Every chapter of The Living Chronicle is generated by AI under human editorial oversight. We are open about how this works because the alternative — pretending the writing is purely human — would be dishonest.
The pipeline
Every chapter passes through ten steps:
- Collect — votes, investigations, and operations from the last 24 hours.
- Community detection — find which plot threads heated up.
- Retrieve — pull the relevant sub-graph from the lore database.
- Plan — the orchestrator drafts a chapter outline.
- Simulate — the DM agent resolves mechanics (rarity, outcomes, state changes).
- Validate — the rules referee checks against canon.
- Update graph — state changes commit.
- Write — the writer agent produces the prose.
- Edit — the editor agent polishes voice and pacing.
- Publish — a human curator signs off, then the chapter goes live.
No step publishes without the curator. That is intentional and load-bearing.
Which models
- Claude (Anthropic) — writing, editorial, the orchestrator.
- Gemini (Google) — validation, structured output, the rules referee.
- OpenAI TTS — audio narration for Chronicle+ subscribers.
The full disclosure, including which models we use for what and why, is at /ai-disclosure.
What the AI does not see
- Your name, email, or any personal data.
- Coven chat transcripts.
- Theories from individual players.
- Anything that could identify a real person.
The AI sees aggregated game state — vote percentages, investigation totals, faction populations — and the canonical lore sub-graph. That is it.
The curator
The human curator reviews every chapter before publication. They can approve, reject and regenerate, or make small inline edits. We picked this design over auto-publish-with-rollback because trust compounds: one harmful chapter that ships and gets rolled back is much more expensive to recover from than one chapter that ships an hour late.
Energy use
A typical chapter costs about 57,500 tokens across the full pipeline, which works out to ~1–3 Wh of electricity — about the same as running a laptop for fifteen minutes. We chose batch generation (one chapter for everyone) instead of per-player AI specifically to keep this number low.