Open any AI image tool, type "cool t-shirt design," and you get the same beige gradient blob everyone else gets. The model isn't the problem. The prompt is.
Four things that fixed this for me:
Pick a niche before you pick a style. "Design for a 40-year-old trail runner who hates motivational quotes" gives the model something to push against. "Vintage shirt" gives it nothing.
Lock the constraints up front. I write a tiny spec first: 2-3 ink colors, no gradients, thick outlines, subject fills the chest area. AI loves adding soft shadows that screen printers hate. Say no to them in the prompt, not after.
Iterate in threes, not ones. Ask for three variations that differ by art style — woodcut, 70s cartoon, brutalist type — rather than three random seeds. You're comparing ideas, not rolling dice.
Shrink it. Before you fall for a design, view it at 200px wide. If the joke or the shape doesn't land at thumbnail size, it won't land on a shirt either.
When I just want a fast starting point, I paste the raw idea into ShirtCraft AI (https://shirtdesigngeneratorai.com). It's free, takes plain text, and hands back print-ready concepts in seconds — no design skills required. Then I take whichever one survives the thumbnail test and refine it somewhere else.
The bottleneck was never ideas. It was giving the model enough detail to have one.