Made Claude keep the books for our D&D world
My friends and I have played the same D&D campaign for six years. Before episode 91 I needed to know what each character knew, so I asked Claude to keep the books.
Episode 90 of our D&D campaign ended with the party in pieces. Two of the five characters were inside a cannon. A third was on the far side of a portal that had just closed, holding the fourth. The fifth was last seen fighting at a crash site. That’s where we stopped.
I run this campaign. My friends and I have played it since 2020, mostly in Greek, three or four hours at a time. So with episode 91 coming up, I had a very practical problem. Before I could pick up that cliffhanger, I needed to know what each character knew, which isn’t always what their player knows. And I needed to know what the world still owed them.
Six years in too many places
None of that lives in one place. My players keep a wiki with a page for most episodes, written from their side of the table. The first year of sessions exists only as video, which I’ve written about transcribing. Years of messages about the story sit in our Discord.
My own planning is spread over scanned notes and documents, and the oldest D&D document in the pile is from April 2017. And I write songs for this campaign, and they hide pieces of the story.
So I gave all of it to Claude Code and asked the question I’d face at the table anyway. Where is everyone, and what do they know?
The wiki remembers what the players saw
A wiki that players write has a trap built in. It records what they saw and what they were told, and in a story like ours, some of what they were told is a lie. When a character tells the party something, the wiki writes it down. That proves the character said it, and nothing more.
A model reading that wiki will happily treat it as the truth, so the first rule was about evidence. Every claim in the reference Claude built carries a label. It’s confirmed canon, a strong inference, unresolved, or old planning that may not be canon anymore. Other labels say where a claim came from, like me speaking out of character or a private conversation with one player.
That sounds like bureaucracy until you see what it’s for. When I ask what’s true, I get the true part, and everything else comes with a warning label attached.
Names are the hard part
Fantasy names are hard enough in one language. We play in Greek, and a name said aloud in Greek gets spelled several ways once it lands in a wiki or a transcript. Claude built a register of names, and it has 134 rows.
Most rows are dull and useful, like the spelling variants of one person. The interesting ones are the traps. Some words are titles, so the same word can mean two different people in two different eras. Some characters change form and come back under another name. And some names only look alike, so they must stay apart forever.
Then there’s the rule I care about most. Sometimes two names might belong to the same person, and the evidence doesn’t settle it. Then the register says to search for them together, and never to merge them as canon.
A model’s instinct is to tidy up. Give it two similar names and a vague hint, and it will tell you they’re the same person. In a campaign with secrets, that’s the kind of answer the story should give later, at the table.
What a player knows and a character doesn’t
I keep secrets, too. Ten of them sit in a register marked for me only.
The subtler problem is the gap between a player and their character. My players compare notes and read the songs closely, so they often know things their characters were never told. The audit Claude wrote names four of these knowledge traps. Now I can see one coming before someone acts on what their character couldn’t know.
The songs needed a rule of their own. I write lyrics for big moments, and they’re full of the story, but a song is poetry. The archive treats a line from a song as evidence about a theme, never as a fact about the world, unless I confirm it separately.
Twelve rulings and a checklist
Where the sources disagreed, Claude didn’t pick a winner on its own. It wrote down each conflict with a ruling it recommended, and left the decision to me. There were twelve.
At the end of the audit is a checklist of ten things that must stay consistent when we pick up from the cliffhanger. Above it are 21 questions marked unresolved and important. That’s as much my to-do list as Claude’s.
The ledger
My favourite part is the ledger. It lists every promise, bargain, debt and obligation still open in the story, and who owes what to whom. There are seventeen of them, collected from six years of play.
When episode 91 starts, two of the characters will still be inside that cannon. I’ll know what each of them knows and what the world owes them. The rest is up to the dice.