The Road to the Council — Proposal
Thesis
The Road to the Council (a working title) is a week-long multiplayer adventure played in a group chat and on a shared map, run by an AI game master that holds a private conversation with every player at once.
Players start alone, scattered across the map, and race for five days to reach a council. Along the way they take side quests, meet each other, and pick up secrets. Only those who arrive join the fellowship and unlock the shared group chat, where the second act begins.
The bet is that a game master who whispers to every player simultaneously was never practical before LLMs. A game built around private threads that converge on a shared story is a new genre, and it can be a profitable small business with frequent, low-value purchases.
Why now
LLMs make nuanced, individual attention cheap. A human game master can run one shared story for a table of five, but cannot hold a separate, ongoing private story with each player while keeping the world consistent. An LLM can, for every player, around the clock.
That unlocks three things tabletop games have never managed together:
Private threads at scale. Each player has their own secrets, perceptions and side deals, and the game master knows all of them.
Selective disclosure. The game master decides what each player learns and when, so tension comes from what one player knows that others don't.
Asynchronous play. The hardest part of tabletop role-play for adults is scheduling a four-hour session. Here, players act for a few minutes a day, whenever suits them, and the story keeps moving.
The same primitives sit underneath the wider mediator project: persistent memory, private reasoning across several people, and revealing only what is useful. The game is the most forgiving place to prove them, because when the game master gets something wrong, players laugh.
The game
A campaign has two acts: five days on the road, alone, then the fellowship together. The group chat is the reward for reaching the council, not the starting point.
Players who reach the council in time form the fellowship; late arrivals still join act two, with a twist.
The map. A web page, linked from the chat, shows the world and every player's position. Players see others exactly when nearby and only by rough region otherwise, so the map raises questions ("why is Sarah heading east?") that bring people back twice a day.
The road (days 1 to 5). Each player has a private thread with the game master, plus a secret: an artefact, a hidden allegiance, or something they know. Every day they get a few actions: travel, explore, rest, or take a side quest. Side quests can win items, allies or secrets, but cost time.
Daily resolution. All actions resolve together at a fixed time each evening, and the map updates on that tick. Players who check in all day get no advantage over those who check in once.
Encounters. When two players' paths cross, the game master opens a temporary private chat for just those two. They can travel together, trade, or lie. Small alliances form before the council.
The council. Players who arrive in time join the fellowship, the group chat opens, and the game master runs the meeting where secrets start to surface and the shared quest is set.
Late arrivals. Missing the council is a role, not an exit. A late player arrives as a suspicious stranger, or returns having been captured and turned. Nobody is left out, and the best drama of act two often comes from them.
Act two, the fellowship. The group pursues a shared goal in the group chat, with private threads continuing underneath. Everyone kept something back from the road.
Group size. Five to eight players. Fewer and encounters are rare; more and the private threads become hard to keep consistent.
The setting, names and rules will be original. The structure echoes classic fantasy, but the game must not borrow any existing world.
What makes it different
AI storytelling already exists, so the game must stand out on structure, not on having an AI narrator.
| Format | Example | What it lacks |
|---|---|---|
| Single-player AI adventure | AI Dungeon | No other people, so no secrets, alliances or betrayal |
| Several players typing into one shared AI story | Common in Discord bots | Everyone sees everything; no private information |
| Human-run play-by-post games | Forum role-play | Depends on a volunteer game master who eventually burns out |
| Tabletop sessions | D&D at a table | Needs everyone free for hours at the same time |
The Road to the Council combines private threads, a shared map, asynchronous daily play and a converging structure. Each of these exists somewhere; the combination does not, because only an AI game master can run it.
How it works
Ordinary code owns the world; the LLM only narrates and judges. If the model is in charge of state, it forgets things and leaks secrets.
The narrator sees only what the disclosure rules allow for that player, and its changes reach the world only if the state engine accepts them.
World state engine (code). Owns the map, positions, the daily clock, inventories, and a ledger of who knows which secret. Every change is stored as an event, which gives a full replay of each campaign.
Disclosure rules (code). Decide what each player may see: their own thread, fog-of-war positions, and secrets only once the ledger says they've learned them. The LLM is given only what the player in front of it is allowed to know.
Narrator (LLM). Writes each player's road, runs encounters, and judges fuzzy actions ("I try to talk my way past the guard"). It proposes changes to the world; the engine accepts or rejects them.
Chat layer. A Discord or Telegram bot for the prototype, with private threads, temporary encounter chats and the fellowship group. WhatsApp needs the Business API, with approval and per-message costs, so it comes later if at all.
Map page. A web page that reads positions and fog of war from the engine.
Cost control is designed in from day one. Sending the full story history to a large model on every message could cost tens of pounds per campaign. Instead:
Send a summary of the world state and recent events, not the whole history.
Use prompt caching for the fixed parts: rules, setting, character sheets.
Use a cheap model for routine turns and a strong model only for key moments: the council, betrayals and the finale.
Log the model cost of every test campaign. The target is a few pounds per campaign.
Business model
Revenue comes from many small purchases made on a card, with no procurement and no large one-off deals.
| Stream | Buyer | How they pay | Notes |
|---|---|---|---|
| Workplace campaigns | Team leads with a social budget | Per campaign, per player | Highest price per player; a week-long game in the team Slack fits the working day |
| Friend-group campaigns | One person in a group of friends | Per campaign for the whole group, e.g. £15 to £25 | Feels like buying a game, not another subscription |
| Campaigns by other writers (later) | Players | Per campaign, revenue shared with the writer | Turns content from a cost into a marketplace |
Illustrative maths, not a forecast. 1,000 campaigns a month at £20, with £4 of model costs each, gives about £16,000 a month gross.
What decides whether that happens:
Spread. Each campaign puts 5 to 8 people through a memorable experience. If one or two of them start a game with another group, growth needs no paid marketing.
Repeat play. Campaigns end, which is a natural place to stop. Sequels, characters who carry over and seasons give a reason to start the next one.
Cost per campaign. At a few pounds the margins are healthy; at tens of pounds they disappear.
Risks
The biggest risk is that the game isn't fun enough to come back to; everything else is secondary.
| Risk | Why it matters | Mitigation |
|---|---|---|
| Players don't return to their private thread | No engagement means no spread and no repeat play | Test this first; daily cliffhangers and a visible map |
| Story loses consistency across threads | Contradictions and leaked secrets break the game | Code owns state and disclosure; the LLM only narrates |
| Model costs too high | Wipes out the margin at £15 to £25 per campaign | Summaries, caching, cheap models for routine turns; log cost per campaign |
| One slow player stalls everyone | Asynchronous play depends on every player acting | Daily deadlines; the game master moves absent characters in the story |
| Hit-driven market | Games succeed or fail on word of mouth | Keep build costs low; judge on early signals, not hope |
| Crowded AI role-play space | Rivals can copy an AI narrator | Compete on the format, which is harder to copy than a narrator |
| Borrowed setting or names | Legal risk and a weaker brand | Original world, names and rules from the start |
First experiment
A scrappy prototype and five test campaigns will show whether players come back to their private threads without being reminded.
Build one campaign: a single map, the five-day road, the council and a short act two.
Build the minimum: a Discord or Telegram bot, a simple map page, the state engine and cost logging. Watch every log.
Run it free with three friend groups and one workplace team.
Ask each group at the end whether they'd play again, and whether they'd pay.
Signals that it's working:
Players message their private thread daily without reminders
At least one group asks to play again without being prompted
At least one team lead or friend group is willing to pay
Model cost per campaign is a few pounds, not tens
Kill criteria. If none of the first three signals appear after five campaigns, stop or substantially rethink the format before building more. The engine still carries over to the wider mediator project.