AI Worldbuilding Comparison for Living Campaigns

ChatGPT, Gemini, Claude and QuestPortal are more useful when they stop trying to do the same job

A village burns at the edge of a contested forest.

The players saved its children three sessions ago, exposed its corrupt reeve last month, and quietly armed its miller before that.

When they return, the ruins should not merely look dramatic.

They should remember why they are burning.

That is where an AI worldbuilding comparison becomes more interesting than asking which model writes the prettiest lore.

Give four AI tools the same sentence — the village has burned; tell me what happens next — and all four may produce perfectly usable ideas.

That is not really the problem.

The problem is that you only have one campaign.

Its miller already has a history. Its church already has relationships. Somebody benefited when the reeve fell. Somebody lost money. Someone remembers that the players armed civilians. Someone else remembers only that they saved children.

A living campaign does not need four machines inventing four plausible villages.

It needs several tools contributing to the same one.

That changes the comparison completely.

In our Mythveil workflow, ChatGPT, Gemini, Claude and QuestPortal are not competitors for a single chair. They have different jobs.

ChatGPT works as the brain and managing editor.

Gemini works as the transdisciplinary researcher.

Claude works particularly well as the storyteller and prose specialist.

QuestPortal works close to the table as an assistant and visualiser.

Around them sits Mythveil as the method and Rexborn as the architecture that keeps the world coherent while all those specialists work upon it.

That division is not an objective league table. All four tools overlap, and their capabilities keep changing.

It is a workflow.

And for long campaigns, workflow matters more than winning a benchmark.

The useful comparison is not “Which AI is best?”

A hammer and a good kitchen knife can both damage a tomato.

That does not make the comparison terribly useful.

AI worldbuilding often suffers from the same problem.

One model is asked to research history, generate factions, create NPC dialogue, decide consequences, summarise canon, write the next session and produce a map. Then another tool receives the same enormous request and the results are compared.

Of course one will appear better at parts of it.

The more useful question is:

> What job are you asking the AI to do?

For a living campaign, we need several different kinds of work.

We need knowledge.

We need continuity.

We need interpretation.

We need language.

We need visual and table-facing material.

And, once the campaign begins moving, we need some way of following what factions and people actually do between one player action and the next.

Those are related jobs.

They are not the same job.

ChatGPT: the brain and managing editor

In our current Mythveil workflow, ChatGPT sits closest to the centre.

Not because it should invent everything.

Quite the opposite.

Its value increases when it has less need to invent.

With our original material, Rexborn documents, campaign records and connected project sources available, its job is to bring different pieces of knowledge into one working picture. ChatGPT can work with connected Google Drive material where that connection and its permissions are available, which makes this kind of source-centred workflow practical rather than theoretical.

That makes the role closer to a managing editor than an oracle.

The editor asks:

What do we already know?

Which document is authoritative?

Does this new research contradict established canon?

Which contradiction is real, and which is merely two characters knowing different things?

What question should we research next?

Which faction does this event actually affect?

Which NPC is in a position to respond?

What happened three months ago that suddenly matters again?

That is also why ChatGPT can act as the hub between the other tools.

We can identify a research question here, formulate a precise research brief for Gemini, receive the material back, compare it with the existing world, and decide what deserves integration.

We can take that accepted material and prepare a focused narrative brief for Claude.

We can take the updated physical state of a place and prepare concrete visual instructions for QuestPortal.

And we can use the same accumulated world model for the Rexborn simulation beneath the campaign.

The weakness is obvious.

Give the brain poor memory and vague source authority and it becomes very good at connecting things that perhaps should never have met.

Fluent synthesis is still synthesis.

If the source hierarchy is weak, plausible bridges can quietly become invented history.

> The central AI should not decide what is canon because it sounds right. It should know what the project has accepted as canon and work from there.

Four clever assistants without that distinction can produce four wonderfully convincing settings.

Unfortunately, you only needed one.

Gemini: the transdisciplinary researcher

Gemini occupies a different chair in our workflow.

It goes looking.

Gemini Deep Research currently uses Google Search as a standard research source and can incorporate additional sources such as Drive when configured to do so. That makes it particularly useful when a campaign question needs breadth before it needs prose.

But Mythveil research is deliberately broader than academic literature or scientific papers.

A living world may need history, archaeology, anthropology, ethnology, religious studies, philosophy, psychology, sociology, economics, law, geography, ecology, climate, architecture, medicine, technology or military history.

Then it may need folklore.

Or oral tradition.

Or saints’ legends.

Or urban myths.

Or occult history.

Or pulp fiction.

Or film.

Or comics.

Or games.

Or roleplaying culture.

Or a dreadful local joke that tells us more about how people see their neighbours than a respectable administrative document ever could.

The campaign decides the question.

The question decides which disciplines and traditions might contain part of the answer.

> Research does not ask only what happened. It asks how people understood it, feared it, exaggerated it, misremembered it, laughed about it and turned it into stories.

Consider our burned village.

A shallow research request might ask:

What was a medieval village like?

That will produce material.

A useful Mythveil research brief asks more specific questions.

How important could a mill be to local survival?

Who might control milling rights?

What did forest access mean economically?

How were burned homes, displaced families and local relief handled in the relevant place and period?

What authority could clergy exercise after a disaster?

How could a dead or disgraced reeve’s obligations survive him?

What kinds of communal memory might turn a fire into blame, martyrdom, divine punishment or local legend?

How quickly could news realistically move to a nearby town?

Now history may meet economics, law, religion, environmental studies, folklore and communication history.

That is transdisciplinary research.

It is not licence to throw everything into the soup.

Breadth is not verification.

A search result is not canon.

A striking anecdote is not representative merely because it is striking.

AI can widen the field quickly. Sources still have to be checked, weighed and understood.

> Research breadth is not research reliability.

Gemini expands the field of knowledge.

The central editorial process decides what belongs in this world.

Claude: give the world a voice

Then there is the problem of language.

A world model can be magnificent and still sound like minutes from a committee meeting.

At some point, somebody has to speak.

This is where we particularly like Claude.

Not as the machine that decides what happened.

As the storyteller who receives a well-defined reality and gives it voice.

Claude’s Projects can work from uploaded project knowledge and instructions, which makes it possible to ground its writing in a supplied body of material rather than beginning cold every time. Anthropic also explicitly positions Claude for creative workflows.

Give Claude:

the speaker,

the speaker’s biography,

what that person actually knows,

what they wrongly believe,

what they want,

who will read the text,

the historical and cultural setting,

the emotional state,

and the facts that must not change.

Then ask for the letter.

Or the sermon.

Or the diary entry.

Or the testimony.

Or the old song.

Or the official version of events written by someone whose continued employment depends upon the official version being terribly reassuring.

That is a different task from determining canon.

For the burned village, perhaps Rexborn tells us that the church has successfully established a moral interpretation of the disaster.

Claude can turn that into a sermon without needing authority to decide whether the interpretation is true.

Perhaps a surviving child writes to one of the player characters.

Perhaps the reeve’s widow dictates a statement accusing the party.

Perhaps the miller gives an account that is blunt, frightened and rather less heroic than the version already circulating in nearby taverns.

The world has already established the conditions.

Claude makes those conditions human.

That distinction protects us from a common problem with AI worldbuilding: beautiful prose can make uncertain information feel settled.

A lie can sound wonderful.

So can a misunderstanding.

That is fine, provided the world knew it was a lie or misunderstanding before we asked somebody to write it beautifully.

> Do not ask the storyteller to decide what the world is. Give the storyteller a world worth speaking for.

QuestPortal: bring the world to the table

Then the players arrive.

They do not need to see the research notes.

They do not need a diagram of the simulation.

They need the burned village.

QuestPortal’s current GM Assistant can work with campaign notes and characters, and it can generate NPC portraits, item art, scene imagery and battlemaps.

That makes it useful in a different direction.

It is close to the table.

But the same rule applies:

feed it properly.

“Make a burned fantasy village” gives us a burned fantasy village.

Probably a perfectly respectable one.

It is not necessarily ours.

Instead:

The mill still stands.

The waterwheel is damaged but intact.

The reeve’s house was looted before it burned.

The chapel roof collapsed.

The well square survived.

Three refugee shelters have been built beside the eastern field.

The new palisade uses timber cut illegally from the disputed forest.

Someone has painted the old reeve’s crest on the remaining wall and crossed it out in black pitch.

Now we are not asking for generic atmosphere.

We are visualising world state.

The same applies to portraits.

Do not ask merely for “a stern priest”.

We already know who this priest is.

Age, dress, period, work, status, fatigue, background, perhaps the scar on his hand and the fact that he has spent four nights helping families bury their dead.

Specificity is what turns generated imagery into continuity.

> QuestPortal can show us the ruins. The world model tells it which ruins must exist.

Rexborn is not the fifth AI

This is the most important distinction in the whole comparison.

Rexborn is not another chatbot.

It is the architecture around the work.

Mythveil provides the research and narrative-design method.

Rexborn keeps track of how the world can respond.

The AI systems are specialists working within that architecture.

That means the world needs more than lore.

It needs actors.

Factions.

NPCs.

Relationships.

Resources.

Knowledge.

False knowledge.

Influence.

History.

And a current world state.

The burned village is therefore not stored as:

Village: destroyed.

That tells us almost nothing.

We need to know who lived there.

Who lost what.

Who now possesses an opportunity.

Who is blamed.

Who is believed.

Who has the ability to do something next.

Build the faction before asking AI to move it

Suppose the reeve was corrupt because he had been protecting the commercial interests of a river guild.

That guild is not simply:

River Guild — hostile.

What does it want?

Secure trade.

Stable tolls.

Access to timber.

Political credibility.

Perhaps it fears the collapse of local authority because informal rule is bad for contracts.

Perhaps another guild wants the opposite because disorder gives smugglers an opening.

Perhaps the church wants relief for the displaced but also fears the spread of armed civilian groups.

Perhaps the miller now commands genuine local loyalty because he organised the defence during the fire.

Perhaps the forest community believes the villagers caused the disaster by cutting disputed timber.

Those are factions.

But factions still do not act.

People do.

The guild has a factor who spent twenty years negotiating contracts and sees every crisis as leverage.

The church has a priest whose brother died in an earlier riot.

The miller has a daughter among the children the players saved.

The forest community has an elder who remembers a previous settlement burning forty years ago.

Now the simulation has something worth processing.

> You cannot simulate a faction you have not understood.

And:

> The institution gives capacity. The person decides what to attempt with it.

This is where AI becomes much more useful.

Instead of asking:

What would a merchant guild do?

we can ask:

Given this guild’s interests, its current losses, its relationships, what it believes about the fire, and the personality of the person currently leading its response, what would it actually try next?

That is no longer random worldbuilding.

It is consequence.

Influence gives the simulation hands

Then Rexborn asks whether these intentions can actually move the world.

Influence is not the same as reputation.

It describes the channels through which a person, faction or institution can reach society and make things happen.

A priest may have Religious Influence.

A guild may possess Trade, Political or Financial Influence.

The miller may suddenly possess substantial local Social Influence even though he has no formal office whatsoever.

A player character may have Police, Church, Street, Academic or other forms of Influence depending upon the setting.

For the human Game Master, we keep this deliberately simple.

For indirect information:

> Did they hear about it?

Influence roll.

For meaningful action:

> Can they do something effective with it?

Influence roll.

If someone actively tries to stop them:

> Opposed Influence.

Hear it.

Act on it.

Done.

That is enough at a normal table.

A simulation assistant can use the same architecture at greater depth.

It can follow several factions.

Several versions of the same event.

Different knowledge states.

Different NPC interpretations.

Actions.

Counteractions.

Failed actions.

Delayed reactions.

New opportunities created by somebody else’s success.

The theory does not need to change.

The resolution does.

> AI does not need a different theory of the world. It allows the same theory to run at greater resolution.

Bardic Resonance: the AI should not give everyone the same truth

Now something interesting happens.

The village burned once.

But the event immediately begins producing several stories.

The children remember that the player characters saved them months ago.

The river guild says the characters destabilised the district when they exposed the reeve.

The miller says they were the only outsiders willing to help villagers defend themselves.

The church says the spread of weapons made an already dangerous conflict worse.

The forest community claims the fire followed the violation of old boundaries.

A nearby noble may not care which version is true. She merely notices that there is no longer an effective local authority.

This is Bardic Resonance.

> An event happens once. Meaning keeps happening.

The simulation should therefore not ask merely:

Who knows that the village burned?

It asks:

Who received which version?

How did it reach them?

What does that version mean to them?

What can they do with it?

A false version can produce a true consequence.

An accurate version can disappear because nobody influential hears it.

A story begun by someone powerless can become dangerous once an institution finds it useful.

Controlling the story is itself a form of Influence.

Now we are simulating a society rather than distributing identical information packets to every NPC.

Bardic News brings the simulation home

There is still a final problem.

A perfect simulation hidden in a database is not much of a campaign.

The players need to experience the changed world.

That is Bardic News.

The results come back in forms appropriate to the setting.

A letter.

A sermon.

A song.

A proclamation.

A newspaper.

A police report.

A rumour.

A changed price.

A closed gate.

An invitation that no longer arrives.

An NPC who suddenly looks away.

This is another place where our specialised AI workflow becomes useful.

The central system knows which consequences need to become perceptible.

Claude can give a particular consequence a convincing voice.

QuestPortal can make the changed place visible.

Gemini may have provided the historical or cultural research that tells us what kind of medium belongs there in the first place.

Rexborn tells us why the news exists.

> Simulation changes the world. Bardic News makes that change playable.

Then the players respond.

And their response goes back into the world.

The burned village, one more time

So let us return to where we began.

The players arrive.

QuestPortal shows them the new village state rather than a generic ruined settlement.

The mill is standing.

The chapel is not.

A rough palisade surrounds the eastern field.

The old reeve’s house has become a communal store.

That was not in the adventure three sessions ago.

It happened because the world moved.

A child recognises the party and runs towards them.

His father pulls him back.

On the remaining wall is a hand-painted notice:

THEY ARMED THE MILLER. THEN THEY LEFT.

Someone underneath has written:

THEY SAVED OUR CHILDREN.

Both statements have survived.

Neither tells the whole story.

The priest will speak to them.

The river guild will not.

The miller now has twenty armed villagers listening when he talks.

A neighbouring lord has offered protection.

The forest community has closed the northern path.

Nobody at the table receives a summary saying:

World state changed by 17 per cent.

They receive a world.

And because that world is built from what they did before, their next decision already carries history inside it.

This is where the four-tool comparison finally matters.

Gemini helped us understand the world beyond our existing knowledge.

ChatGPT brought that research together with canon, factions, NPCs, Influence and the existing campaign.

Claude gave selected voices inside that world something worth saying.

QuestPortal helped make the changed place visible at the table.

Mythveil gave the work its research and narrative method.

Rexborn kept the consequences connected.

No single tool had to be everything.

That was the point.

You do not need four AIs to run a good campaign

None of this means a Game Master needs a small digital corporation behind the screen.

One AI can be enough.

No AI can be enough.

A notebook can be enough.

The larger workflow becomes useful when the campaign itself becomes larger: more history, more factions, more NPC relationships, more research, more simultaneous consequences and more years of play.

What matters is understanding the job.

Use a researcher when you need to know more.

Use an editor when too much knowledge needs to become one coherent world.

Use a storyteller when established reality needs a human voice.

Use a table assistant when that reality needs to become visible and playable.

Use simulation when the world contains enough autonomous actors that simply deciding everything by authorial instinct would begin to flatten them.

And keep a human responsible for what the project is actually trying to become.

AI should increase the amount of world we can understand and maintain.

It should not make us care less about which world we are making.

The real AI worldbuilding comparison

So which is best for worldbuilding?

ChatGPT?

Gemini?

Claude?

QuestPortal?

Wrong question.

The better question is:

> Which part of worldbuilding are you trying to solve?

In our workflow, Gemini expands the field of knowledge.

ChatGPT turns that field, existing canon and campaign history into coherent architecture.

Claude gives selected parts of that architecture a voice.

QuestPortal helps make it visible and usable at the table.

Mythveil determines how we research and design.

Rexborn allows the world to keep moving after the scene ends.

Then players disturb it again.

The world responds again.

Stories travel.

Influence collides.

Bardic News comes back.

And slowly, almost without anybody sitting down to write a grand history of the campaign, yesterday’s choices become today’s circumstances.

That is the standard worth comparing AI worldbuilding against.

Not how much lore it can produce.

Not how quickly it can name twenty taverns.

Not which model writes the nicest paragraph about moonlight on a ruined tower.

Ask something harder.

> Can this workflow remember enough, understand enough and simulate enough for yesterday’s play to become today’s world?

Because a village can be generated in seconds.

Making it remember why it burned is the interesting part.

MYTHVEIL

We create worlds people want to live in.

Stories that remember. Worlds that change.

Read: https://mythveilchronicles.com/ Play: https://startplaying.games/gm/dunchan Watch: https://www.youtube.com/@MythveilChronicles Support: https://www.patreon.com/MythveilChronicles Collaborate: https://www.linkedin.com/in/francisco-jos%C3%A9-z-b05aab371/