The challenge
One week to author the Tarot deck that I always wanted.
Researching how to make generative AI reliable, I chose to make my own Tarot deck. In a week.
The Thoth deck is an extremely rigorous system of correspondences between imagery, color, and composition, built on a technical system of interacting hyper-specific signifiers.
I have been an illustrator and a symbology nerd my whole life. I’ve always dreamed of painting the Tarot deck I’d most want to own.
Could AI bring that project to life, under my own art direction, at a level I could actually respect? Well, yes. But it took some engineering…
The idea
How to formally compile meaning?
My essay Rigorously Joyful explores how formalizing knowledge could preserve an organization’s identity while giving its people more freedom to discover.
The deck brought the same question into my own visual practice. Meaning lives in the relationships between symbols: what governs an image, what supports it, how color changes its reading.
Preserving those relationships across hundreds of generations required making the tradition’s internal logic explicit. The research became a grammar the agents could work with, connecting each card’s correspondences to its role in the composition while leaving me room to steer its visual expression.
01 / Research
Retrieval before generation
I used a prototype of my app Brush to break the mission into structured tasks. Agents produced a symbolic bible, a visual identity study, a fusion framework, and a typography bible.
Further research covered color, composition, and the Minor Arcana. The surviving corpus spans 16 research and design documents that define the context boundaries.
02 / Structure
Dynamic prompting
The research became card briefs, shared instructions, and structured data. Each Major Arcana card had four image roles: its governing act, atmosphere, structure, and material traces. A TypeScript resolver combined these with symbolic correspondences and color strategies to produce inspectable prompts. Recorded diagnoses could inform the next pass. Machines manage it; I direct.
03 / Production
Create, compare, refine
Generation ran in documented waves, often producing entire sets of cards, followed by visual critique and my art direction. Once the visual language stabilized, I moved to controlled variations of individual cards. Prompts, outputs, and approval records preserved what changed and why.
04 / Color
The color problem
Color is not expressive in this system, every card has rules. By separating monochrome generation from coloring and building a reusable Python engine to control its placement, color could be solved for all cards, in any version, programatically. Segmentation, brightness thresholds, and positioned masks directed color into specific regions, exported as transparent layers. These could be adjusted independently beneath the original engraving.
Outcomes
A repeatable process for carrying symbolic requirements from research through generation, critique, and final artwork.
- A corpus of 16 research and design documents defining the symbolic, visual, and typographic constraints.
- Structured card briefs and a TypeScript resolver that compiles correspondences and image roles into inspectable prompts.
- Documented generation and critique cycles, preserving prompts, outputs, and art-direction decisions.
- A reusable Python color engine that exports independently editable layers beneath the original engraving.
- Reusable card layouts in two print sizes.
- A finished deck of 78 unique cards, held together by a coherent visual language.