Glyphhaus

1. Brief

Numinous Current Tarot is a secular deck for attention, clarity, and the everyday strange. The goal was to ship a complete, coherent visual world: a 78-card deck with consistent woodcut/engraving aesthetics, a repeatable card template, and enough range to hold across wildly different subjects without style drift.
‍
Constraint: speed and volume. The output had to feel designed, not “AI-generated”. That meant treating AI as a candidate generator — and treating art direction, selection, and integration as the craft.

2. Workflow

Each card followed a repeatable loop:
Micro-brief (subject, mood, composition, what must stay constant) → candidate generation under strict style constraints → high-rejection curation (most candidates fail subtly) → selection → integration into the card system (typography, borders, labels, marks, suit logic).
‍
The guiding rule was simple: consistency beats novelty. One impressive outlier is useless if the deck collapses at card 27.

3. Process Proof
‍
Selection (misses vs. final):
AI output is cheap; coherent direction is not. Most candidates were rejected for small but fatal reasons: the linework gets too smooth, the lighting turns cinematic, the composition becomes “AI-clever”, or the mood slips out of the deck’s register. The final images were chosen because they obey the deck’s visual constitution — not because they were the flashiest option.

Prompting as a template system:
The prompts were intentionally formulaic to keep the deck consistent at scale. I worked with a repeatable base prompt (“retro ink drawing on white background”) and then swapped only a few controlled variables: color, subject, and style notes specific to the card. Consistency was reinforced by iterating from previous successful cards (using the last result as a visual anchor), not by chasing novelty.

Curation: the key decisions were choosing the right subject framing and keeping the output inside the established look — rejecting images that introduced “AI sheen”, modern rendering cues, or compositional drift.

Integration (image → shippable asset):
The finished deliverable isn’t the raw image — it’s the card as a product unit: typographic hierarchy, border rules, title/numbering, monogram placement, and suit color logic that stays stable across the full set.

4. Design System
‍
The deck is built as a template so the visual language survives scale: consistent header + category label, fixed margins, stable border treatment, emblem placement, and the NC monogram as an anchor. This is what turns “a collection of images” into a coherent library of assets that can be printed, published, and extended.

5. Editorial integration
‍
The same visual language was applied to long-form content: a companion handbook and a 400-page course. This is where AI work often fails — not at generation, but at layout discipline. The goal here was cohesion at scale: readable page rhythm, consistent typography, and illustrations that support the text instead of hijacking it.

The point isn’t “AI can generate images.” The point is: the pages remain designed — consistent hierarchy, spacing, and tone across a long document.

6. Tools

Sora — candidate generation under strict style constraints (high rejection rate by design)
ChatGPT + Claude — drafting support and structure for handbook + course content (human-edited for voice + coherence)
Illustrator — template system, marks, borders, production assets
InDesign — editorial layout and consistency at long-form scale