Build the part of the world your next session touches, from what your table has already made true
November 8, 2026

The most useful way to use an AI world builder for a campaign is to have it build the part of the world your next session will touch, starting from what your table has already made true, and to leave the rest blank. Ask an AI for a continent and you'll get one, with a history nobody at the table will hear. Ask for the place the party reaches next session, with five things the players can learn there and five that only you know, and you get something you can run.
Building less also matters because AI is bad at keeping an invented world straight. In a 2024 study, researchers gave AI models whole recent novels and asked them to judge pairs of true and false claims about each one, a task the people who'd read the books found easy. The best model got 56% of the pairs right, where guessing gets 25%. The models did worst on fantasy and science fiction, the books "that contain extensive world-building." Every page of world you generate is another page that you and the AI both have to keep consistent.
Most pages about AI world builders sell a tool that generates the whole world up front. The GMs who write most about prep say to do the opposite, so this guide starts with them: how much world a campaign actually needs, seven ways to use AI that follow from that, and one place built from a single line of notes into something ready for the table.
Less than most GMs expect, and the rest gets made in play. Sly Flourish's advice in Lazy World Building is to "build each part as you need it for the next session," and to "Clear the path in front of the characters as they walk it." The Angry GM's limit for starting a homebrew campaign is stricter: "you come up with no more than five facts for the players and five more for yourself." He means it literally: "And when I say facts, I don't mean paragraphs or pages. I mean sentences."
The Alexandrian puts the minimum before play at "three to six pages," and says what happens after that is the better part of the job. In his fast homebrewing guide, a setting "will naturally expand through the scenarios you run in it. This is, in fact, the most efficient and arguably the best way to build a setting."
A world built that way grows in two places: your prep, and whatever gets said at the table. The second kind needs somewhere to live. The Alexandrian keeps a "Known Facts" section in his campaign status document for details improvised in play, and he describes it plainly: "This is literally just a bullet-pointed list of stuff." That list matters more once an AI is adding to the world, as way 7 below shows.
None of these GMs is selling AI. Sly Flourish says LLMs "offer the illusion of help," and the Alexandrian argues that they "can't handle the fictional continuity of an RPG adventure". Their advice still tells you what to ask an AI for, and what to keep for yourself.
Asked for a setting with nothing to go on, AI gives you the average fantasy setting. In an EN World thread from 2023 on whether ChatGPT can create a campaign setting, one GM summed up its attempt: "It's telling us the place is dangerous and riddled with factions but can't actually give us what makes it unique."
It also writes the wrong kind of material for a game. The Alexandrian's oldest rule is "Don't prep plots, prep situations", and a GM in a 2026 D&D Beyond thread on AI found that AI "seems to want to write stories not situations."
And it loses track. Matthew Neagley's notes on prepping a campaign with ChatGPT at Gnome Stew warn, "Don't expect complete consistency. It can change details from paragraph to paragraph." When a 2025 study had AI rewrite classic stories in a modern setting, plot holes turned up in its versions up to twice as often as in the originals.
Give it your campaign and the first problem mostly goes away. A new one replaces it: it fills every gap. The Angry GM wasn't writing about AI, but his warning fits it: "When things fit perfectly and make perfect sense, when there's never any weird inconsistencies or loose ends dangling around, the world feels like a façade." The example below ran into exactly that.
Ask about the place the party will be next session, not the region around it. The coast, the empire, and the gods can wait until someone at the table goes looking for them. If you know where the party is headed, ask for that place. If you don't, ask for the two or three places they could reach, each in a few sentences, and build out the one they pick.
The same goes for groups. A living campaign needs about three groups with plans, the ones nearest the party, and the rest can stay a name on the map.
Some facts will decide the campaign: who the villain is, what the secret is, whether the missing brother is alive. Decide those yourself and tell the AI. Left open, the AI fills them with its own answers, and they arrive looking just as finished as the parts you meant it to invent. In the example below, it decided how a missing sailor's coat reached the salt cotes, and that was exactly the kind of answer I wanted to make myself.
The same advice works for characters. The guide to making characters with AI has the AI pitch while you decide what's true.
Paste in or point it at the notes the next session depends on: the last few session logs, the people the party is chasing, and your list of known facts. Then ask for the Angry GM's ten sentences: five things the players can learn, and five only you know. A sentence limit keeps the output to what you'll actually use, and it keeps the place small enough to check.
If you want options before you commit, ask for ten short versions of the place, each a different kind of place rather than ten variations on one, and keep the one you like.
Leave blanks on purpose, and say so in the prompt. Dungeon World's GM principles include "Draw maps, leave blanks" and "Leave room for the unknown." The Angry GM's version is a place you've only named: "If the only thing you know about the Valley of Daggers is that it's called the Valley of Daggers… you're as curious as the players to find out what the hell is there."
AI won't leave blanks unless you ask. Tell it what to leave alone ("leave the rest of the coast blank"), and ask it to list what it left undecided, so you know where the gaps are.

History only reaches the players when they find it. Sly Flourish suggests you take a page of lore and "break it down to ten one-sentence bits you can drop into your game wherever you need it." AI does this well. Give it your GM-only facts and ask for one-line clues, each in a different place or from a different person.
For anything the party has to work out, ask for at least three clues that point to it, so missing one doesn't stall the session. The campaign ideas post covers why three.
Players can fill in parts of the world too. Sly's campaign checklist says to "Think up any worldbuilding questions you want to bring to your players," and to "Focus these questions on the things you actually want answers to." Dungeon World makes it a principle: "Ask questions and use the answers."
The AI's job here is small. Give it your list of blanks and ask for one question per character that would fill one of them, like "Who taught your character to sail, and what did they never tell you?" for a blank about the coast. Pick the questions you want answered, then write down what the players say as known facts.
After every session, add what was said at the table to your known-facts list, especially anything improvised. Before you use anything new from the AI, give it the list and ask what contradicts it. Some GMs on EN World already use it this way. One uses it to "point out contradictions in descriptions of the same character."
Then read the list yourself. In the same 2025 study, once the stories got longer and the model had to say where the plot hole was, the best model only slightly beat a guess of "no errors", so the AI's check is a second reader, not the only one.
I set up a campaign for this post in Campaign Arks: Bishop's Lenn, the river trade city from the characters guide, with its three merchant houses, the party, and three sessions written up as if they'd been played. The thread that matters is Magot's. Her brother Hankin went to sea on the Margaret last spring and never wrote. In session 2, a sailor on the quay was wearing Hankin's coat, and said he'd bought it downriver, "at the salt cotes," from a man with a burned hand. In session 3, the party decided to go and find him, leaving at first light. The known-facts list had one line that would matter: below the city, no barge goes past the last sluice after dark.
The salt cotes were one sentence in the notes. First, I asked the way most world-builder pages suggest:
Build out the salt cotes and the country downriver of Bishop's Lenn: its history, who lives there, the factions, and what's going on.
It read the session logs and known facts first, and nothing that came back was generic. It had the tide, the last sluice, and the barge skipper the party met in session 2. It was also about 800 words, in 38 suggested additions to the salt cotes document: a history of the salt works, a debt court with its own officer and tally wall, five named parts of the settlement, a layout with movement rules for the mud, and two new documents for the country downriver and the court. It also wrote a GM-only secret that decided how Hankin's coat got to the salt cotes. It flagged that secret as its own invention, which was fair, but every gap was filled. It had done what I asked, and the ask was too big.
So I dismissed all 38 suggestions, decided the important answer myself (way 2), and asked for a session's worth (ways 3 and 4):
Next session the party reaches the salt cotes at first light, looking for the man with the burned hand. I've already decided a few things: the man is Hankin. He burned his hand when the Margaret was set alight at sea, and he's working the pans under another name. He doesn't want to be found, not even by Magot. Give me only what this session touches: what the party sees as they arrive, five things the players can find out here, five things only I know, and up to three people they'll meet, each with one thing they want. Keep it to about 300 words, leave the rest of the country downriver blank, and don't contradict the session logs or the known facts.
This time it came back at about 300 words, with Hankin working the pans as Wat Sallow. Two of the five things the players can find out, unedited:
- Sallow came up the causeway in the dark, soaked and stinking of smoke, with a bandaged right hand.
- No one at the Cotes will say the Margaret's name.
It left blanks where I hadn't decided anything: "Who set her alight is still undecided," and "Whether Wesenham knows he is alive is undecided." It withdrew its two new documents "so the country downriver stays blank."
Before accepting anything, I asked it to check its suggestions against the session logs and known facts (way 7). It found two things I'd have tripped over at the table. It had given the barge skipper a berth at the salt cotes, but her notes keep her registered boat in the city harbor. And the new night cargo had to go out to sea, since no barge passes the sluice upriver after dark. It also flagged the timing: with the sluice rule, arriving at first light means waiting out the night above the sluice. That sent me back to my own prompt. Session 3 has the party leaving the city at first light, so I'd rather they arrived in the afternoon.
So I made my changes. You can reword a suggestion in Campaign Arks before accepting it, but I asked for these in the chat:

I also asked it to turn the five GM-only facts into one-line clues, each in a different place or from a different person (way 5). Three of them, cut to the clue itself:
- Tilney's cote, eastern row: A tarred sailor's knife under the bunk, marked H.B., with a fresh burn scraped off its handle.
- The rakers' healer: Dressed Tilney's right hand and says it is a fire burn, not a rope or tar burn.
- A raker's child, at night: Heard Tilney cry out in his sleep about hatches and smoke…
That's three clues pointing to the one conclusion the session turns on, that Tilney is Hankin. The knife is its invention, and it said so. I kept it, because Magot will know those initials. Here's what I'd take to the table:
The country downriver is still blank. After the session, whatever gets said at the salt cotes goes on the known-facts list.
Swap in your own campaign. Each one lines up with one of the seven ways.
A general chatbot only knows what you give it, so building the next session means pasting or uploading the logs and the known facts every time. Campaign Arks is a session-prep assistant for GMs, and it keeps a campaign as documents: places, NPCs, factions, secrets, and a log for every session. In the example, the assistant on Pro read the session logs and the known-facts list itself, and its drafts arrived in the salt cotes document as suggestions. You can reword a suggestion or dismiss it, and nothing enters the campaign until you accept it. The free plan gives you one campaign, with an AI brainstorming partner that knows the page you're on.
For how it fits into a week of prep, see the GM session prep cycle.
Prep my next session · Free for one campaign, forever · No credit card
Can an AI world builder remember earlier sessions?
A chatbot only knows what you've given it, and in a long conversation, details from early on get lost or changed. Keep your session logs and known facts outside the chat, and give it the parts the next session depends on each time you ask. A tool that reads your notes saves you the pasting, but the notes are still the record of what's true.
How do I stop my campaign lore contradicting itself?
Keep one known-facts list, add to it after every session, and check new material against it before you use it (way 7 above). Generating less helps too: ten sentences about one place have far fewer chances to contradict anything than a ten-page history.
Will using AI make my world feel generic?
Mostly when it has nothing to go on. Give it your session logs and your own decisions, and its first answer stops being the average fantasy setting. Names need their own fix, because AI hands out the same favorites across campaigns. The guide to AI D&D names gives each culture its own naming style.
Isn't worldbuilding the fun part?
For a lot of GMs it is, and nothing here asks you to give it up. Keep the parts you enjoy and hand the AI the chores, like turning lore into clues, writing up a place you've already decided, or checking new material against the list. One GM on EN World found ChatGPT's creative work lacking, but said it "has a good amount of utility in saving time on more straightforward-but-time-consuming tasks."
Can I use an AI world builder with a published setting?
Yes. Tell it which parts of the setting your campaign uses and what you've changed at your table, and check anything it adds against the book before it goes on your known-facts list.
Does this work for systems other than D&D?
Yes. Everything here is about the world, not the rules, so it works the same for Pathfinder, Daggerheart, Call of Cthulhu, or a homebrew game.