How to Write Better AI Image Prompts: The 5-Part Formula

A better AI image prompt is a short, specific description of the finished image: the subject first, then the style, the light, the camera angle, and the mood — written as a picture, not as instructions to the model.

Most weak prompts fail for the same two reasons. They are too vague ("a nice photo of a woman"), or they read like a chat message ("please make me a beautiful image of..."). This guide gives you a repeatable five-part formula, three copy-paste templates, and a 15-minute iteration loop that works across Midjourney, GPT Image, Nano Banana, or any image model.

The 5-Part Formula

Every strong prompt is built from five layers, in this order:

1. Subject: who or what is in the frame, and what it is doing
2. Style and medium: photograph, oil painting, 3D render, editorial, watercolor
3. Light: direction, quality, and color of the light
4. Camera and composition: lens, angle, framing, depth of field
5. Mood and details: atmosphere, color palette, small props that sell the scene

Order matters. Image models give the most weight to the words that come first, so the subject always leads. A prompt that starts with the style ("cinematic, moody, 8k, ultra-detailed photo of a...") wastes its strongest position on adjectives.

Here is the same idea written two ways. Weak: "a beautiful portrait of a woman in a cafe, cinematic." Strong: "A young woman with curly dark hair, laughing over a ceramic coffee cup at a marble cafe table, warm morning light through a tall window, soft shadows, documentary photography, 85mm lens, shallow depth of field, cream and amber tones." The second prompt gives the model something to draw.

1. Start With the Subject, Not the Tool

Name the subject with concrete nouns. "A glass perfume bottle" is a start; "a square glass perfume bottle with a brushed-gold cap and a cream linen label" is a subject. Add one action or state when it helps: laughing, pouring, mid-stride, resting on a stone ledge.

Be specific about materials, because materials are what make an image feel real: frosted glass, brushed metal, wet marble, rough linen, polished concrete. If your image keeps coming back flat, the missing ingredient is usually a material, not a style word.

2. Name the Style and the Medium

Style is the single choice that changes everything, so name it plainly. "Editorial product photography," "35mm film photo," "watercolor illustration," "isometric 3D render," "cinematic still" — each one points the model at a different visual language. Mixing two styles ("oil painting photograph") confuses it; pick one and commit.

A useful habit is to anchor the style to a reference the model already knows: a magazine look, a film look, a decade. "Shot like a 1970s fashion editorial" gives the model a coherent palette, framing, and mood in a few words. When you want a consistent series, keep the style phrase identical across every prompt and change only the subject — that is how you get a set of images that actually look like they belong together.

3. Light Does the Heavy Lifting

Lighting is the layer most beginners skip, and it is the highest-leverage part of any prompt. It controls texture, depth, emotion, and whether the image feels flat or cinematic. One lighting phrase can rescue a prompt that is otherwise perfect.

Think in three dimensions: direction, quality, and color. Direction: "soft window light from the left," "backlit silhouette," "overhead studio light." Quality: "soft diffused," "hard directional," "golden hour glow." Color: "warm amber," "cool blue morning light," "neutral daylight." Combine one from each group and you have a complete lighting description in under ten words.

4. Camera and Composition

The camera line tells the model how to frame the scene. A few phrases cover almost every case:

• Lens: "85mm portrait lens," "35mm wide angle," "macro lens"
• Angle: "eye level," "shot from above," "low angle looking up," "three-quarter view"
• Framing: "close-up," "waist-up," "wide shot with negative space," "centered composition"
• Depth: "shallow depth of field," "everything in sharp focus"

Composition also means deciding what is not in the frame. "Clean background with negative space on the left" is how you get an image you can put text on. "Centered product, minimal styling, no props" is how you get a catalog shot instead of a lifestyle scene.

5. Say What to Leave Out

Negative prompts — telling the model what to exclude — fix most warped and cluttered results. Depending on your tool, you can use a dedicated negative field, a "--no" flag, or a plain sentence at the end: "no text, no watermark, no extra fingers, no blurry background." Keep the list short — three or four items that match the actual problems you are seeing. A wall of twenty negatives dilutes the prompt and often makes things worse.

The most useful negatives are the ones that repeat across your generations. If every portrait comes back with garbled text on a shirt, "no text on clothing" goes into your standard negative list. Build that list once, reuse it everywhere.

Three Copy-Paste Templates

Use these as starting points. Replace the bracketed parts, keep the structure, and keep each prompt under about 60 words — after roughly 75 tokens most models stop paying attention to the rest.

Portrait template

[Subject with one distinctive feature], [action or expression], [setting], [lighting direction and quality], [style or medium], [lens and framing], [color palette].

Example: "A young woman with silver-blonde hair, laughing over a ceramic coffee cup at a marble cafe table, warm morning window light, soft shadows, documentary photography, 85mm lens, shallow depth of field, cream and amber tones."

Product template

[Product with material and label detail], [surface or setting], [props], [lighting], [style], [camera angle], [background].

Example: "A frosted glass serum bottle with a brushed-gold pump and cream linen label, standing on a travertine stone plinth, a sprig of eucalyptus beside it, soft diffused window light from the left, editorial product photography, three-quarter angle, clean beige background."

Lifestyle template

[Person or group], [activity], [place with two concrete details], [time of day and light], [style], [composition].

Example: "A surfer girl in a yellow swimsuit carrying a surfboard past a beach cafe, late afternoon golden light, candid travel photography, wide shot with the ocean in the background."

Taste Is the Part No Formula Covers

A formula can make your prompts competent. Taste is what makes them good. The difference between a prompt that produces a usable image and one that produces a striking image is usually a handful of judgment calls: which detail to keep, which cliche to cut, when the scene is finished and when it is over-decorated. Beginners overstuff prompts because every added word feels like control; experienced prompt writers cut words until only the ones that change the image remain.

If you work with AI design tools regularly, it is worth building that judgment deliberately. The Taste Skill for Claude Code (https://skillstackhq.com/en/skills/leonxlnx-taste-skill) is a good reference for what "good taste" looks like as a set of rules — it teaches an AI to avoid generic, slop-looking design, and the same principles apply to the images you prompt for: fewer cliches, stronger composition, one clear idea per image.

The 15-Minute Iteration Loop

Good images come from iteration, not luck. Here is a loop that works:

1. Write the prompt with the five-part formula and generate four variations in one batch.
2. Pick the best one and name what is wrong with it — one sentence, specific: "the light is too flat" or "the label is unreadable."
3. Change one layer only. Fix the light, or fix the framing, never both at once. If you change everything, you cannot tell which change helped.
4. Lock what works. When a prompt produces a keeper, save the wording — including the negative list — in a file or a note. Reuse the proven skeleton and swap only the subject next time.
5. Stop at good enough. After three rounds without a clear improvement, keep the best version and move on.

This loop is also why ready-made prompts are worth using: they let you skip the first two rounds. If you want a deeper look at prompt craft in general, the guide on how to write AI prompts that rank (https://skillstackhq.com/en/blog/how-to-write-ai-prompts-that-rank) covers the same thinking applied to text prompts.

Common Mistakes

• Writing instructions instead of descriptions. "Please make a picture where a woman..." wastes words. Describe the picture itself.
• Overstuffing. More than about 60 words, and the model starts ignoring the tail of your prompt. Cut adjectives that do not change the image.
• Copying the same prompt across models. Midjourney likes short, comma-separated keyword phrases; GPT Image and Nano Banana handle natural-language sentences better. Adapt the wording to the model.
• Rerolling blindly. Generating twenty versions of a broken prompt is slower than fixing one layer and generating four.

Conclusion

Writing better AI image prompts is a craft with a short learning curve: describe the finished image, subject first, in five layers — subject, style, light, camera, mood — then say what to leave out, and iterate one layer at a time. Start from the templates above, or from a proven prompt you adapt, and save the ones that work. After a few rounds you will spend less time re-rolling and more time shipping images that look art-directed, not generated.