How to Write AI Image Prompts That Actually Work

Quick answer

Learning how to write AI image prompts comes down to one habit: describe the finished picture in a fixed order — subject first, then style, lighting, composition, and what to leave out. Most bad results fail on one of those five slots, not on the whole idea. A prompt that names a specific subject, a visible style, a light source, a camera position and a short list of exclusions will beat a long string of adjectives almost every time. The examples below show the formula working, and the fixes section covers the five failures that waste most people’s time.

What changed with Images 2.6

OpenAI launched Images 2.6 on 4 November 2026, and it is the reason this guide exists now rather than last month. The update is available in ChatGPT, Codex and the API across all subscription tiers, generates images up to 50% faster than its predecessor (OpenAI’s own claim), and adds animated GIF creation, markup-guided edits and SVG export. For API users there are two new models: GPT-Image-2.6 Flare for speed and Sunburst for detail.

That matters for prompting because the models now follow instructions more closely. A vague prompt used to produce a vague picture; a vague prompt now produces a confident picture of the wrong thing. OpenAI’s own prompting guide for its GPT Image models makes the same point in plainer terms: write the prompt the way you would write a short spec for a designer — subject, scene, details, constraints, in that order — and keep settings like size and quality out of the words themselves. The rest of this guide follows that structure, and it works across the other major tools too, with the per-tool differences noted near the end.

The five-slot formula

Every prompt in this guide uses the same five slots, in the same order. Image models give more weight to the words that come first, so the subject leads and the exclusions close.

Slot	What it answers	Example fill
Subject	Who or what, doing what, where	a red-haired woman reading on a park bench in autumn
Style	The medium and art direction	35mm film photo, flat vector illustration, oil painting
Lighting	Light source, direction, mood	soft window light, golden-hour backlight, neon rim light
Composition	Framing, angle, placement	close-up portrait, eye-level, subject left, space right
Constraints	Quality cues and what to exclude	sharp focus, film grain; no text, no watermark, no extra fingers

Here is one prompt built slot by slot, so you can see the order doing its work:

“A golden butter croissant on a white marble counter, with a few flakes of pastry beside it. Close-up from a slight top-down angle, the croissant in the lower right of the frame. Bright bakery photo style, soft morning light from a window behind, pale cream and gold colours. No text, no hands, no other food.”

Nothing in that prompt is fancy. It works because each slot answers its own question and nothing is left for the model to guess.

Three prompts, fixed

The fastest way to learn the formula is to watch a bad prompt turn into a good one. Each pair below keeps the same idea and changes only the description.

A product photo. The bad version: “a nice bottle of perfume.” The model has to guess the bottle, the surface, the light and the shot — and it will guess all four badly. The fixed version:

“Product photo of a matte black perfume bottle on a pale stone surface, soft morning window light from the left, clean neutral background, realistic contact shadow, premium minimal composition, no text, no hands.”

What changed: a concrete subject (matte black bottle, pale stone), a named light source and direction, a composition, and two exclusions. The words “premium” and “minimal” are the only mood words left, and they sit on top of visible detail instead of replacing it.

A portrait. The bad version: “a woman in a city.” The fixed version:

“A woman in her 40s with silver hair, wearing a tailored navy suit, standing in a rain-soaked Tokyo street at night, neon signs reflected in wet asphalt, cinematic 35mm photo, shallow depth of field.”

What changed: age, hair, clothing, weather, time of day and a reflection that tells the model what the street looks like. “Cinematic” earns its place because the rest of the prompt already shows what cinematic means here.

An illustration. The bad version: “a cosy cabin.” The fixed version:

“A small wooden cabin in a snowy pine forest at dusk, warm light in the windows, smoke curling from the chimney, flat vector illustration, deep blue and amber palette, no text.”

What changed: the mood word “cosy” became three visible things — warm window light, smoke, dusk. That swap is the single most reliable fix in this whole guide, and it gets its own section below.

Five ways prompts go wrong

“It looks generic”

The cause is almost always a vague subject. “A dog” gives the model a thousand dogs to choose from, so it picks the most average one. Name the breed, the age, the pose and the place: “a golden retriever puppy wagging its tail on a wet beach at dawn.” If the subject still comes out wrong, add one more concrete detail rather than three more adjectives. Adjectives describe; nouns decide.

“The mood is wrong”

Mood words like cosy, luxury, epic and premium mean something different to every reader, so the model has to guess what you meant. Turn each mood word into two or three things a camera could actually capture. Cosy becomes a woollen blanket, steam rising from a mug, lamplight on wood. Luxury becomes brushed brass, soft shadows, a single object on empty marble. Keep the mood word if you like it — but never let it stand alone.

“The light looks flat”

Lighting is the strongest visual variable you can control, and the one most prompts skip entirely. Name the source and the direction: a single window on the left, an overcast sky, a warm lamp behind the subject. The same scene under golden-hour backlight and under flat midday sun looks like two different pictures. If your image feels dull, the missing slot is usually this one, not the subject.

“The text is gibberish”

Small image models are poor at lettering, and even the newer ones need help. GPT Image models render text far better than their predecessors, and Ideogram is widely recommended for images that must contain readable words — but the habit is the same everywhere: put the exact words in quotation marks, keep them short, say where each line sits, and ask for no extra text, since stray words are the most common problem in text-heavy images. For unusual words or brand names, spell them out letter by letter. And when the lettering has to be perfect, generate the image without text and add it afterwards in an editor. That is not a workaround; it is the normal workflow.

“Every retry looks different”

Each generation starts from random noise — a seed — so the same prompt produces a different image every time. That is normal, not a bug. Generate several variations and pick the best one, then change one thing at a time when you edit: “change only the background to a plain grey studio wall, keep the product’s shape, label text, colours and reflections.” OpenAI’s guide is explicit about this: start from a clean base prompt and refine with small, single-change follow-ups. If you change five things at once, you will never learn which one fixed the picture.

The same prompt behaves differently in each tool

The formula holds everywhere, but the dialect changes. ChatGPT’s image generation (GPT Image) takes plain-language paragraphs and follows quoted text well, so write the way you talk and put exact wording in quotes. Midjourney responds best to short, high-signal phrase sequences — dense keywords rather than sentences — and takes negatives through a –no parameter, which reads each word on its own. Stable Diffusion rewards structured, weighted keywords and has a dedicated negative-prompt field. Ideogram is the strongest choice when the image has to contain readable text. None of these is a verdict on which tool is best overall; it is a note on which phrasing each one rewards, so the same idea does not have to be rewritten from scratch when you switch tools.

Negative prompts: what to leave out

The last slot is the one beginners skip and experienced users never do. A negative prompt tells the generator what to avoid, and it works by steering the image away from named traits. The usual list is short: “blurry, extra limbs, text, watermark.” In Stable Diffusion it goes in the negative field; in Midjourney it goes after –no; in GPT Image it goes in the constraints as plain language (“no text, no watermark, no extra fingers”). One caution: a negative that contradicts the subject — “no water” in a beach scene — will confuse the model more than it helps. Negatives are for flaws and clutter, not for arguing with your own subject.

How to write AI image prompts: a checklist you can reuse

Copy the five slots, fill each one, and read the prompt back before you generate:

1. Subject: the main thing, doing something specific, somewhere specific.
2. Style: the medium — photo, painting, illustration, render — named, not implied.
3. Lighting: source and direction, in one short phrase.
4. Composition: framing, angle, and where the subject sits.
5. Constraints: quality cues plus what to exclude, including “no text” unless text is the point.

Ten to sixty words is the range most guides suggest, with the subject first. After roughly 75 tokens most models stop paying attention to the rest, so a prompt that runs to a paragraph is usually a prompt that has not decided what matters yet.

If you use AI tools for research as well as images, our guide to using NotebookLM covers a similar step-by-step workflow for Google’s notebook tool.

Your next step is a small one: take a prompt that disappointed you, run it through the five slots, and change only the slots that were empty. Most failed prompts are missing two slots at most, and the fix takes less time than generating the same vague prompt a sixth time.

Sources

• OpenAI, GPT Image prompting guide (cookbook, current): https://developers.openai.com/cookbook/examples/multimodal/image-gen-1.5-prompting_guide
• Moneycontrol, “OpenAI launches Images 2.6 with faster image generation, GIF creation support,” 5 Nov 2026: https://www.moneycontrol.com/technology/openai-launches-images-2-6-with-faster-image-generation-gif-creation-support-article-14071178.html
• NextG Design, “AI Image Prompt Tips: How to Write Prompts That Work,” 19 Oct 2026: https://nextgdesign.com/guides/ai-image-prompt-tips.html
• Vogoo AI, “How to Write AI Picture Generator Prompts (A Copy-Paste Formula),” 28 Jul 2026: https://vogoo.ai/blog/how-to-write-ai-picture-generator-prompts