Type "a cool logo for my coffee shop" into any AI image tool and you will get something: a brown circle, a steaming cup, some mangled text that almost says coffee. Technically an image. Practically useless.
Now try this instead:
minimal flat logo mark for a coffee shop, single steaming cup icon inside a circle, thick clean lines, warm brown and cream palette, centered on white background, no textSame tool, same ten seconds. Completely different result. The model did not get smarter between the two attempts. The prompt did.
That gap is what this guide closes. It covers one formula that works across ChatGPT, Midjourney, FLUX, Gemini, Firefly, Leonardo and Ideogram, why each piece of it works, 50 prompts you can copy, 200 keywords that actually change the output, and the mistakes that quietly ruin otherwise good prompts.
There is a one-page PDF cheat sheet of everything in this guide at the bottom, free to download and keep next to your keyboard.
Why Most Prompts Fail
Almost every disappointing AI image comes from one of four prompt problems, and none of them are about vocabulary.
The prompt is a wish, not a description. "A cool logo" describes how you want to feel about the result. The model needs to know what to draw: what object, what style of mark, what colors, what background. Models generate averages of everything that matches your words, and the average of "cool" is mush.
The subject is buried. Ask for "a beautiful cinematic moody atmospheric shot, golden light, incredible detail, of a lighthouse" and you have spent your strongest position, the start of the prompt, on decoration. The lighthouse arrives as an afterthought, and it shows.
The prompt contradicts itself. "Minimalist detailed intricate simple design" asks the model to go in two directions at once. You get a compromise nobody wanted.
It says what you don't want, in the wrong place. Writing "no people" inside a normal prompt often adds people. Models respond to the concepts you mention, not the grammar around them. Mention people, get people. Exclusions belong in a negative prompt, and we cover exactly how those work later in this guide.
The ImgCruncher Prompt Formula
Every strong image prompt is built from the same twelve blocks. You will rarely use all twelve at once; the skill is knowing which ones your image needs.
Written out, the skeleton looks like this:
[subject] [doing what] in [environment], [composition], [lighting],
shot on [camera or lens], [style], [mood], [color palette],
[quality], [aspect ratio] | negative: [what to exclude]Order matters more than most people expect. Models weigh the beginning of a prompt more heavily than the end, so the formula puts the things that define the image first and the polish last.
The Twelve Building Blocks, One by One
1. Subject
The single most important phrase in the prompt, and it should come first. Be specific in ways that change pixels: "a dog" could be ten thousand images, "an elderly golden retriever with a grey muzzle" is one. Specificity is not length; it is picking details the model can actually draw.
2. Action
A subject doing something beats a subject existing. "A chef" gives you a stock-photo pose. "A chef flambeing a pan, flames leaping" gives you energy, a light source, and a story in one clause. Verbs are the cheapest drama available.
3. Environment
Where the scene happens sets background, lighting logic and mood at once: "in a cramped Tokyo alley kitchen at night" has already decided half your image. If you skip it, the model picks a generic studio void or invents something random.
4. Composition
How the frame is arranged: close-up, full body, low angle, overhead flat lay, rule of thirds, centered symmetry, negative space. Without direction, models default to a centered medium shot of everything. One composition phrase breaks that habit.
5. Lighting
The highest-leverage block in the whole formula. "Golden hour", "soft window light", "studio softbox", "neon glow" each transform an image more than any style keyword, because lighting decides color, shadow, depth and mood simultaneously. If you only add one Craft block, make it this one.
6. Camera
Camera language is compressed instruction: "85mm lens" implies portrait framing and background blur, "wide angle" implies drama and context, "macro" implies extreme detail, "drone shot" implies scale. You are not naming hardware, you are borrowing the look that hardware produces.
7. Style
The medium and rendering approach: photorealistic, film photography, watercolor, flat vector, 3D render, pixel art, anime. Pick one. Stacking five styles ("photorealistic watercolor 3D anime") is the fastest way to get a muddy hybrid.
8. Mood
The emotional temperature: cozy, eerie, triumphant, melancholic, playful. Mood words steer dozens of small decisions, from color warmth to how much shadow the model allows. One mood word is plenty.
9. Colors
Name a palette and the whole image obeys: "muted earth tones", "teal and orange", "pastel palette", "high contrast black and white". This is the block that makes a set of images look like they belong together, which matters the moment you generate more than one.
10. Quality
Tags like "highly detailed", "sharp focus" and "professional photograph" nudge the model toward its better outputs. They help modestly and only at the end of a prompt that already works. They cannot rescue a weak subject, which is why they sit near the top of the pyramid, not the bottom.
11. Aspect Ratio
The shape of the canvas changes the composition itself: 16:9 gives the model room for environment, 4:5 pushes it toward the subject, 1:1 centers everything. Decide where the image will live before you generate, not after. Midjourney uses --ar 16:9, most other tools have a setting or accept it written in plain words.
12. Negative Prompt
The list of things the model should avoid: blur, watermarks, extra fingers, text. Powerful where supported, useless where not, and frequently misused. It gets its own section below.
How the AI Actually Reads Your Prompt
Understanding three mechanics will improve your prompts more than any keyword list.
Your words become coordinates, not instructions. The model does not read your prompt like a person. It converts the text into a numerical representation of meaning, then generates an image that lands near those coordinates. This is why "no elephants" fails: the word elephants pulls the coordinates toward elephants regardless of the "no". It is also why word choice beats word count; each phrase either moves the coordinates somewhere useful or adds noise.
Early words carry more weight. Attention is not distributed evenly across a prompt. The opening tokens anchor what the image is about, which is why the formula leads with subject and action and why a prompt that opens with ten adjectives produces beautiful pictures of nothing in particular.
Everything you mention tends to appear. Models are additive. Casually mention rain in a mood clause and you will get wet streets. This cuts both ways: it is why rich descriptions work, and why sloppy ones backfire.
The Same Prompt Is a Different Prompt in Every Model
The formula holds everywhere, but each model has a personality: what it excels at, how it likes prompts phrased, and whether it supports true negative prompts.
| Model | Strongest at | Prompt style that works | Negative prompts |
|---|---|---|---|
| ChatGPT (GPT Image) | Instruction following, text in images, edits by conversation | Full sentences, as if briefing a designer | Written in plain words ("without any text") |
| Midjourney | Artistic look, stylized beauty, strong defaults | Short evocative phrases, comma-separated | Yes, via --no parameter |
| FLUX | Photorealism, prompt accuracy, natural skin and hands | Detailed natural language descriptions | Depends on the app hosting it |
| Gemini | Conversational editing, combining reference images | Plain sentences, then follow-up corrections | Written in plain words |
| Adobe Firefly | Commercially safe output, Photoshop integration | Clear descriptions plus its style controls | Via exclusion settings, not prompt syntax |
| Leonardo AI | Game assets, concept art, consistent character work | Keyword-rich prompts with model presets | Yes, dedicated negative prompt field |
| Ideogram | Text rendering: logos, posters, signage | Describe the design and quote the exact text | Limited; keep prompts positive |
The practical rule: write the prompt once using the formula, then adapt the delivery. For Midjourney, tighten it into phrases and move exclusions into --no. For ChatGPT and Gemini, relax it into sentences and iterate conversationally. For Ideogram, put the exact text you want rendered in quotation marks. If you are unsure which tool fits which job, our field guide to AI tools by task breaks it down.
Building a Prompt From Scratch, Step by Step
Here is the formula applied in five passes, starting from the laziest possible prompt.
Why each pass earns its place:
- "a cat" leaves every decision to the model: breed, pose, setting, light, style. You will get the internet's average cat.
- Scene: "a ginger cat stretching on a sunlit windowsill" locks subject, action and environment. The image now has a story and a light source.
- Craft: "low angle close-up, soft morning window light, 50mm lens" frames the shot like a photographer instead of a security camera.
- Look: "warm cozy mood, film photography style, muted amber tones" makes it feel like something rather than a document of a cat.
- Polish: "highly detailed, 3:2 aspect ratio, negative: blur, extra limbs" shapes the output file and trims common failure modes.
Notice what did not happen: no wall of fifty keywords, no mystical incantations. Each pass answered one question the model would otherwise have answered randomly.
Generate after pass 2, not after pass 5. Seeing what the model does with the core tells you what the later passes actually need to fix. Prompting is a conversation, not an exam with one attempt.
50 Prompt Examples You Can Copy
Every prompt below follows the formula and works as a starting point in any major model. Swap the bracketed pieces for your own subject and adjust the aspect ratio to where the image will live.
Portraits (5)
portrait of an elderly fisherman with deep wrinkles, weathered face lit by golden hour sun, 85mm lens, shallow depth of field, film photography, warm muted tones, 4:5candid street portrait of a young woman laughing, neon signs reflected in rain behind her, 35mm lens, night photography, cinematic teal and magenta, 4:5studio portrait of a dancer mid-spin, fabric flowing, single dramatic rim light on black background, high contrast black and white, sharp focus, 4:5environmental portrait of a potter at her wheel, clay-dusted hands in focus, soft window light from the left, earthy palette, documentary style, 3:2close-up portrait of a child blowing a dandelion, backlit seeds glowing, macro detail on seeds, dreamy soft focus background, pastel palette, 1:1
Product photography (5)
a matte black wireless earbud case on wet slate stone, water droplets, single softbox from above, dark moody commercial photography, sharp macro detail, 1:1a glass perfume bottle on a mirrored surface, caustic light patterns, pastel gradient background, high-key studio lighting, luxury advertising style, 4:5a leather backpack on a wooden cabin table beside a window, morning light, steam from a coffee mug nearby, warm lifestyle photography, 3:2a running shoe frozen mid-splash in colored powder, high-speed photography look, black background, vivid orange and blue, dramatic side light, 16:9a skincare jar surrounded by fresh citrus slices and leaves, overhead flat lay, bright even daylight, clean white background, fresh green and yellow palette, 1:1
Landscapes and nature (5)
a lone lighthouse on a cliff during a storm, huge waves crashing below, dramatic dark clouds with a single beam of light, wide angle, cinematic, desaturated blues, 16:9terraced rice fields at sunrise, morning mist in the valleys, a farmer walking a ridge line, drone shot from high above, lush greens and gold, 16:9a desert of white sand dunes under a full moon, one twisted dead tree, long shadows, minimalist composition with vast negative space, monochrome blue, 3:2autumn forest stream with slow silky water, long exposure, fallen red maple leaves on wet rocks, overcast soft light, rich oranges and browns, 3:2northern lights over a frozen lake, a tiny glowing tent on the shore for scale, stars visible, ultra wide angle, deep greens and purples, 16:9
Logos and icons (5)
minimal flat logo mark for a mountain hiking brand, geometric peak inside a circle, thick uniform line weight, two colors only, centered on white, no textapp icon for a meditation app, abstract lotus made of soft gradient waves, rounded square, calm blue to purple gradient, clean and simple, 1:1vintage badge logo for a barbershop, crossed scissors and comb, circular text border reading "TRIM & CO", distressed letterpress texture, black on creammascot logo of a friendly robot head for a coding academy, flat vector, bold outlines, limited palette of navy orange and white, symmetrical, no textmonogram logo combining letters A and V, single continuous line, elegant thin stroke, luxury fashion feel, gold on deep charcoal, centered, 1:1
Flat illustration and vector (5)
flat vector illustration of a remote worker at a desk by a big window, cat on the windowsill, plants everywhere, simple geometric shapes, 4-color palette, generous negative space, 3:2isometric illustration of a tiny smart home cutaway, labeled rooms with glowing devices, soft shadows, pastel tech palette, clean lines, 1:1line art illustration of two hands forming a heart, single continuous line, one accent color, white background, minimal and elegant, 1:1flat illustration of a city street converting to a bicycle lane, people cycling, split scene before and after, bold shapes, optimistic bright palette, 16:9children's book illustration of a fox and a rabbit sharing an umbrella in the rain, soft textured shapes, warm and gentle, watercolor paper texture, 4:5
3D and render styles (5)
3D render of a cute miniature ramen shop diorama, glowing lanterns, tiny steam, soft clay material, isometric view, warm night palette, high detail, 1:1abstract 3D composition of glass and chrome spheres on liquid ripples, studio HDRI lighting, shallow depth of field, iridescent accents, 16:9claymation style character of a grumpy cactus in a tiny pot, visible fingerprints in clay, single soft key light, pastel studio background, 1:1low poly 3D landscape of a mountain valley with a river, flat shading, sunset gradient sky, drifting geometric clouds, minimal, 16:93D product render of a transparent mechanical watch exploding into parts, technical lighting, dark background, precise reflections, 4:5
Architecture and interiors (5)
modern glass cabin in a snowy pine forest at dusk, warm light glowing from inside, long exposure snowfall, architectural photography, cool blue with amber accents, 3:2sunlit reading nook with floor-to-ceiling bookshelves, worn leather armchair, dust motes in a light beam, cozy interior photography, warm earth tones, 4:5brutalist concrete museum atrium, dramatic skylight shadows, one person for scale, symmetrical one-point perspective, monochrome, 4:5Mediterranean courtyard with white walls and blue shutters, bougainvillea, harsh midday sun and crisp shadows, travel photography, saturated, 3:2futuristic vertical garden skyscraper, plants cascading from every balcony, golden hour, low angle looking up, hopeful sci-fi, green and gold, 9:16
Food photography (5)
rustic sourdough loaf torn open showing airy crumb, flour dust in the air, dark moody backdrop, side window light, dark food photography, 4:5ramen bowl with soft egg and chashu, steam rising, chopsticks lifting noodles, overhead 45 degree angle, warm izakaya lighting, rich and glossy, 4:5colorful smoothie bowls on a marble table, overhead flat lay, scattered berries and coconut, bright natural daylight, fresh vibrant palette, 1:1chocolate lava cake being cut open, molten center flowing, extreme close-up, dramatic spotlight, deep browns with gold, slow motion feel, 4:5street tacos on a paper plate, lime wedges and radish, food truck bokeh lights behind, handheld documentary style, warm evening tones, 3:2
Fantasy and concept art (5)
ancient library inside a hollow tree, spiral staircases of roots, floating glowing books, a tiny librarian gnome, painterly fantasy concept art, emerald and amber, 4:5sky whale migrating through clouds at sunset, small airships escorting it, epic scale, volumetric god rays, cinematic concept art, coral and violet, 16:9abandoned robot overgrown with wildflowers in a meadow, a bird nesting in its shoulder, melancholic and peaceful, soft impressionist light, 3:2underwater city in a giant glass dome, whale silhouette passing above, bioluminescent streets, deep blue with cyan glow, wide establishing shot, 16:9a knight's armor stand in a sunlit ruined chapel, ivy through broken stained glass, colored light patches on stone, quiet after-the-battle mood, 4:5
Social media and marketing graphics (5)
bold announcement background with abstract flowing 3D gradient shapes, deep purple to electric blue, dark clean layout with empty center space for text overlay, 1:1flat illustration banner of diverse hands raised holding lightbulbs, optimistic startup energy, simple shapes, brand-friendly palette of coral and navy, 16:9photo backdrop of a pastel podium stage with soft shadows, empty product pedestal center frame, studio lighting, minimalist, ready for product compositing, 4:5energetic sale graphic of paper confetti exploding from a gift box, high-speed photo style, bright red and yellow on clean white, negative space top third, 9:16calm newsletter header of a desk from above with coffee, notebook and glasses, soft morning light, muted neutral palette, space on the right for a headline, 16:9
Models render text poorly unless they were built for it. For the logo and badge prompts that include words, Ideogram and ChatGPT handle lettering best; in other models, generate the mark without text and add typography yourself afterwards.
200 Keywords That Actually Change the Output
These are not magic words. Each one is shorthand for a visual decision, grouped by the formula block it belongs to. Steal freely, but add them one at a time so you can see what each is doing.
Lighting (25)
golden hour, blue hour, soft window light, overcast diffused light, studio softbox, rim light, backlit, silhouette, candlelight, neon glow, volumetric light, god rays, dappled shade, harsh midday sun, moonlight, bioluminescence, light through fog, lens flare, chiaroscuro, low key lighting, high key lighting, practical lights, firelight, stage spotlight, reflected bounce light
Camera and lens (25)
85mm portrait lens, 50mm lens, 35mm street lens, wide angle, ultra wide, telephoto compression, macro lens, fisheye, tilt-shift, drone shot, aerial view, GoPro POV, shallow depth of field, deep focus, f/1.8 bokeh, long exposure, motion blur, high-speed freeze, double exposure, film grain, anamorphic, handheld documentary, security camera view, periscope view, microscope detail
Composition and framing (25)
close-up, extreme close-up, medium shot, full body shot, establishing shot, low angle, high angle, overhead flat lay, dutch angle, rule of thirds, centered symmetry, golden ratio, negative space, leading lines, frame within a frame, over-the-shoulder, profile view, three-quarter view, from behind, worm's eye view, bird's eye view, panoramic, tight crop, off-center subject, foreground framing
Style and medium (25)
photorealistic, cinematic still, film photography, analog photo, watercolor, gouache, oil painting, impressionist, flat vector, line art, charcoal sketch, ink drawing, 3D render, claymation, stop motion, pixel art, voxel art, isometric, low poly, anime, studio ghibli style, comic book ink, art deco, bauhaus poster, ukiyo-e woodblock
Mood and atmosphere (25)
cozy, serene, dreamy, melancholic, nostalgic, eerie, ominous, tense, triumphant, playful, whimsical, chaotic, tranquil, romantic, lonely, hopeful, mysterious, majestic, intimate, energetic, somber, ethereal, gritty, festive, contemplative
Color and palette (25)
muted earth tones, pastel palette, jewel tones, teal and orange, monochrome blue, high contrast black and white, sepia, warm amber tones, cool desaturated blues, neon palette, candy colors, autumn palette, iridescent, duotone, limited 3-color palette, complementary colors, analogous greens, faded vintage colors, rich saturated color, chalky matte tones, metallic gold accents, blush and sage, midnight palette, sun-bleached, technicolor
Texture and detail (25)
intricate textures, weathered surface, rough brushstrokes, smooth gradients, visible fabric weave, cracked paint, polished chrome, frosted glass, wet surfaces, dust particles, paper texture, film scratches, embossed detail, carved wood grain, brushed metal, soft fur detail, translucent skin, condensation droplets, crumbling stone, glossy ceramic, matte finish, iridescent scales, woven wicker, hammered copper, velvet sheen
Quality and rendering (25)
highly detailed, sharp focus, ultra sharp, 8k, masterpiece, professional photograph, award-winning photography, octane render, unreal engine, ray tracing, physically based rendering, studio quality, magazine editorial, fine art print, hyperrealistic, crisp edges, clean composition, perfectly lit, color graded, HDR, photoreal skin texture, accurate anatomy, coherent perspective, balanced exposure, gallery quality
Negative Prompts: What They Fix and What They Cannot
A negative prompt is a second list the model steers away from. Where it is supported, it is the correct home for every "no" and "without" you are tempted to write.
What negative prompts reliably reduce: watermarks and text artifacts, blur and low-quality outputs, extra fingers and limbs, oversaturation, cluttered backgrounds. A dependable starter set:
blurry, low quality, distorted, extra fingers, extra limbs, watermark, text, logo, oversaturated, deformed face, bad anatomy, jpeg artifacts, cropped, out of frameWhat they cannot do is redesign the image. "Negative: ugly" does not make an image beautiful, and negating a concept your positive prompt implies ("negative: people" on "a busy market") sets up a fight the positive prompt usually wins.
Support varies by tool. Midjourney takes --no text, watermark at the end of the prompt. Leonardo and most Stable-Diffusion-based tools give you a dedicated field. ChatGPT and Gemini respond better to plain-language exclusions inside the sentence ("with no text anywhere in the image"), because they process instructions rather than raw keyword lists.
Keep negatives short. A fifteen-word negative prompt is a filter; a hundred-word one starts deleting things you wanted, and you will not be able to tell which word did it.
Ten Mistakes That Ruin Good Prompts
1. Style soup. "Photorealistic watercolor 3D render anime style" forces an average of four incompatible looks. Pick one medium per image.
2. Burying the subject. If your subject appears after word twenty, the model has already decided what the image is about. Lead with it.
3. Writing negatives into the positive prompt. "No cars, nothing in the background" mentions cars and background. Move exclusions to the negative prompt or phrase the scene positively: "an empty street".
4. Contradictions. Minimalist but intricate, dark but bright and airy, candid but perfectly posed. The model will not warn you; it will just average.
5. The keyword wall. Fifty comma-separated tags dilute each other. Twelve deliberate choices beat fifty reflexive ones.
6. Vague adjectives doing heavy lifting. Beautiful, epic, stunning, cool: none of them draw anything. Replace each with the concrete detail you actually mean.
7. One-shot expectations. Professionals generate, look, adjust one block, and generate again. Treating the first output as the final verdict on your prompt wastes the fastest feedback loop in creative software.
8. Copying syntax across models. --ar 16:9 --no text pasted into ChatGPT is just noise in the sentence. Parameters are Midjourney grammar; other tools want settings or plain words.
9. Demanding rendered text from the wrong model. Long sentences of in-image text still come out garbled in most tools. Keep rendered text short, quote it exactly, and use a text-strong model like Ideogram, or add type yourself afterwards.
10. Ignoring the aspect ratio. Generating square and cropping to a banner later throws away composition the model would have happily built for the wide frame. Set the ratio first.
Copy-Ready Prompt Templates
Fill the brackets, delete what you do not need, and adjust the ratio to the destination.
PORTRAIT
portrait of [person], [expression or action], [environment],
85mm lens, [lighting], shallow depth of field,
[style], [mood], [palette], 4:5
negative: blurry, distorted face, extra fingers, watermarkPRODUCT SHOT
[product] on [surface], [prop or context detail],
[studio softbox / natural window] lighting, [angle],
commercial photography, [palette], sharp macro detail, 1:1
negative: watermark, text, clutter, harsh reflectionsLANDSCAPE
[place] at [time of day], [weather or atmosphere],
[one focal element for scale], [wide angle / drone shot],
[cinematic / documentary] style, [palette], 16:9
negative: people, buildings, oversaturatedFLAT ILLUSTRATION
flat vector illustration of [concept or scene],
simple geometric shapes, [n]-color palette of [colors],
generous negative space, clean lines, [mood], [ratio]
negative: gradients, photorealism, textureLOGO MARK
minimal [flat / badge / monogram] logo of [symbol] for [brand type],
[line weight], [2-3 colors], centered on [background], no text
negative: text, letters, photorealism, gradients, clutter3D RENDER
3D render of [subject], [material: clay / glass / chrome / plastic],
[isometric / studio] view, [lighting], [palette],
high detail, [ratio]
negative: blurry, noisy, distorted geometryDownload the Cheat Sheet
Everything above compressed onto one printable page: the formula, the template skeleton, six keyword banks, the negative prompt starter and four fill-in templates.
Download the free AI Prompt Cheat Sheet (PDF) — no signup, no email, just the file.
After You Generate: Make the Files Usable
A prompt that finally works usually hands you a heavy PNG at a size no platform wants. The last step of the workflow is making the file fit where it is going.
- AI tools often export multi-megabyte PNGs even for photographic images. Run them through the image compressor, or convert photographs to JPG or WebP with the converter, and they shrink dramatically with no visible change.
- Need the image bigger than the model generates? The upscaler doubles dimensions cleanly.
- Generated a product or character you want on a clean backdrop? The background remover cuts it out in the browser.
All of it runs on your device, so your generations never leave your computer. For the full post-generation workflow, see our guide to optimizing AI-generated images.
The Takeaway
Prompting is not a secret language. It is the discipline of describing an image the way the person who would have made it describes it: what the subject is, what it is doing, where, in what light, through what lens, in what style and mood. The twelve blocks exist so you never stare at an empty prompt box again; the pyramid exists so you spend your words where they matter.
Start with the subject. Add one block at a time. Generate early and often. And keep the cheat sheet within reach until the formula becomes reflex.
