Two people use the same AI generator; one gets murky clip-art, the other gets images people ask about. The difference is rarely the model — it's the prompt. Prompting is a small craft with a handful of learnable rules.
Subject → Setting → Style → Light. Each layer answers one question the model must otherwise guess:
Order matters: models weight early words most, so lead with what's essential.
Some words steer harder than others. Styles: watercolor, oil painting, flat vector, pixel art, cinematic photo, isometric 3D, ink sketch. Light: golden hour, neon glow, soft overcast, dramatic rim light. Quality nudges: highly detailed, sharp focus. Camera terms work on photographic styles: close-up portrait, wide-angle, shallow depth of field. Steal freely from this list — everyone does.
The negative prompt lists what the image must avoid — and it fixes the classic AI artifacts. A dependable starter set: "text, watermark, blurry, low quality, extra fingers, deformed hands". Then add case-specific exclusions: generating a landscape? Negative "people". A daytime scene that keeps going dark? Negative "night, darkness". One exclusion per problem you actually see — don't stuff it preemptively.
When a result disappoints, diagnose before rewriting: wrong subject → sharpen layer one; right subject, wrong vibe → swap the style or light words; right everything plus junk → extend the negative prompt. Change one layer per run so you learn what each word does. Ten minutes of this and you'll have a personal prompt template that works across every idea.