A prompt example is only useful if you can tell which words must stay and which are placeholders. Every template below marks that split: the slot names stay, the bracketed values change. Copy one, replace the brackets, run it, then change exactly one bracket at a time.
The templates follow the writing order from How to write GPT Image 2.5 prompts: goal, subject, light and style, composition, constraints. Nothing here is official prompt text. OpenAI publishes templates as a product feature with preset formats, not as a library of prompt strings, so the wording below is ours and the facts it rests on are quoted from the help centre.
How to read these templates
Four conventions, and they are the whole trick:
- Angle brackets mark the values you replace.
<product>becomes the thing you are actually shooting. - Plain sentences stay as written. “One subject only” is a constraint, not a suggestion.
- Quoted strings are text that will appear in the image, character for character.
- The order is fixed. Goal first, constraints last, for the reason the basics article explains.
A minimal template, expanded once, so the shape is clear:
Goal: <format> <ratio> for <destination>.
Subject: <one subject, described physically>, one only.
Light and style: <light direction and quality>, <rendering style>.
Composition: <camera height and angle>, <placement in frame>.
Constraints: no text, no logo, no watermark, no <unwanted extra>.Product and e-commerce prompts
Marketplace images have two hard requirements: the product must be recognisable, and the background must not fight it. Both come from stating the light and the surface, not from adding praise words.
Goal: 1:1 marketplace listing image for <product category>.
Subject: <product> standing upright on <surface>, one unit, label facing camera.
Light and style: soft diffused light from the upper left plus a weak fill on the
right, product photography look, realistic material texture, no stylised effects.
Composition: eye level, 85mm equivalent, product filling about 55 percent of the frame,
centred, generous margin on all four sides.
Constraints: pure <background colour> background, no text, no props, no hands,
no reflection on the surface, no second product.The one thing people change too often is the margin. A marketplace crop is decided by the platform, so generous margins on all four sides survive cropping better than a tight frame.
Poster and campaign prompts
A poster is text plus empty space. If either is unstated the model fills the gap with its own wording and its own layout.
Goal: A3 poster, portrait, 3:4, for <event or campaign>.
Subject: <one visual subject>, placed in the middle third of the frame.
Light and style: <light setup>, flat graphic rendering, two-colour print look.
Composition: lower third kept calm and empty for a date or venue line.
Text, set exactly: headline "<HEADLINE>", large, centred, upper third;
subhead "<SUBLINE>", small, centred under the headline.
Constraints: no logo, no watermark, no extra lettering, no second subject.Quote the lettering every time. An unquoted headline request is an invitation for the model to write its own line, and the result is usually a poster that would have worked with someone else’s copy.
Character and illustration prompts
Consistency is the hard part here, and it is a wording problem before it is a model problem. Name the features that must not move, and re-state them in every follow-up turn.
Goal: <format and ratio> for <destination>.
Subject: <character>, <age band>, <build>, <hair>, <clothing>, seen from
<angle>; the same character as any reference image supplied.
Light and style: <light direction>, <illustration style>, consistent line weight.
Composition: <camera height>, character occupying <share> of the frame.
Constraints: keep face, hair and clothing identical to the reference; no text,
no logo, no watermark, no additional characters.That “keep … identical” line is worth re-sending with every edit. Keep a fixed description block and paste it unchanged into each turn; the character consistency guide covers the multi-turn version of this.
Two worked examples to copy
Both examples below are complete and runnable. The first is a product shot with a fixed evaluation step; the second is a poster where only the headline is allowed to change.
Example one: a product shot
The point of this example is the last line. Decide what you will inspect before you generate, so the decision is not made by how the picture feels.
Goal: 4:5 product image for a marketplace listing, portrait.
Subject: a matte ceramic mug in <colour>, standing on a plain surface, one unit,
handle to the right, no other object in frame.
Light and style: soft light from the upper left with a weak fill on the right,
product photography look, visible clay texture, no gloss.
Composition: eye level, product centred, filling roughly half the frame,
even margin on all four sides.
Constraints: seamless <background colour> background, no text, no props, no hands,
no shadow cast onto the background, no second mug.
After: check the handle outline, the rim ellipse, the label area and the background.Acceptance is four checks: handle outline unbroken, rim ellipse symmetrical, label area clean, background unbroken top to bottom. If the background has a gradient seam, change only the background sentence and resend.
Example two: a poster edit
Here the layout is already right, so the only variable is the headline.
Goal: A3 poster, portrait, 3:4, for a Saturday market.
Subject: a shallow wooden crate of late-summer vegetables on plain linen, eye level.
Light and style: warm morning light from the upper left, screen-printed look,
two flat inks, visible paper grain.
Composition: lower third left calm and unbroken for a date line.
Text, set exactly: headline "HARVEST MARKET", large, centred, upper third;
subhead "Saturdays, 8am to 1pm", small, centred below the headline.
Constraints: no logo, no watermark, no second crate, no extra lettering.
Change one thing: replace the headline with "<NEW HEADLINE>" and keep everything else.Because the whole poster is restated, swapping the headline does not disturb the layout. That is the payoff of the fixed order: the variable sits in one line and everything around it stays pinned.
When a copied prompt misses
Work down this list before rewriting anything:
- The frame is wrong. Add the ratio and the camera height as their own sentence; do not add adjectives.
- Extra objects appear. Move the exclusion to its own line and name the specific object, not the category.
- The lettering is wrong. Quote the exact string, name the size and the position, and resend only that line.
- Something outside the edit moved. The documented limit applies: “Highlights are not always precise, and edits may extend beyond the area you selected.” Resend with a tighter selection rather than a longer instruction.
- The style is generic. Replace “high quality” with a rendering word — screen print, 85mm product optics, flat ink — because the model acts on nouns.
One more entry point is worth remembering when wording runs out: a drawing can carry the layout instead of words, which is what Sketch is for. And if the image contains text in Chinese, check the lettering against the Chinese text rendering guide rather than assuming a prompt fix will settle it.
GPT Image 2.5 prompt examples FAQ
- Can I paste these prompts as they are? Yes — every fence above is complete. Replace each bracketed value; leaving one in place usually produces a literal reading of the placeholder.
- Do the examples work for other models? The order and the slot split carry over, but the wording is tuned for GPT Image 2.5 and there is no cross-model guarantee.
- How many examples should I keep on hand? Three is enough: one product, one poster, one character. Reuse beats collecting, because you learn which line changed the result.
- Why quote the lettering instead of describing it? A quoted string is a text instruction; a description is a subject instruction, and only the first one is reliably reproduced.
- Where do the official templates fit? They are a product feature with preset formats you customise in the app, not prompt text you can copy. See Templates.
Where to go next
- Prompting basics explains the five parts and their order.
- Character consistency covers keeping one subject across turns.
- Chinese text rendering covers lettering outside the Latin script.
- Sticker pack is a worked case for a set of images instead of one.