Most people open an AI art tool, type three words, get something weird, and quit. The gap between a so-so result and one you'd actually print isn't the model — it's the prompt. Here's how to write prompts that give you something usable on the first or second try instead of the twentieth. If you'd rather experiment as you read, you can start generating here and follow along.
Describe the scene, not the vibe
New users type "cool dragon." The tool has no idea what "cool" means to you, so it guesses. Instead, describe what a camera would see: a dragon perched on a broken clock tower at dusk, scales catching orange light, fog rolling over the rooftops. Concrete nouns and spatial details do more work than adjectives. You're not writing poetry — you're giving instructions. Every placeholder word like "cool," "nice," or "epic" is space you could have filled with something the model can actually picture.
Name the style and the medium
Style words change everything. "Watercolor on cold-press paper," "35mm film photo, slight grain," "flat vector illustration with thick outlines" — each pushes the model toward a different world. Pick one medium and commit to it. Mixing three styles in one prompt usually produces mush. If you want a photo look, say photo. If you want a painting, say painting and name the kind. Naming the medium is often more powerful than naming the subject, because it changes the entire rendering approach rather than just the content.
Control light like a photographer
Light is the single biggest lever. Golden hour, overcast diffused light, harsh noon sun, neon reflections on wet asphalt, a single candle from the left. Add where the light comes from and how soft it is. Prompts with a clear light source beat prompts without one almost every time, because lighting is what makes a scene feel real rather than rendered. Photographers spend their careers chasing light — borrow their vocabulary and your images improve immediately.
Set the framing
Tell the tool how close you are. "Close-up portrait, shallow depth of field," "wide establishing shot," "top-down flat lay," "eye-level medium shot." Framing words stop the model from inventing a composition you didn't want. If you're making a thumbnail or a hero image, mention the aspect ratio and leave breathing room in the area where text will go. Describing the camera angle and distance is one of the fastest ways to move from a random result to a deliberate one.
Add what you don't want
Most tools let you list things to avoid: no text, no watermark, no extra fingers, no busy background. Use it. Also keep negatives short — a wall of exclusions tends to confuse the model more than it helps. Three or four targeted ones usually do the job. The most common useful negatives are text, watermarks, and duplicate limbs, since those are the artifacts people notice first and complain about most.
Iterate with small changes
Don't rewrite the whole prompt when something's off. Keep everything, change one thing — the lighting, the framing, the medium — and regenerate. That way you learn which word caused the result. People who treat prompting like a conversation, adjusting one variable at a time, get to a good image far faster than people who start over each round. This is slow at first and fast overall, because the knowledge sticks.
Save your winners
When a prompt finally works, copy it somewhere. Build a small library of prompts that produce consistent results in the style you like. Over a few weeks that library becomes more valuable than any single image, because you can reuse the structure for new subjects and skip the guessing game entirely. Many people keep a plain text file sorted by subject — portraits, products, landscapes — and paste the closest match as a starting point.
Match the resolution to the use
Think about where the image goes before you generate. A social post needs different proportions than a desktop wallpaper or a print. Generating at the wrong shape and cropping later throws away details and often cuts off the best part of the composition. Decide the destination first, then describe the framing and ratio to match it. It's a small habit that saves a lot of rework.
Keep prompts readable
You don't need comma-separated keyword soup. A readable sentence or two — "a weathered fisherman mending a net on a wooden dock, overcast morning light, 35mm photo, shallow depth of field" — communicates more clearly than a jumble of tags, and it's far easier to edit one piece at a time. Write for a person, and the model tends to follow along.
That's the whole method: describe the scene, lock a style, control the light, set the frame, exclude a few problems, and change one thing at a time. Try it on your next idea — head to the generator and put a real prompt through its paces. The difference on the first attempt is usually obvious.