How to write AI image prompts that don't fall apart: six jobs, five real examples

Most AI images look generic for one reason: the prompt left the decisions to the model. So the model picked the most average option for everything it wasn't told. Here are the six jobs a prompt has to cover, and five real prompts that cover them well, each shown with the image it made.

Why prompts go flat

A weak prompt is rarely wrong. It is just quiet about the things that matter.

Leave out the light and you get flat, even light. Leave out the framing and the subject lands dead centre. Leave out the background and you get something vaguely pleasant and forgettable.

The fix is not a longer prompt. It is a prompt where every phrase has a job.

The six jobs

• Subject. Who or what, doing what, made of what. "A woman in a structured black suit" beats "a stylish woman".
• Style or medium. Photo, line art, 3D render, risograph print. Put it early. It moves the picture more than almost anything.
• Lighting. The one people skip. Name the source, the direction and the quality: "single direct flash from camera-left".
• Camera. Lens, depth of field, angle. "85mm" and "shallow depth of field" mean something specific to a model.
• Composition. Where things sit, how much empty space, and the aspect ratio.
• Limits. What must not appear. Plain sentences like "no text, no watermark" work in most tools.

On top of all six sits a seventh question: which model are you writing for? More on that near the end.

Five prompts, taken apart

All five come from Promptclay, a place where AI creators share prompts, remix each other's with the credit kept on, and keep the good ones in collections. Disclosure: it's my project. These five were collected into the library by its library accounts, and each page credits them.

Each prompt is quoted exactly. Words in {curly braces} are the parts meant to be swapped.

Example 1: light and lens do the heavy lifting

Prompt page: https://promptclay.com/prompts/studio-fashion-portrait-with-high-contrast-4vyxbn | Made with Nano Banana Pro in Gemini | Credited to @library.social

Studio fashion portrait of a woman in a structured black suit, white shirt, and black tie against a matte red backdrop. Hands in pockets, relaxed posture, direct gaze. Single direct flash from camera-left creating sharp shadows and strong contrast. Clean commercial fashion style with crisp detail. 85mm portrait lens. Aspect ratio 4:5.

Result from the prompt page: Studio Fashion Portrait with High Contrast on Promptclay (credit @library.social)

• Under 60 words, and each sentence covers a different job: wardrobe, pose, light, style, lens, ratio.
• "Matte red backdrop" sets the colour and tells the model the surface won't shine, so the flash shadows land clean.
• The light reads like a line on a photographer's call sheet. One source, one direction, one effect.
• "Aspect ratio 4:5" is written out, because Gemini has no ratio menu.

Example 2: open the right parts, lock the rest

Prompt page: https://promptclay.com/prompts/golden-hour-slatted-shadow-portrait-0nmnqc | Made with Nano Banana Pro in Gemini | Credited to @library.social

A close-up cinematic portrait of a young woman with natural beauty, {eye_color} eyes, and slightly parted lips. Her wavy, {hair_color} hair is loosely tied back with soft, messy tendrils framing her face. She is wearing a {clothing}. Golden hour sunlight streams through a window, casting sharp, dramatic slatted shadow patterns diagonally across her face and neck. High-contrast chiaroscuro lighting, warm golden tones with deep shadows, visible skin texture and subtle freckles, shot on 85mm f/1.4 lens, shallow depth of field, soft muted background, moody editorial photography.

Result from the prompt page: Golden Hour Slatted Shadow Portrait on Promptclay (credit @library.social)

• Eye colour, hair colour and clothing are open. The light, the shadow pattern and the lens are locked. The light is the whole point of this image, so that split is right.
• "Slatted shadow patterns diagonally across her face" is a physical instruction the model can draw. "Moody shadows" is not.
• "Visible skin texture and subtle freckles" pushes back on the over-smoothed look many models default to.

Example 3: count things out loud

Prompt page: https://promptclay.com/prompts/simple-hand-drawn-koala-mascot-s6d01t | Made with GPT Image 2 in ChatGPT | Credited to @library.game.art

Create a simple centered children's-book illustration of a cute full-body koala character on a plain white background. The koala has soft {fur_color} textured fur with a rough crayon or felt-tip marker fill, thick irregular black hand-drawn outlines, a large rounded head, two oversized round ears, and a small upright oval body. Use exactly two ears, each with one solid black inner ear patch; no pink inner-ear lines. The face has exactly two round white eyes with black circular pupils, one large vertical oval black nose in the center, and one small curved smiling mouth below it. The arms hang down at the sides with rounded mitten-like paws, and the short legs end in small three-toed feet. Keep the style naive, flat, slightly wobbly, and hand-drawn, like a child-friendly mascot sticker; no shadows, no scenery, no text, no accessories, no extra characters.

Result from the prompt page: Simple Hand-Drawn Koala Mascot on Promptclay (credit @library.game.art)

• "Exactly two ears", "exactly two round white eyes", "one large vertical oval black nose". Each one closes a door the model might otherwise wander through.
• The limits are concrete: "no shadows, no scenery, no text, no accessories, no extra characters". Worth keeping as a stock line for mascots and stickers.
• The medium is described by its tools ("crayon or felt-tip marker fill"), so you get a texture, not just a label.

Example 4: style and camera in four words

Prompt page: https://promptclay.com/prompts/3d-isometric-floating-home-office-illustration-h6keby | Made with Nano Banana Pro in Gemini | Credited to @library.product

A 3D isometric pastel home office scene where objects appear gently lifted, featuring a modern desk with a laptop slightly hovering, plants slowly floating upward, shelves detached from the wall, and a coffee cup resting in mid-air without spills. Smooth rounded surfaces, soft matte materials, {color_palette} palette, minimal and clean details, subtle depth, soft ambient lighting, gentle shadows that hint at weightlessness, peaceful futuristic mood, illustration-style render, modern UI-friendly composition.

Result from the prompt page: 3D Isometric Floating Home Office Illustration on Promptclay (credit @library.product)

• "3D isometric" up front decides the viewpoint and the style at once.
• Weightlessness is broken into things the model can draw: a hovering laptop, plants floating up, a cup "without spills". That last bit stops a dramatic splash.
• "Modern UI-friendly composition" is handy for landing pages and empty states. It tends to mean clean edges and room around the subject.

Example 5: a reference photo and a negative list

Prompt page: https://promptclay.com/prompts/cinematic-spring-cherry-blossom-portrait-1zf7kd | Made with GPT Image 2 in ChatGPT | Credited to @library.avatars

Use the uploaded portrait only to preserve the person’s facial identity, hairstyle, hairline, skin tone, and natural presence. Create a realistic cinematic portrait under a {tree_type} at {time_of_day}, warm golden backlight, soft silhouette, falling {petals}, the person gently holding a red rose near the face, quiet romantic atmosphere, shallow depth of field, natural skin texture, realistic body proportions, soft spring outfit, elegant composition, 3:4 vertical photo, high-quality DSLR look, no text, no watermark

Negative Prompt:

anime, cartoon, plastic skin, over-smoothed face, fake beauty filter, distorted face, wrong identity, bad hands, extra fingers, unnatural body, harsh lighting, low resolution, overexposed, messy background, logo, watermark, text

Result from the prompt page: Cinematic Spring Cherry Blossom Portrait on Promptclay (credit @library.avatars)

• The first sentence scopes the upload: keep the face and hair, take nothing else from the photo.
• The negative list reads like a catalogue of portrait failures: plastic skin, wrong identity, extra fingers. A solid starting list for your own portraits.
• "No text, no watermark" also appears in the main prompt. Chat-style tools don't always treat a "Negative Prompt:" heading as special, so saying it twice is cheap insurance.

Writing for Gemini vs ChatGPT

Same words, different results. These notes come from Promptclay's model pages for Gemini and ChatGPT image:

• Gemini: say the ratio in words. Upload the photo first, then paste. Say what to keep ("same face, same pose"). One change per message.
• ChatGPT image: ask for the ratio in words too ("square", "wide 16:9"). Put any words you want rendered in quotes. Name colours plainly. Edit in the same chat.
• Midjourney flags like --ar 9:16 mean nothing to other tools. Strip them and write the ratio out.

Takeaways

• Give every phrase a job: subject, medium, light, camera, composition, limits.
• Light is the most skipped job and the biggest mood lever.
• Count anything that has to be exact.
• Lock the parts that define the look. Open only the parts that should change.
• Write for the model you are using.

Try one swap

Pick a prompt above, change only the lighting line, and run it again. Then change only the lens. One swap at a time is how you find out which words a model listens to.

If your version turns out well, you can post it on Promptclay as a remix. The original creator's name stays on it automatically, and you can keep your favourites in a collection. Copying any prompt is free, no account needed.

Prompts are better shared. Post what your swap taught you, so the next person starts where you stopped.