Most people who start making board game prototypes with AI hit the same wall around the third card: every piece looks like it came from a different game. The color drifts, the line weight jumps, and the tiny icons stop matching. You do not need a design degree to fix this. You need to lock a few decisions before you type a single prompt.
Decide the physical size first
Before anything else, work out the real dimensions. A standard poker card is 63 by 88 millimeters. A tarot card is 70 by 120. Mini cards for a travel game might be 44 by 67. This matters because prompt language about detail scales with the print size. A prompt asking for "intricate filigree" is useless if the final print is four centimeters wide and the filigree becomes gray mud.
Write the dimensions down and keep them. Every asset in the set should be composed for that same rectangle, with the same safe margin of about three millimeters inside the trim. If you generate at a large size and shrink down, the detail survives. If you generate small and stretch, it does not.
Build a one-paragraph style anchor
The single most effective trick for consistency is a short block of text you paste into every prompt unchanged. Something like:
"Flat vector illustration, two-tone line art over solid fill, palette of deep indigo, warm sand, and muted terracotta, no gradients, no drop shadows, clean silhouettes, centered composition with generous negative space."
That paragraph is your style contract. Change the subject line, keep the contract. When you paste it into a tool, the model has far less room to wander. If you want to see how much prompt structure changes results, there is a walkthrough on the AI image generator site about writing prompts that hold together across a series.
Make a reference board before you make art
Collect eight to twelve existing images that share the feeling you want. Not to upload them, but to look at while you write. Note three things about each: the dominant colors, the line thickness, and the amount of empty space. You will find that good card art is far emptier than you expect. New designers fill every corner; production games leave room for icons, numbers, and text.
Once you have those notes, compress them into the style anchor paragraph. The board is scaffolding. You throw it away once the paragraph does the job.
Generate one hero card, then reverse-engineer it
Do not generate twenty cards in a batch. Generate one that feels right, then look at what the model actually produced and adjust. Maybe it gave you a heavier outline than you asked for. Fine, that is now your outline weight. Update the style anchor to describe what you got, not what you hoped for. From that point on the anchor describes reality, which is what makes it reproducible.
This inversion is the difference between fighting the tool and steering it. You are not specifying an imaginary style; you are codifying an existing output so it can repeat.
Use a fixed seed or reference image when available
Many tools accept a seed number, which keeps the underlying randomization stable. Others accept an image reference or a style reference. Either one is worth more than any adjective in your prompt. If your tool supports a reference image, use the hero card as the reference for every sibling card and describe only what changes: "same style, subject is a compass rose instead of a lantern."
Plan the set on a grid, not a list
Before generating, sketch the whole deck as a rough map. Which cards are resources, which are events, which are characters? Group them so that similar items get generated in the same session, while your prompt and your eye are calibrated for that category. Character cards want a consistent body proportion and framing. Resource cards want a consistent icon scale. Mixing the two in one batch is how sets end up feeling incoherent.
Check for the three common failure modes
When something feels off but you cannot say why, look for these:
- Scale drift. Put six cards side by side at final print size. Does the main subject occupy a similar amount of the frame on each? If one card's figure fills the card and another floats in the middle, you have scale drift.
- Palette creep. Sample the background color of each card. If one is warm off-white and another is cool gray, the deck will look dirty when shuffled and dealt. Fix it in post by matching the background values.
- Weight inconsistency. Zoom to 100 percent. Are the outlines the same thickness everywhere? Thinning a thick line is easy. Thickening a thin one across a full card is tedious.
Do a print test before you commit
Print six cards at final size on the paper stock you intend to use. Not on screen, not on plain office paper if you plan a linen finish. Color shifts on uncoated stock, and small text fills in. Look at the test cards under the lighting where the game will actually be played, which is usually a dim room with overhead light, not a bright monitor.
Keep the final file exports generous: 300 DPI, with bleed, and text as a separate layer if you can. You will want to move a number two millimeters later, and rebuilding the art for that is a waste of an afternoon.
What to do when a card still looks wrong
Regenerate rather than repair, unless the flaw is a stray shape on an otherwise good card. Repainting an AI image in an editor often produces a patch that is technically correct and stylistically off, and you will see it every time you play. A fresh generation with the same anchor is usually faster and cleaner.
The full set rarely needs to be perfect. It needs to be consistent, legible at arm's length, and cheap to redo when a rule changes. Once those three are true, you can move on to the part that actually matters, which is whether the game is fun. If you want to explore different visual directions before committing to a style anchor, try a few test prompts on the image generation tool and keep the ones whose palette survives being shrunk to card size.