An AI co-GM for WFRP 4e, and the Empire around it

The bookkeeping is handled. The Old World is not.

Ask GMs what makes this game heavy and they will talk about tracking, not about tone. Advantage on every combatant, Conditions that stack a number each time they land, Psychology, a Critical arriving on top of the injury that caused it. That is the real weight of a Warhammer fight, and your Foundry system took it off you years ago.

Your system runs the opposed Test as a single flow, treating the defender's next roll as the defence and posting what comes of it. Success Levels get counted by whichever method your table chose. Advantage clears itself when the fight ends. Fear and Terror get prompted when it begins, and at the far end the system comes back around for Corruption, infection and disease. Career advances deduct their own experience as you spend it. Encumbrance keeps pace with the packs.

One honest note about all that, because it changes what to expect from any tool near this system. The wfrp4e system is machinery rather than content: it knows to reach for a critical or a miscast table, and the table itself arrived with the module you bought. Familiar works the same way round, which is the point. What it searches is the compendia and journals your world actually holds, so a name it gives you belongs to the Empire you paid for.

None of that is a negotiation, then. Everything above stays exactly where it is, and what Familiar takes is the Empire those numbers happen inside.

• A room brought up from one sentence, with the light, the weather through it, and something playing underneath
• A voice held per NPC across a whole evening, so a Reiklander and a Middenlander never come out sounding like the same person
• The session written back into your world afterwards, searchable, instead of stacking up in a folder somewhere
• A cast the party can ask about months later, because the answer lives in the campaign rather than in a conversation
• A portrait, a document or a torn corner of a map drawn when the table needs to see the thing

• The prep, read and run loopWhat a journal page needs on it so the AI narrates your Empire and not a generic one.

The upkeep a long campaign in the Empire runs up

Warhammer campaigns are long, and the published ones are very long. The work moves accordingly: away from the fight in front of you, and into the two years of small print behind it.

None of this is arithmetic. All of it is attention, and attention is what runs out first on a Tuesday evening.

• A published campaign arrives with a cast of named people who exist mostly in relation to each other, and tonight four of them are in the same room
• Somebody wants to know what the chancellor actually agreed to, and the agreement is buried in a write-up from fourteen months back
• The book maps the towns and leaves the days between them to you, so three river legs need something on them worth stopping for
• A thread laid down in the first book pays off in the third, and nothing joins the two together except your own recall
• The coaching inn has to smell of something and sound like somewhere, again, for the ninth time this campaign

What one sentence gets you

"Before we start: what did the party promise the coachman, and who did they put offside in Ubersreik last time? Then put us on the road out of the city at dawn." Two questions and an instruction, and not a die anywhere in them.

Familiar goes to the journals for the answers rather than to its own recollection of the conversation. The promise is sitting in a write-up from six weeks ago and comes back in the wording you used at the time, not as a summary of a summary. So does the grudge, along with who was standing there when it was earned.

Then the road. First light, the traffic already on it, a bed of sound under the talking, and the gate guard who wants to know your business before any of it starts.

The coachman holds one voice for the length of the journey, and he holds the thing you decided he would rather not discuss.

Say the word afterwards and the evening goes back into a journal page in your own world, which is what makes the question answerable the next time somebody asks it.

Every Test on that road stays where it belongs. The roll to read the guard, the Success Levels that decide how much he lets slide, the Fortune point somebody burns to have another go: the wfrp4e system handles all of it, and nothing here reaches for any of them.

Register is the thing it has to get right here

Warhammer is not heroic fantasy with grubbier boots, and the difference shows up in the first paragraph of anything written for it. The people are ordinary. The odds are poor. The joke lands, and then something worse happens. An AI reaching for grandeur gets every one of those wrong at once, and a GM who plays this game will hear it immediately.

The terminology is the easy half of the problem. Tests and Success Levels, Careers and their tiers, Fate against Fortune and Resilience against Resolve, Advantage, Corruption, the Conditions under their own names: all of it comes through in this system's vocabulary, and a rules question sends Familiar looking through the compendia and journals your world has installed before the ruling comes back to you.

Tone is the half that has to be told. That is what the starter below is for, and it spends more of itself on register than any other starter in the set, which is an accurate reading of what this game is difficult about. Paste it in and it becomes the brief Familiar works from for the rest of the campaign.

"These are ordinary folk scraping by, not chosen heroes." That one line out of the starter is the whole difference between an NPC who begs and an NPC who postures, and it is doing more work than any glossary would.

Where the rules engine stops

On Warhammer Fantasy 4e, Familiar runs everything around the rules, and the rules are the one thing it does not touch.

Refereeing the rules is D&D 5e (2024) only. In combat the tracker, the turns, and the tokens still work; the rulings stay yours.

One setting to change: turn the combat auto-pilot off in Familiar's settings. It drives NPC turns through the 5e engine, so it does not know your system; the monsters' turns stay yours to run. Playing solo, that is the part you will feel.

This table is read from Warhammer Fantasy 4e's own data model and Familiar's shipped code. It has not yet been run at a live Warhammer Fantasy 4e table, so nothing here is described as observed.

The starter your table begins from

Familiar carries a Table Rules field per world, under Settings, System Rules. Whatever you put there is appended to every request, and it reaches your MCP client too.

Warhammer Fantasy 4e has its own starter, offered first when Familiar sees your world. It is a starting point rather than a ruleset: the terms, the table culture, and slots for your setting, your tone, and your house rules. Paste it, then cut what does not match your table.

We play Warhammer Fantasy Roleplay 4e. Use this system's terms: percentile Tests, Characteristics, Success Levels, Fate and Fortune points, Advantage, Careers, and Talents. Look rules, careers, and lore up in this world's compendium and journals, and defer all rules adjudication to the GM.
The system automates Tests, opposed rolls, damage, criticals, and conditions, so let it resolve the mechanics.
Standing guidance:
- These are ordinary folk scraping by, not chosen heroes; reward cunning and a timely retreat.
- Treat each Career and its Status as identity; let station open some doors and bar others.
- Play Chaos and corruption as slow creeping dread whose marks surface over many sessions.
- Let gallows humour lighten the gloom, then let something grim follow the laughter.
- Keep the Old World grubby and grounded; magic stays feared and rare, never a fix on tap.
- Track names and past deeds from the world's journals, not memory, and admit what was never written down.
Setting: [e.g. Ubersreik in the Reikland, the Empire].
Tone: [e.g. grim and perilous, with gallows humour].
House rules:
- [rule]
Never speak or decide for the player characters.

Two places it stops at the edge of your sheet

One of these is a number this system keeps and Familiar does not. The other is a field name it goes looking for and does not find. Neither is a failure being reported as a success.

• Nine of the fifteen Conditions carry a count, and that count is what the end-of-round scripts act on. Familiar can put Bleeding on a token and take it off again. Setting it to Bleeding 2, or telling you it is already at 2, is outside what the tool takes, so the stack stays a job for the sheet.
• Wounds do reach Familiar. They arrive on the character read at the path the capability table prints, and they arrive counting down from a maximum, which is the direction that makes them readable at all. Where they do not arrive is anywhere else: the scene and token listings and the GM screen are all looking for a name this system does not use for that number, so they show nothing there rather than something wrong.

Your Foundry system already puts Wounds and Advantage on the token bars, so the two numbers you watch in play are on screen with Familiar nowhere in that loop. Progression is the same story. Ask for a rest, an experience award or a Career advance and the tool names the system your world is running and hands the job back, because that is a ledger the system keeps properly and a chat message has no business editing.

The campaign that outlasts the context window

People run this game for years. A published campaign in the Empire can carry a party from a bad night in a border town all the way into the politics of a city, across several books, with a cast that keeps turning up and a conspiracy that only makes sense looking backwards.

That shape is the one an AI is worst at unaided. A model holds a conversation, not a campaign, and by the fortieth session the first one is long gone.

So the record does not live in the conversation. Write-ups, the cast, what the party worked out and roughly when, all of it goes onto journal pages in your own world, and it comes back the next time the name does. Careers move on, Corruption accumulates, and the notes keep pace because they are notes rather than memory.

• How the campaign record worksWhat gets written down, where it is kept, and how it comes back months later.

Careers are the social map, not just a sheet

The starter tells Familiar to treat a Career and its Status as identity, and to let station open some doors and close others. That is a small instruction with a long reach in this setting.

It means a rat catcher and a physician get answered differently at the same door, and that the difference is the scene rather than a modifier. It means the watch sergeant has an opinion about which of the party he addresses. Ask for a crowd and you get people with trades, because in the Empire what someone does is most of who they are allowed to be.

That is table culture rather than mechanics, which is exactly the sort of thing a Table Rules brief is for and exactly the sort of thing no rules engine would ever hold.

The parts that are the same everywhere

Connecting an AI, choosing a model, and what it costs do not change with your game system.

• Everything Familiar can doThe full tool reference. Everything outside combat refereeing runs on your system too.
• Connect Claude, ChatGPT, or GeminiThe connection steps, identical on every game system.
• Choosing an AI modelWhich model, and how much thinking headroom a hard scene needs.
• AI voices for your NPCsTTS setup and voice pinning, the same on every game system.
• Campaign memoryThe record and the recall, unchanged by your game system.