Notes in, Patreon release out: how Crunchtime runs FrozenDice on AI agents

Creative work dies in production friction. The idea is there — campaign notes exist, monster descriptions are drafted, lore fragments assembled. What's missing is the last step: gathering it all, formatting it, and actually getting it published.

For FrozenDice — Crunchtime's own D&D hobby project at patreon.com/frozendice — this was the exact problem. Three years of Nordgaard campaigns have produced enough homebrew material to sustain a Patreon channel with monthly releases. But for a founder simultaneously running an AI consulting firm, there aren't enough hours in the week to manually format Homebrewery PDFs, write image prompts, and assemble post copy.

The solution is the same one we deliver to clients: a structured AI agent pipeline that handles the repetitive production steps, inserts human review where creative judgment actually matters, and delivers a publication-ready draft.

This post walks through that pipeline — with the first release, Nordic Valkyries, Vol. 1 — The Storm Sisters, as evidence.

What FrozenDice is and what it publishes

FrozenDice is a tabletop RPG hobby project rooted in three Nordgaard campaigns and two one-shots spanning several years. All content is 5e-compatible homebrew: monsters, NPCs, magic items, and mini-adventures built around Nordgaard — a Norse-inspired setting with its own gods, history, and factions.

The publishing format is Homebrewery — a web tool that renders Markdown into professionally formatted RPG PDFs. Distribution is via Patreon.

The first release through the new pipeline is Nordic Valkyries, Vol. 1 — The Storm Sisters: three Valkyries — Ròta, Hildr, and Geirdriful — inspired by Old Norse skaldic tradition and the aesthetic of God of War: Ragnarok. Sisters from a fractured Valkyrie choir, each with their own agenda, combat method, and role in Nordgaard lore.

The release contains 3 complete 5e character entries at challenge ratings CR 11, CR 12, and CR 13, available at patreon.com/frozendice.

The problem the pipeline solves

Producing a single Homebrewery release manually involves:

• Retrieving source material from D&D Beyond (login-gated, no public API)
• Structuring lore text and stat blocks into a consistent format
• Writing Homebrewery markdown with correct {{monster,frame,wide}} blocks and \page breaks
• Generating image prompts, iterating through candidates, selecting and uploading
• Writing Patreon post copy with teasers, tags, and tier notes
• Exporting the PDF via Homebrewery and assembling everything for publishing

For one character with full lore, stat block, and image: 2–3 hours of work. Multiply by three sisters in one release, and that's a weekend of production for a hobby project that should feel like a break.

The pipeline compresses this to roughly 45–60 minutes of active, creative decision-making — everything else is automated.

The pipeline: 7 stages, 5 human review gates

The pipeline is built in Notion around two parallel structures: a Nordgaard world hub for campaign canon (Bestiary, NPCs, Places) and a Patreon Releases hub for publication artifacts. The two are cross-linked via Notion relations so each Bestiary entry knows which release it belongs to.

Stage 1 — Source consolidation (human-led)

D&D Beyond material is login-gated. Christian manually copies relevant stat blocks and lore fragments into the Source Material page for the release. The pipeline reads from here and normalizes to structured objects per character.

Stage 2 — Research and style-loading (automated)

The agent fetches four existing style guides from Notion and distills them into a ~500-word Style Contract — the system prompt context for all subsequent generation steps. Simultaneously, a Norse Reference Pack is generated: names, themes, combat flavor, visual motifs. Christian reviews and approves both (Gate G2).

Stage 3 — Enrich and canonicalize (automated, with review)

For each character the agent generates a Bestiary page in Notion containing: ~200 words of lore text, appearance and personality description, full stat block with gaps filled based on 5e balance logic, tactics section, and 3 playable plot hooks. Cross-linked to the release page. Gate G3: Christian edits Bestiary pages in Notion until approved.

Stage 4 — Format to Homebrewery markdown (automated)

The agent reads the approved Bestiary pages and produces a single coherent Homebrewery document with correct \page breaks, {{monster,frame,wide}} blocks, table of contents, foreword, and Credits section. Stored in Notion as one code block — ready for copy-paste to naturalcrit.com.

Stage 5 — Image generation (automated)

1 cover image + 1 hero illustration per character = 4–5 images total. Generated via Gemini CLI with the Nano Banana extension. 2–3 prompt variants per image; Christian selects from candidates (Gate G4). Selected images upload as Notion image blocks.

Stage 6 — Patreon post draft (automated)

The agent generates a post title, 3–4 paragraphs of teaser copy, tags, and tier notes for the Patreon post. Stored on a Patreon Post Draft subpage.

Stage 7 — Manual publishing (human-led)

Christian copies the Homebrewery markdown → pastes into naturalcrit.com → exports PDF → uploads as a Patreon draft. Gate G5: final review. The release page is marked Published.

The five human review gates (G1–G5) aren't optional — they're architecture. AI produces stylistically consistent markdown without syntax errors. It doesn't always produce gameplay-balanced CR 13 in practice. Creative ownership and final judgment belong to the creator.

Storm Sisters: the artifacts that prove the pipeline

Nordic Valkyries, Vol. 1 is published at patreon.com/frozendice as a member-tier post. This article is the documented account of how the pipeline produced it.

Ròta, Valkyrie of Storms (CR 13) is the sister the choir calls down when subtlety has already failed. Her stat block reflects this: she opens with call lightning, escalating to chain lightning or destructive wave under pressure. The passive Storm-Born grants immunity to thunder damage. Legendary action Wing Buffet pushes nearby creatures away with a DC 18 Strength save. She doesn't retreat. A plot hook, "The Sister Who Stayed", references inscriptions above the gates of Nordgaard fortresses where she lingered long after her duty was done — and no skald has explained why.

Hildr, Blizzard Valkyrie (CR 12) works inside the blizzard. Blizzard Camouflage makes her heavily obscured to creatures more than 5 feet away and lets her vanish and reappear 60 feet away as an action. The tactics section is specific: open with sleet storm to create whiteout conditions, never approach the same target twice, prioritize the healer. Legendary action Snowstorm cycles the blinded condition (Constitution DC 17). She asks every mortal she kills one question: "Did you know, at the end, what you were?"

Geirdriful (CR 11) marks from the high air currents above battle and sometimes withholds names from the muster roll for reasons that are hers alone. The least martial of the three, and the most cryptic.

What the pipeline actually does — and doesn't do

The AI agent handles well:

• Style consistency across all three characters (the Style Contract works)
• 5e mechanical syntax and balance logic in stat blocks
• Homebrewery markdown without formatting errors
• Image prompts and variants within style constraints
• Plot hooks and roleplay scenarios that are actually playable

Human oversight is still required for:

• Gameplay balance: CR 13 on a stat block isn't the same as CR 13 at the table
• Aesthetic selection: image candidates vary in tone; some hit the Norse feel, some don't
• Lore precision: Nordgaard world history belongs to the creator
• Final publishing: Homebrewery and Patreon lack public APIs for programmatic posting

The net effect: instead of 6–8 hours of manual production per release, the pipeline compresses active creative investment to ~45–60 minutes. The creator makes decisions — doesn't format.

The tool stack

From D&D hobby to the pattern for content agents

FrozenDice is a hobby project — but the pipeline architecture isn't hobby-specific. The same pattern appears in any creative production process where source material is scattered and output is formatted:

A fiction author serializing short prose on Substack needs the same: source draft → style contract context → formatted chapter draft → human review → publishing.

A podcast network that wants blog posts per episode needs the same: audio content → transcription → structured blog draft → editorial review → publishing.

An agency with one creative brief that needs ten campaign variants needs the same: brief in → style contract context → variant drafts → client approval → finalization.

The pattern: source in → style definition → structured generation → human approval per stage → formatted output → publishing.

How we'd build this for you

FrozenDice is our internal proof of concept — not a client case study.

We don't claim to have delivered content pipelines for fifty clients. We claim to be running one autonomously on a real creative project — with documented pipeline design, produced artifacts, and published results. That's the highest standard for "we know what we're doing": actually doing it ourselves.

A typical content pipeline engagement with us is a 3–5 week pilot: map your production process, identify which stages suit automation, and deliver a working pipeline with 2–3 automated stages and defined human review checkpoints.

Book a Discovery Sprint →