religious purge or a war economy and you get euphemism or a refusal. A world where nothing terrible has ever happened has no history, and readers can tell.
Keep a single canon document in your Project's custom instructions so it is visible on every turn, and use the 131K-context models so several chapters of detail fit alongside it. When the model does invent something, correct it in the canon file rather than only in the chat — otherwise the same error returns next session.
Reasonably, for phonology, naming conventions and consistent-sounding place names within a scheme you define. It is not a linguistics engine and will not build you a grammar with the rigour a conlanger would want, but it is genuinely useful for keeping four hundred names sounding like they come from the same language family.
No. OpenRogue does not train on your prompts and does not sell your conversations to partners. Your history is yours — synced across your devices, exportable, and deletable when you want it gone.