A D&D session-prep and post-session documentation toolkit powered by the Claude API. Runs as a FastAPI + Vue 3 web UI or a suite of CLI tools — generate encounter documents, keep campaign canon consistent, process Zoom VTT transcripts, and produce multi-voice narrative session documents.

Requirements: Python 3.10+, Node.js (web UI only), uv

uv venv
source .venv/bin/activate
uv pip install anthropic pyyaml pyperclip pyvis fastapi uvicorn
export ANTHROPIC_API_KEY=sk-ant-...

# Build the frontend and start the FastAPI server
./startup

Then open http://localhost:8000.

No build step needed — run scripts from a campaign workspace:

python new_workspace.py ~/campaigns/mycamp --name "My Campaign"
cd ~/campaigns/mycamp
python ~/CampaignGenerator/prep.py --beat "The party enters Icespire Hold"

python new_workspace.py ~/campaigns/mycamp --name "My Campaign"
cd ~/campaigns/mycamp

Generates a config.yaml with absolute paths — every script auto-detects it when run from the workspace directory. No --config flag needed.

# Single encounter beat
python ~/CampaignGenerator/prep.py --beat "The party enters Icespire Hold"

# Three-stage pipeline: Lore Oracle → Encounter Architect → Voice Keeper
python ~/CampaignGenerator/prep.py --mode pipeline --beat "The party enters Icespire Hold"

# Full session arc
python ~/CampaignGenerator/prep.py --session "1. Travel to Hold 2. Confront boss 3. Dragon reveal"

# Convert Zoom transcript to summary
python ~/CampaignGenerator/vtt_summary.py session.vtt -o summaries/session_12.md

# Regenerate grounding docs
python ~/CampaignGenerator/campaign_state.py summaries.md -o docs/campaign_state.md
python ~/CampaignGenerator/distill.py summaries.md -o docs/world_state.md

# Session doc pipeline — see sd_*.py section below

The recommended path for producing narrative session documents is a four-script pipeline with a human-review checkpoint after each stage:

VTT transcript + recap
        │
 enhance_summary.py   ← enhance raw summary
        │
 scene_extract.py     ← extract per-scene dialogue and action beats from VTT
        │  [human review]
 sd_consistency.py    ← Pass 1: consistency check vs. campaign state / world state
        │
 sd_plan.py           ← Pass 3: narrative plan — assigns characters to scene chunks
        │  [human review]
 sd_narrate.py        ← Pass 5: per-scene first-person narration in each character's voice
        │  [human review]
 assemble.py          ← combine narrated scenes into the final session document

Each script is standalone — run it, review the output, then proceed. See docs/cli/session_doc_pipeline.md for flags, voice files, player-name mapping, token scaling, and design rationale.

The RLM pipeline lets prep and narration tools pull statblocks, encounter tables, and canon lore from a local library without hallucination:

1. Retrieve — rpg_retriever.py searches in tiers: MemPalace drawer → statblock → cost-tagged candidate
2. Propose — dossier_proposer.py writes docs/dossier_proposal.md — a ranked candidate list for human review and approval
3. Render — approved proposals are consumed by render pipelines (prep.py, sd_narrate.py, planning.py)

The human checkpoint between retrieval and rendering is enforced by a CI test — render pipelines cannot call retrieval functions directly. See docs/rlm/rlm_pipeline.md and docs/rlm/rlm_architecture.md.

All scripts auto-detect config: CWD config.yaml first, then config/config.yaml in the script directory. Running from a campaign workspace means no --config flag needed.

Override with --config path/to/config.yaml when necessary.

python -m pytest tests/

All scripts default to Claude Sonnet (claude-sonnet-4-6) and accept --model for overrides. Narration scripts accept --fast for Claude Haiku (~4× cheaper), useful for iterating before a final run.

python sd_narrate.py ... --fast                       # Haiku — iterate quickly
python sd_narrate.py ... --model claude-opus-4-8      # Opus — highest quality

MIT