How do you write AI image prompts that work?

Published October 8, 2026 · Updated October 8, 2026

Describe the picture in sentences: its purpose, subject, light and exact words. To edit your own photo, name one change and list what stays.

A good AI image prompt describes a scene in sentences, not a pile of keywords: what the picture is for, who is in it, where they are, the light, and the exact words to print. To edit your own photo, say what changes and list everything that stays. For a video, say what moves.

What we measured ourselves
2026-10-08 → 2026-10-08
9 to 268 wordswords in the lemonade poster prompt, from the one-line first try to the finished poster

We wrote the same idea four times, each version adding one kind of decision, and drew every version with GPT image. The table below shows what each step adds and how long the prompt got, for the lemonade poster and for a manga cover.

Source: N33 engines, staff run
The four steps of a prompt, and how long each one was
StepWhat the step addsLemonade posterManga cover
1. Minimalonly the product and the person9 words3 words
2. Structuredthe brand name, the action and the headline37 words15 words
3. Proa strict palette, the layout and typography rules163 words71 words
4. Polishedthe camera, exact counts and every line of text268 words236 words

To see what each part of a prompt does, we wrote two ideas, a lemonade poster and a manga cover, at four levels each and drew every step on our own engines on 8 October 2026. The shortest prompt was 3 words, the longest 268. Short prompts do not fail: the model fills every gap on its own, so you get a picture, just not yours. Each step after that takes one decision back from the model. The full ladders, with every picture, are on our AI image prompts page.

What we measured ourselves
2026-10-08 → 2026-10-08
21pictures drawn in our first batch for this guide and the prompt hubs

One prompt was refused by moderation in the Studio and went through after we rewrote it. One generation failed for a transient reason and succeeded on a retry. Every picture in this guide comes from this run.

Source: N33 engines, staff run

What should a good AI prompt include?

Six parts cover most prompts: what the result is for, the subject and what it does, the setting, the light and camera, the style, and any exact text. For an edit, add what stays. For a video, add what moves.
The parts of a prompt, with examples from our own ladders
PartExampleWhat goes wrong without it
What it is forA finished vertical advertising posterThe model guesses the format, often a plain photo with no layout
Subject and actionA girl in a white bucket hat sips through a striped strawA generic person in a generic pose
SettingAgainst a flat bubblegum-pink wallA busy background that fights the subject
Light and cameraHard on-camera flash, shot from slightly below on a 35mm lensFlat, even light that reads as stock
Style and paletteStrict palette: bubblegum pink, lemon yellow, mint green, whiteColours wander, and every version looks different
Exact textThe words "SOUR SUN" in quotes, once, across the topInvented words or misspelled lettering
What stays (edits)Keep her face, skin tone, hair, lighting and framing unchangedThe face changes along with the edit
What moves (video)The can tips toward the camera, slow push inRandom motion, or none at all

How do you write a prompt for an AI image?

Start with the finished thing you want, such as a poster or a magazine cover. Then fix the palette, place each section on the canvas, set the camera and light, and put every word that must appear in quotes. Count what matters: one can, two hands.

The order matters less than the habit of deciding. Our lemonade poster went from one line to a brief. The second step named the brand, the girl and the headline. The third fixed a strict palette, placed the photograph in the upper two-thirds and the slogan on a yellow panel, and added typography rules. The fourth set a lens and a camera angle, gave exact counts, and added two stickers and a button. Google gives the same advice for Nano Banana in plain terms: stop writing tag lists like "cool car, neon, city, night" and describe the scene instead.

A winking girl in a bucket hat holds a pink lemon can to the lens beside flying lemon slices, 3D SOUR SUN lettering, stickers and a black button
The last step of our lemonade poster ladder, drawn with GPT image. Every step of this ladder and the others is on the prompt hub.
PromptGPT image
A finished vertical 4:5 advertising poster for a fictional sparkling pink lemonade brand. Loud, glossy, maximum saturation, hard-flash look. STRICT PALETTE: bubblegum pink, lemon yellow, mint green, white, black outlines. No other colors.

PHOTOGRAPH in the upper two-thirds: a laughing girl in a white bucket hat and a mint tee against a flat bubblegum-pink wall, one eye squeezed shut from the sour taste, lips puckered around a pink-and-white striped straw. She holds a tall slim can toward a 35mm lens shot from slightly below; cold drops run down the can, three lemon slices hang in the air beside it. Hard direct flash: hot highlights on cheeks, a sharp black shadow on the wall behind her shoulder.

GRAPHICS: "SOUR SUN" set enormous across the top on one line in a chunky extruded 3D rounded display face: lemon-yellow faces, mint-green extrusion, thick black outline, white gloss streak. The lettering overlaps the brim of her hat.

TWO glossy sticker badges with soft drop shadows: a mint starburst reading "NEW" in white caps, and a yellow lemon-shaped badge reading "zero sugar" in black lowercase. Exactly two stickers.

LOWER THIRD on a solid lemon-yellow panel: "PUCKER UP, SUMMER" in black bold caps, centered; below it one line of small black text: "Real lemons. Pink fizz. Zero chill."; below that a black pill button with yellow text "Find a cooler".

Typography rules: English, spelled exactly as given (the brand is S-O-U-R  S-U-N), at most two typefaces, straight baselines, no doubled glyphs, no gibberish, "SOUR SUN" appears exactly once. Exactly one can, one straw, two hands, two stickers, one button. No watermark, no extra logos.
A mid-century screenprint travel poster of sharp Norwegian peaks over a fjord, LOFOTEN in black serif and a black tagline band at the bottom
A travel poster drawn with GPT image and written the same way: the finished format first, then the layout, the illustration, a held palette and the typography rules.
PromptGPT image
A vertical 4:5 travel poster filling the frame edge to edge, with a thin bone-cream inner border. Style: mid-century screenprint, flat layered color separations, clean graphic shapes, light stipple texture, visible paper grain as if printed on aged uncoated paper and photographed flat.

Layout: the name LOFOTEN at upper left in a large high-contrast serif, black ink, one line. Under it NORWAY in small widely spaced black caps. A solid black band across the bottom carries the tagline WHERE THE LIGHT STAYS UP in bone-cream spaced caps, centered.

Illustration in the lower three quarters: sharp green-black peaks rising straight out of a still fjord, a row of red wooden fishing cabins on stilts along the shore, drying racks of fish beside them, a midnight sun low over the water in a banded sky from deep teal through coral to pale gold, a single small rowing boat leaving a white wake.

Palette held to teal, coral, pale gold, red ochre, slate green, bone cream and black.

Typography: exactly two typefaces, straight baselines, correct spelling, Latin characters only. One place name, one country line, one tagline, each once. No people, no watermark.

How do you write a prompt to edit your own photo?

Name the one change, then list what stays: the face, skin tone, hair, pose, light and framing. Describe people by how they look rather than as "she" or "him", and make one change per turn instead of several at once.

Edits tend to fail the same way: you ask for a new jacket and get a new person. Two habits prevent most of it. First, write the keep list every time, even on a later turn, because each turn can drift a little further from the original. Second, ask for one change, check it, then ask for the next. Google's own advice for editing is to be direct and specific, as in "change the man's tie to green". Use your own photo or a generated one, never a picture of someone who has not agreed to it.

The same woman with two pale pink rose petals and water droplets on her cheekbone and a soft rose blush, face and light unchanged
An edit made with Flux Kontext on our own generated window light portrait: petals, droplets and a blush added, the face, closed eyes, hair and light kept by name.
PromptFlux Kontext
Add two pale pink rose petals resting on her cheekbone and small clear water droplets beaded across the petals and her skin, one droplet leaving a short wet trail. Brush a soft rose blush high on the cheekbone. Keep her exact facial features, skin tone, closed eyes, hair, lighting and framing unchanged.
The same platinum haired woman on the pastel street, now in a mustard corduroy jacket over a cream turtleneck, face and pose unchanged
Another Flux Kontext edit in the same grammar: two named pieces of clothing change, and the person, the bicycle, the street and the light are listed as kept.
PromptFlux Kontext
Change the woman's faded denim jacket to a fitted mustard-yellow corduroy jacket with brown buttons, and change her white tee to a thin cream turtleneck. Keep the same pose, the same mint-green bicycle, the same street and lighting, and keep the woman with short platinum hair exactly as she is, while keeping the same facial features, freckles and expression.

How do you write a prompt for an AI video?

Describe what moves, not the whole picture again. Keep to one scene and one camera move per clip, and put the action in the first sentence. When you start from a still, the still becomes the first frame, so the prompt only needs the motion.

Video prompts are short for a reason: the later an instruction sits in a long prompt, the less reliably it is followed, so the subject and its action go first and the camera move second. One Seedance guide puts the budget at 60 to 100 words. To try one, write it on our Seedance video generator page or turn a still into a clip with image to video.

How do prompts differ between Nano Banana, GPT image, Flux, Grok and Seedance?

Nano Banana and GPT image take long, structured briefs and handle text well. Flux Kontext is for a precise change to a photo you already have. Grok Imagine likes short, direct prompts. Seedance makes video: one subject, one action, one camera move.
How to write for each model
ModelBest forHow to write for itEdits and references
Nano Banana, the image model in the Gemini appEdits of your own photo, seasonal portraits, postersFull sentences: subject, composition, action, location, styleBlends up to 14 images, depending on where you use it
GPT imagePosters, covers and layouts with a lot of textThe intended use first, then the scene, the subject, the details and the constraintsRefer to each input image by number and say how they combine
Flux KontextOne precise change to an existing photoAn instruction: change, replace, add, keepName the person by how they look and list what stays
Grok ImagineFast stylised stills and quick animationShort and direct: the subject plus a styleAn uploaded picture becomes the first frame; ratios from 1:1 to 9:16
SeedanceVideo from text or from a pictureSubject, action, scene and one camera move, in 60 to 100 wordsStart from a still when the look must stay fixed

From the model makers' own guides and tutorials published in the last twelve months, and our own run on 8 October 2026.

How do you get AI to spell text correctly?

Put every word in quotes or capitals, say where it goes and how often it appears, spell hard words letter by letter, and allow two typefaces at most. Pick a model known for text, such as GPT image or Nano Banana, and keep small print rare.

Our manga cover ladder is the hard case: Japanese lettering, a Latin subtitle and a starburst on one cover. The last step named each piece of text, where it sits and how many times it appears, and asked for everything else on the cover to be too small to read, so the model has no reason to invent words to fill the space. Small text is still the weak spot of every model, and Google says as much about its own.

A red, yellow and black manga cover with halftone print texture: a grinning girl with a black bob takes a hard turn on a red sport motorbike
The last step of our manga cover ladder, drawn with GPT image: the masthead, the subtitle and the starburst each named once, everything else too small to read.
PromptGPT image
A finished weekly manga magazine cover, portrait 4:5. Printed look: visible halftone dots in the midtones, slight ink misregistration on the reds, faint newsprint paper texture over everything.

STRICT PALETTE: signal red, acid yellow, black, cream newsprint. No other colors.

ILLUSTRATION fills the cover: a girl with a short black bob and a torn acid-yellow scarf leans hard into a turn on a red sport motorbike, knee slider nearly touching the road, grinning with her teeth showing, eyes locked on the reader. Hard cel shading, crisp black outlines, heavy spot blacks. Dense black speed lines radiate from a vanishing point behind her helmet-free head. Sparks spray from the footpeg at the lower right in red and yellow dots. Low angle, we look slightly up at her.

LAYOUT:
MASTHEAD: the two kanji 疾風 rendered once, correctly formed, huge, top left, in black brush lettering with a red offset shadow, partly overlapped by her scarf.
Under it, in small clean red Latin capitals: GALE RIDER.
Top right: a yellow starburst with a black outline reading NEW SERIES in black capitals, two words exactly.
Bottom edge: a solid black band with tiny cream decorative type that is too small to read.

TYPOGRAPHY RULES: at most two typefaces, Latin text on straight baselines, exact spellings (疾風 once, GALE RIDER, NEW SERIES). No speech bubbles, no panel borders, no other readable text. Exactly two arms, one bike, two wheels. No watermark.
Common prompt mistakes and how to fix them
MistakeKindFix
A list of tags: neon, city, night, ultra detailedImageWrite the scene as sentences
No purpose statedImageOpen with the finished thing: a poster, a cover, a product shot
Long text in a small spaceImageFewer words, bigger type, each line quoted and placed
Several changes in one editPhoto editOne change per turn, check it, then the next
No keep listPhoto editName what stays (face, skin tone, hair, pose, light, framing) and repeat it every turn
"She" and "him" instead of a descriptionPhoto editDescribe the person by look: the woman with short platinum hair
Several camera moves in one clipVideoOne move per clip, and a new clip for the next move
Describing the still again in an image to video promptVideoThe still is the first frame: write only what moves

More prompts to copy and adapt, each with the picture it made, are on our AI image prompts page. To try one now, open the GPT image generator or the Nano Banana image generator.

Questions and answers

Do negative prompts work?

Partly. A short exclusion list such as "no watermark, no extra text" works on many models, and OpenAI's own prompting guide uses them. For anything bigger, describe the clean state you want instead: "an empty street" rather than "no cars".

Can you use several reference images in one prompt?

Yes, on most image models. Give each image a job and refer to it by its number: "the person from image 1, the coat from image 2". Google says Nano Banana Pro can blend up to 14 images, depending on where you use it.

What aspect ratio should you use for short videos and social posts?

Vertical 9:16 for short video feeds and stories, 3:4 or 1:1 for feed posts, 16:9 for widescreen. Write the ratio into the prompt or pick it in the settings. When you turn a still into a clip, draw the still in the same ratio, because it becomes the first frame.

Is it free to try these prompts on N33 AI?

Writing and polishing a prompt in the chat is free and needs no sign-in. Drawing pictures and clips in the Studio uses energy, and a free account comes with some to start.