How to Write AI Image Prompts
Published September 5, 2026
Most AI image prompt guides are written for artists exploring styles. A business needs the opposite: one image that fits a specific slot — a product shot at the right ratio, a blog header with room for a headline, a social card that matches the brand — without burning twenty generations hunting for it. The prompt is what turns that hope into a specification.
This guide gives you the formula first, then worked prompts for the four jobs that come up constantly, then how to pick the model, size and ratio. Every dimension quoted below is what the generator actually produces — we generated the images and measured them — and the last step (compressing before upload) matters more for page speed than most guides admit.
The short answer
A working AI image prompt names five things: the subject, the context around it, the style, the composition, and the lighting — in one to three sentences. “Glass water bottle on wet stone, studio product photo, centered with copy space, soft diffused light” beats a paragraph of adjectives. Iterate at 1K, then re-run the winning prompt at 2K, and compress before it goes on a page.
The five-part formula that does most of the work
Model providers' own prompting guides converge on the same structure: describe the subject, the context around it, the style of image, the composition, and the lighting. Each part removes one guess the model would otherwise make — and every guess it doesn't have to make is one less place the picture goes strange.
Length is the surprise: OpenAI's own image tutorial says one to three clear sentences are enough. A prompt is not a thesaurus exercise. Stacking twenty adjectives muddies every choice the model makes; five concrete decisions beat fifty vague ones.
| Part | What it answers | Example fragment |
|---|---|---|
| Subject | What is in the picture | a glass water bottle |
| Context | Where it is, doing what | standing on wet dark stone |
| Style | What kind of image | studio product photograph |
| Composition | Framing and layout | centered, empty space on the left |
| Lighting | Light quality | soft diffused light, subtle reflection |
Worked prompts for the jobs a business actually has
The four prompts below are ready to paste. Each one names all five parts, and each one is paired with the ratio and size that fit the slot — which is the detail most people leave until after generating, then discover the crop is wrong.
| Job | Prompt | Ratio & size |
|---|---|---|
| Product hero | Matte ceramic mug in sage green, three-quarter view on a light oak table, morning window light, shallow depth of field, e-commerce product photo | 1:1 · 2K |
| Blog featured image | Flat illustration of a tidy desk with laptop and notebook, muted pastel palette, generous negative space, minimal vector style | 16:9 · 2K |
| Social post | Single scoop of strawberry ice cream against a flat pastel background, high-key studio lighting, playful advertising photography | 3:4 · 2K |
| Banner / hero strip | Wide cinematic photo of an open road at golden hour, empty left half for text, warm tones, crisp horizon | 21:9 · 2K |
Picking the size and ratio (measured, not guessed)
Resolution tiers and aspect ratios are not cosmetic — they decide whether the image fits the slot without cropping. The dimensions below are the generator's real outputs, measured from generated files: a 1K 16:9 image is 1312×736 pixels, and a 2K square is 2048×2048.
The workflow that saves the most credits: draft at 1K on the fastest model while you are still changing the prompt, then re-run the winning prompt unchanged at 2K on the best model. The prompt is the specification — once it is right, the same text produces the same picture at higher resolution.
| Placement | Ratio | Working size | Pixels you get |
|---|---|---|---|
| Product / square social | 1:1 | 2K | 2048×2048 |
| Blog hero / YouTube thumb | 16:9 | 1K draft, 2K final | 1312×736 at 1K |
| Story / reel cover | 9:16 | 2K | portrait, same tier math |
| Banner / hero strip | 21:9 | 2K | ultra-wide, crop-free |
The words that move the picture most
Not all words pull equal weight. Style and lighting vocabulary changes the whole render for one or two words, which makes it the cheapest lever to iterate on: keep the subject fixed and swap the style or the light until the mood is right.
| You add… | The picture gets… |
|---|---|
| “golden hour” | warm directional light, long shadows |
| “soft diffused light” | even, nearly shadowless — the product-shot standard |
| “shallow depth of field” | blurred background, subject isolated |
| “flat vector illustration” | clean shapes, no photographic realism |
| “negative space on the left” | room where a headline can go |
| “isometric” | technical 3D-looking illustration, good for diagrams |
Five mistakes that waste your generations
Every one of these costs a round trip to the model. They are listed in the order people meet them.
- Describing what you don't want. Text-to-image models do not process negation reliably — “no hands, no text” often adds both. Describe what you do want, and regenerate when a stray element appears.
- Prompt stuffing. Twenty adjectives dilute every decision; one to three sentences with five concrete parts win. This is the model providers' own guidance, not a style preference.
- Expecting long text inside the image. A single short word sometimes renders correctly; sentences almost never survive. Generate the image with negative space and add real text in your editor, where fonts are crisp and editable.
- Chasing an exact repeat. The same prompt produces a different image every run — that is how these models sample. Save the prompt of every winner instead of trying to nudge a near-miss back.
- Generating at final size on attempt one. A 2K image is ~2 MB of PNG; iterating at 2K is slow and pointless. Draft at 1K, and only the winner goes big.
Make it, then make it web-ready
The generator turns these prompts into images; the optimizers get the file down to upload weight — all free, all in your browser.
Turn any prompt built with this formula into an image — three models, 1K–4K, eight ratios, your own free API key.
A 2K PNG is around 2 MB. Get it to web weight in your browser before it goes on a page.
Exact pixel dimensions when a platform crops — no upload, no resample surprises.
Photographic generations shrink dramatically as JPG; keep PNG only when you need transparency.
Frequently asked questions
- What makes a good AI image prompt?
- Five named parts in one to three sentences: subject, context, style, composition, lighting. “Glass water bottle on wet stone, studio product photo, centered with copy space, soft diffused light” is a complete prompt — concrete decisions, no adjective piles.
- How long should an AI image prompt be?
- One to three sentences. OpenAI's own image-generation tutorial says clear beats long: each extra clause dilutes the ones before it. If a prompt needs a paragraph, the image usually needed to be two images.
- How do I write a prompt for a product photo?
- Name the product and its finish, the surface it sits on, the light, and say “e-commerce product photo” or “studio product photograph” to fix the style. Example: “Matte ceramic mug in sage green, three-quarter view on a light oak table, morning window light, shallow depth of field, e-commerce product photo.” Generate 1:1 at 2K.
- Can AI images contain readable text?
- Short single words sometimes render correctly; full sentences almost never do. The reliable workflow is to generate the image with deliberate negative space (“negative space on the left”) and add real text in an editor — crisper, editable, and translatable.
- Why did the same prompt give me a different image?
- Generation is sampled, not replayed — the same prompt yields a different result every run, by design. Keep a list of winning prompts instead of trying to reproduce one image exactly; re-running the winner is the repeatable part.
- What resolution should I generate at?
- Draft at 1K while changing the prompt, publish at 2K. In this generator 1K 16:9 is 1312×736 and 2K square is 2048×2048 — measured from real outputs, not estimates. 3K and 4K are for print and very large displays; a web page never shows more than 2K.
- Can I use AI-generated images commercially?
- That is governed by the AI provider's terms, not by this site — check the current terms of the service that generated the image before using it in paid work, and keep trademarks, logos and real people's faces out of prompts for commercial assets.
Sources
- OpenAI — GPT image models prompting guide (OpenAI Cookbook)
- The model provider's own prompting documentation for image generation: prompt structure and lighting/composition control for production-quality visuals.
- OpenAI Academy — Creating images with ChatGPT
- States the length guidance quoted here: good image prompts do not need to be long, typically one to three clear sentences.
- Google Cloud — A developer's guide to Imagen 3 on Vertex AI
- Google's own guidance on prompting Imagen for photorealistic output, converging on the same subject-context-style-composition structure.
- ToolNest — measured generator outputs (2026-09-05)
- The dimensions in the sizing table were verified by generating a 1K 16:9 image (1312×736) and a 2K square (2048×2048) and reading the file headers. Unit tests pin the guide to these figures.
General information only, not financial, tax or legal advice. Rates and rules vary by jurisdiction and change over time — verify anything consequential with a qualified professional. See our disclaimer.