On the 17th of October, people sat down in the Great Hall of Stirling Castle to play Call of Cthulhu.

I was in Scotland at the time. I didn’t make it. That still annoys me.

What caught my attention wasn’t the castle (well fine, it kind of was the castle). It was what sat underneath the show. The game came out of A Viking in the Sun, a University of Edinburgh research project on Harald Hardrada, and its sister show in Istanbul earlier this year was built from real archaeological objects, historical sources and Icelandic sagas. Historians did months of careful work so that four players could walk into a room and go wherever they liked.


The prep paradox


I’m building a Delta Green campaign for next year. Delta Green is the Call of Cthulhu offshoot where the players are federal agents trying to contain things that shouldn’t exist, usually at a terrible cost to themselves. Think True Detective or The X-Files. The person running the game is called the Handler, which is a much better job title than Dungeon Master - which can turn heads if said in the wrong room - “Im a Dungeon Master on weekends”.

Every game runner learns the same lesson early: players will ignore your plot. They’ll interrogate the wrong witness, burn down the wrong barn, and get emotionally attached to the bloke who was meant to be a one-line clue. That’s fine. It’s what makes collaborative storytelling work, and the chaos is where the memories come from. So why prep at all?

Because in my experience, the sessions that feel real are the ones where the world pushes back with reasons. When the players do something unexpected and an NPC responds the way that person would actually respond, the table feels it. Nobody claps. They just lean in.

Delta Green also needs the cosmic horror to land, and that’s slow work. Seeds planted early are what let the background creep in later without you forcing it.

You can’t improvise a reason you never wrote down. You can only improvise a reaction.


This is a data problem wearing a cloak


Most campaign notes look like most data platforms. There’s a 40-page setting document nobody re-reads, a notebook with half the NPCs in it, a Google Doc with the other half, and a lot of important facts living in one person’s head. Everything exists. Nothing joins.

Now there’s a new consumer of that mess. If you want an LLM to help you build the world, it needs context, and it can only use what it can find and understand. There’s been a steady stream of writing this year about making data ready for agents. A campaign setting is the friendliest sandbox I can think of for trying that out, because the stakes are a ruined Thursday night rather than a ruined quarterly report.

Here’s how I’m structuring it.


One note, one thing


Every NPC, organisation, place, object and event gets its own note. One thing per note. If you’ve done any dimensional modelling, you already know this move: pick the grain first, and everything after gets easier.

A big setting document is a table with no keys. You can read it, but you can’t point at a row in it.

Here’s an NPC note (She’s invented for this post, so no spoilers for anyone in my group).

---
type: npc
name: Dr Ruth Calder
role: Archivist, county historical society
location: "[[Ashby Falls]]"
faction: "[[Calder Family Trust]]"
status: canon
wants: Keep the 1952 flood records sealed
fears: Her brother's name turning up in them
knows:
  - "[[The Drowned Chapel]]"
  - "[[Agent Morrow]]"
secret: She burned three ledgers in 1998 and kept the fourth
---

Polite, precise, and helpful right up to the moment anyone
asks about the chapel. Offers tea. Watches the door.

Look at the fields. name, role and location are what most people write down. wants, fears, knows and secret are what make her playable.

Strip those four out and you’ve got a dimension with a surrogate key and a job title. Leave them in and you know what she does when an agent turns up asking about 1952, even though you never planned for that scene. She stalls. She calls her brother. She doesn’t mention the fourth ledger.

That’s the “why”, and it’s the part nobody writes down.



Obsidian’s [[wiki links]] turn the notes into a graph. Ruth links to the chapel, the chapel links to the flood, the flood links to three other people who’d rather it stayed buried. Backlinks show you every note that points at a given place or person, for free.

Add the Dataview community plugin and the frontmatter becomes queryable:

TABLE wants, fears
FROM "npcs"
WHERE contains(knows, [[The Drowned Chapel]])

You can run this query you run at the table, mid-session, when the players knock on a door you didn’t plan for. Who in this town knows about the chapel, and what are they afraid of? Ten seconds, and you’re improvising from something solid.


Sessions are an event log


Session notes are dated and append-only. I don’t go back and edit what happened in session three, any more than you’d update last month’s transactions.

The current state of the world gets derived from the log. Ruth’s status changes because something happened in a session, and the session note is the record of why. If you’ve spent any time with slowly changing dimensions, you’ll recognise the shape. The players’ choices are the facts. Your prep is the dimension they land on.


Canon, draft and rumour


Every note carries a status:

  • canon: true in the world, and it’ll stay true
  • draft: an idea I haven’t committed to
  • rumour: something people in the world believe, which may be wrong on purpose

Ideas start as drafts. They get checked against canon for contradictions, then promoted. Write, audit, publish, applied to fiction.

That gate matters a lot more in Part 2, when Claude starts proposing ideas. Nothing it writes goes straight to canon. Ever.


Why bother with a vector index?


Text search finds words. Embeddings find meaning.

Ask “who would be nervous if the flood records came out?” and a keyword search returns nothing, because none of the notes contain the word nervous. An embedding search finds Ruth, because her fears line is about exactly that.

This is also why the one-note-one-thing rule pays off twice. A note about a single person is the right size for retrieval. A 40-page document chopped into arbitrary chunks gives the model half of Ruth and half of someone else, and then it confidently tells you Ruth runs the pub.

(Yes, you could just use Ctrl-F. Right up until you can’t remember whether you called it the Drowned Chapel or the Flooded Church.)


The tradeoff


None of this is free.

What you get: a world that answers back. When the players go off the map, you’re improvising from motives you wrote down, not making them up under pressure. The vault is also ready for an AI collaborator, because it’s already in small, linked, labelled pieces.

What it costs: prep time, and the real risk of building an iceberg. Most of what you write the players will never see. Some Handlers will find that satisfying. Others will find it maddening.

What to watch for: the moment you catch yourself defending a note against a player’s choice. When the note and the players disagree, the note loses. The vault exists to serve their story, not to script it.

The players at Stirling Castle got to wander wherever they liked because historians had done the slow work first. That’s the job of the vault: be the slow work, so the table can be fast.

Im still finding my feed, becoming a Handler. Having a co-Handler helper gives me the confidence, it reads the vault, checks motives, generates rumours from secrets, and helps the world react when the players do something nobody expected. Everything it writes goes into drafts, and only I decide what comes out.