Prompts That Keep AI Graffiti Readable

Most bad AI graffiti is not a “model problem.” It is a brief problem. People paste a paragraph of lore, demand six art directions, and then act surprised when the letters look like wet bread. Treat prompts like can control notes for a writer on a deadline.

A prompt that fits on one sticky note

Word + style + spray behavior + wall + light + 2–3 colors. Example: “Wildstyle tag NOVA, fat-cap fill, soft overspray, heavy but controlled drips, dark brick, sodium streetlight, pink and cyan.” That is enough. If your tool already has style/drip/wall dropdowns (ours does), do not restate every dropdown in the custom box—use custom text for the missing flavor only.

Language that nudges spray, not fantasy art

• Caps: fat cap, skinny cap, soft fill, spatter
• Paint: opaque coverage, chrome fill, candy fade, keyline
• Mess (on purpose): wet drips, runs, overspray halos
• Surfaces: glazed brick, raw concrete, plywood boards, metal shutter

Change one dial per credit

Lock the word. Run A. Then change only palette, or only drip level, or only wall. Side-by-side comparison teaches you what the model actually heard. Blasting a brand-new novel prompt every time guarantees you learn nothing and empty the credit balance.

Hard no’s

1. Full lyrics or mission statements as the “tag.”
2. “Make it iconic / viral / masterpiece” with no visual constraints.
3. Asking for microscopic calligraphy on a mural canvas.
4. Referencing a specific living writer’s famous piece as the style target.