Every tarot deck starts with a vision and ends with a spreadsheet of 78 cards. The Fool, the Magician, the High Priestess — each one needs art, and each one has to look like it belongs to the same deck. That's why decks from professional artists cost hundreds of dollars and take months to commission. It's also why more readers and creators have started reaching for an AI image generator to build their own. You can produce a full deck's worth of cohesive art in a weekend, no illustration skills required. The catch is consistency: a deck where every card looks like a different artist drew it is unusable. Here's the workflow that keeps all 78 cards looking like one set.
Why tarot is a surprisingly good AI project
Tarot asks for exactly what AI does well. Each card is a single, self-contained image with one strong symbolic subject, so you don't need complex scenes with characters interacting — one figure, one setting, one meaning. The Rider-Waite tradition gives you a shared visual language that readers instantly recognize, and you can reimagine those classic symbols in your own style without copying anyone's art. The format is forgiving too: 78 small images, not one giant mural. A card that doesn't work gets regenerated, not re-commissioned, and small flaws barely show at playing-card size. Thousands of AI-assisted decks already sell on Etsy and Gumroad, so there's a proven market if you decide to publish yours.
Write a style bible before the first prompt
This is the step everyone skips, and it's the difference between a deck and a mess. Before you generate anything, write down exactly what the deck looks like: the art style (watercolor, linocut, art deco, woodcut), the color palette (muted earth tones, jewel tones, black and gold), the era (Victorian, mid-century, cyberpunk), and the line work (soft edges, heavy outlines, none at all). Then pick a handful of fixed style descriptors and reuse them verbatim in every single prompt. If you type "vintage art deco illustration, deep blue and gold palette" into every prompt, the model keeps returning to the same visual space. If you describe the style differently each time, you end up with 78 different decks. The style bible is the single highest-leverage thing you can write, so do it on paper, not in your head.
Lock the major arcana before touching the minors
The 22 major arcana cards carry the deck's identity, so nail them before you generate a single minor card. AI tools don't remember your previous prompts, so each one has to stand alone: describe the Fool's outfit, the Magician's posture, the Empress's throne in full detail every time, using the same words. Generate the majors first and review them as a group. If the High Priestess comes out in pastels while the Emperor is all moody blacks, fix your style descriptors now, before the inconsistency multiplies across the minors. This is also the stage to check the small stuff — a crown that floats, a hand with six fingers, a sword that bends like rubber. Regenerate the offender and move on; don't try to fix it in an editor, because the fixed version won't match the rest of the deck's texture.
Keep the minor arcana on a system
The 56 minor arcana are where AI decks fall apart, because that's a lot of near-identical images. Give yourself a rigid system: one background per suit, one color scheme per suit, one standard composition. Cups in cool blues, wands in warm oranges, swords in grays, pentacles in greens — whatever you pick, don't vary it. The pip cards (Ace through Ten) are mostly symbolic scenes, so you can generate them in batches and swap only the suit element between runs. The court cards (Page, Knight, Queen, King) need the most consistency, so give each suit a costume template and a pose template, and reuse them word for word. It feels repetitive while you're typing, but that repetition is exactly what makes the finished deck look intentional.
A prompt skeleton you can steal
Here's a structure that holds up well for tarot: "[Style descriptors], [card name] as [subject description], [composition], [background], [palette], no text, no letters, centered figure, full body, tarot card art". For the Three of Cups it might become: "art deco illustration, Three of Cups as three women raising golden goblets, centered composition, sunburst background, deep blue and gold palette, no text, no letters, tarot card art." Locking the last three phrases — no text, centered figure, tarot card art — in every prompt gives the whole deck a shared frame. Keep a note of the exact skeleton and every descriptor you use, because you'll want them again for the card back, the box art, and any second edition.
Deal with text before you print
Two things will go wrong in every AI deck: hands and letters. Hands occasionally come out with extra fingers — work around them by keeping them gloved, sleeved, or holding objects close to the body. Text is worse, because the model happily spells "VII" as "VVI" and garbles titles. The clean fix is to generate every card with no text at all, then add the number and title in a simple layout tool. That gives you tidy placement for the card name on the frame, which a printed deck needs anyway, and it means your Roman numerals are always correct. Do the card back as one final generation: a symmetrical pattern that reads the same upside down, because half your deck will be reversed in every reading.
From 78 images to a real deck
Once the art holds together, the rest is assembly. Lay out the framed cards, order a single proof deck before committing to a full print run — ink and paper change how colors read, and one cheap proof saves you from a box of cards that look washed out. If you're selling, keep your style bible and prompt skeleton safe; you'll need them to recreate the deck when the tool updates or when you want a second edition. A deck that used to take a professional artist a year can now take you a weekend, and the style bible is what makes it look like one artist's year of work. If you're curious what other projects AI image generators handle well, the use-case guides on this site are a good place to keep exploring.