Every AI image tool — Midjourney, DALL-E, Gemini, Stable Diffusion — responds to the same fundamental prompting principles. Learn how to write good prompts for AI image generation once, and the skill transfers everywhere. Here’s the universal formula.
The Universal Prompt Formula
Subject + Details + Style + Lighting + Composition + Mood
Most bad prompts fail because they’re missing 3-4 of these. “A dog” gives the AI nothing to work with. Watch what happens as we add layers:
- “A dog” → generic, boring.
- “A golden retriever puppy” → better, still generic.
- “A golden retriever puppy playing in autumn leaves” → scene emerging.
- “A golden retriever puppy playing in autumn leaves, golden-hour sunlight, shallow depth of field, professional pet photography” → that’s an image.
Each layer removes a guess. The AI’s guesses average toward cliché; your specifics pull toward interesting.
Each Layer, Explained
Subject: Nouns Over Adjectives
Concrete nouns render best: “weathered fisherman” beats “a man.” If the subject matters, spend words on it — age, clothing, expression, activity. Vague subjects get vague images.
Style: The Biggest Lever
One style phrase transforms everything: photorealistic, cinematic, watercolor, pixel art, art nouveau, editorial photography, isometric 3D, charcoal sketch, ukiyo-e, synthwave. When in doubt about why your image looks wrong, the style word is usually the culprit — or the missing piece.
Lighting: The Mood Machine
“Golden hour,” “neon glow,” “soft morning light,” “dramatic side lighting,” “overcast.” Lighting controls emotion more than any other single element. Two identical scenes at noon vs. dusk are different images entirely.
Composition: Think Like a Photographer
“Close-up portrait,” “wide landscape,” “aerial view,” “centered composition,” “shallow depth of field.” These tell the AI how to frame, not just what to show.
Tool-Specific Notes
- Midjourney: loves style words and parameters; see the full Midjourney guide.
- DALL-E / ChatGPT: best at following long, detailed natural-language descriptions. Write conversationally.
- Gemini: great at conversational editing — “now make it nighttime” actually works well.
- Stable Diffusion: rewards structured, comma-separated tag lists and negative prompts.
What to Avoid (Negative Prompts)
Specifying what you don’t want is as important as what you do: “no text, no watermark, no blurry, no extra fingers, no distorted faces.” Most tools support negative prompts directly — the full technique is covered in negative prompts explained. Hands and text remain the classic AI failure modes; excluding them preemptively saves generations.
The Iteration Loop (How Pros Actually Work)
- Generate 4 variations of your base prompt.
- Pick the closest, identify what’s wrong specifically (“lighting’s great, but the pose is stiff”).
- Adjust one thing and regenerate.
- Repeat 2-3 times. Stop when it’s good enough — perfectionism burns credits.
Beginners generate once and accept it. Pros generate, select, refine. The loop matters more than the initial prompt.
How to Write Good Prompts for AI Image Generation: The Pro Iteration Loop
- Generate 4 variations of your base prompt.
- Pick the closest, identify what’s wrong specifically (“lighting’s great, but the pose is stiff”).
- Adjust one thing and regenerate.
- Repeat 2-3 times. Stop when it’s good enough — perfectionism burns credits.
Beginners generate once and accept it. Pros generate, select, refine. The loop matters more than the initial prompt.
Prompting for People (The Hard Part)
Humans are the hardest subject. Specific techniques:
- Describe, don’t name: “a woman in her 60s with silver hair in a bun, laugh lines, wearing a denim jacket” beats “an old lady.” (Don’t prompt real people’s names.)
- Specify the feeling: “candid laughter,” “quiet contemplation,” “confident stance” — emotion words shape the result enormously.
- Hands: the classic failure. “Hands in pockets,” “holding a coffee cup naturally,” or simply “–no deformed hands” preempts the worst outcomes.
- Diversity deliberately: models default to training-data averages. Specify age, ethnicity, style when it matters for your purpose.
- One person, clear action: group scenes multiply failure modes. Single subjects with clear poses are most reliable.
The Ethics of AI Portraits
- Don’t create fake people for deception: AI-generated “team members” on company sites, fake testimonials with AI faces — dishonest and increasingly illegal.
- Disclose AI generation where audiences expect real photography (news, dating profiles, portfolios).
- Respect likeness: generating images of real people without consent is a legal and ethical minefield.
- Stock photo replacement: the legitimate use — custom illustrations and conceptual imagery where no real photo exists.
Building a Personal Style Library
As you find style phrases that work, save them: your go-to lighting setups, favorite art styles, reliable negative prompts. Over months, this becomes a personal visual language — prompts that produce your aesthetic consistently. Share them or keep them secret; either way, they’re an asset. The photographers of the AI era aren’t the ones with the best camera — they’re the ones with the best prompt library.
Building a Personal Style Vocabulary
The fastest way to improve your image prompts is building a style vocabulary — words that reliably produce aesthetics you like. Start collecting: when a generated image nails the look you wanted, save the style words from that prompt (“cinematic still,” “editorial photography,” “ukiyo-e inspired”). Organize by category: lighting (golden hour, rim light, volumetric), composition (rule of thirds, symmetrical, dutch angle), mood (melancholic, triumphant, serene), era (art deco, 70s film grain, cyberpunk). After a month you’ll have a personal lexicon that makes every prompt stronger. This is also where reusable prompt templates pay off — save your best-performing style combinations as templates and remix them for new subjects. The pros aren’t better at describing; they’re better at reusing what works.
Anthropic’s prompt engineering documentation is another excellent deep-dive, with techniques specifically tuned for Claude’s strengths.
Bottom Line
How to write good prompts for AI image generation: concrete subject, decisive style, deliberate lighting, photographic composition — then iterate. The formula works in every tool; only the syntax (parameters, negative prompts) changes. Master the layers, embrace the iteration loop, and your images will jump a level within days.
