A GPT Image 2.5 prompt is five decisions: what the picture is for, what is in it, how it is lit and styled, how it is framed, and what must not appear. Written in that order, the model has less room to guess and the prompt stays editable when one part turns out wrong.
That structure is ours, not OpenAI’s. The help centre names three things the model follows as instructions — add text, add details within the image, make the background transparent — and one sentence about being specific: “Be specific about the details that matter. For example, if you’re creating a poster, describe what it’s for and any text you want it to include.” This page is built from those sentences plus the release notes, and it never supplies a number the vendor has not published.
The five parts of a prompt
The five parts, in the order they earn their place:
- Goal. The format and the destination — an A2 poster, a 1:1 marketplace listing, a 4:5 social frame. This decides ratio, empty space and how large the subject sits.
- Subject. One noun phrase, described physically: how many, what they are made of, what they are doing.
- Light and style. Where the light comes from, how hard it is, and the rendering you want.
- Composition. Camera height, angle, and where the subject sits in the frame.
- Constraints. The exclusions list: no logo, no watermark, no second object, no extra lettering.
The labels are optional; the sequence is not. The first two parts describe what exists, the last three how it should be drawn — a split models follow more reliably than a pile of adjectives.
Order of the five elements
Order is not decoration. Each position removes a class of ambiguity, and a prompt that leads with style leaves the subject and the frame unspecified — which is where most drift comes from.
| Position | Element | What it settles |
|---|---|---|
| 1 | Goal | Format, ratio, empty space |
| 2 | Subject | What is physically in the frame |
| 3 | Light | Direction, hardness, mood |
| 4 | Framing | Camera height, angle, placement |
| 5 | Exclusions | What must not be added |
A skeleton worth saving, with the variable parts left as placeholders:
Goal: <format and destination, plus ratio>.
Subject: <one subject, described physically>.
Light and style: <direction and quality of light, rendering>.
Composition: <camera height, angle, placement>.
Constraints: no text, no logo, no watermark, no <unwanted extra>.The popular eight-step order
Chinese-language guides often circulate a longer sequence — purpose, subject, composition, style, light, material, text, negative prompt. It is a useful mnemonic and it is not an official list; no accessible OpenAI page publishes a step order at all. Keep the eight buckets if they help you remember, but do not quote them as vendor guidance.
The narrowest official order is the three instructions the model accepts, plus the aspect ratio, which you set with the picker or write into the prompt. The GPT Image 2.5 overview covers what changed in the model itself.
Text, ratios and restrictions
Lettering is the part that fails most visibly, and the fix is quoting. Ask for "HARVEST MARKET" in quotes, and say where it goes and how large it is; ask for a headline and the model invents the wording. Size, casing and position are all worth stating, because a poster that is right except for the typography is not usable.
Ratios work in two places: the aspect ratio picker, or the prompt itself — the second version travels with the prompt when you save or share it. Restrictions belong last for a reason. They only matter once the rest is decided, and a long exclusion list at the top tends to pull the model toward the things it is being told to avoid.
Two prompt examples to copy
Both run as written. The first makes a picture from nothing; the second changes one region of a picture you already have.
Example one: a market poster
A poster is a fair test, because ratio, empty space and lettering all have to be stated before the model can get them right.
Goal: A2 poster, portrait, 3:4, for a Saturday farmers market.
Subject: a shallow wooden crate of late-summer vegetables on plain linen, eye level,
one crate only, nothing else on the surface.
Light and style: soft warm morning light from the upper left, gentle falloff,
screen-printed poster look, two flat inks, visible paper grain.
Composition: crate in the middle third, lower third left calm and unbroken.
Text, set exactly: headline "HARVEST MARKET", large, centred; subhead
"Saturdays, 8am to 1pm", small, centred below the headline.
Constraints: no logo, no watermark, no second crate, no extra lettering.If the lettering comes back wrong, change only the quoted line and resend. Rewriting the whole prompt teaches you nothing about which part was ambiguous.
Example two: a product still
Here the object is right and the surroundings are wrong, so the edit is worth doing as a change rather than a regeneration.
Reference: the uploaded photo of one product on a wooden desk.
Change: replace the wooden desk with a seamless pale grey studio sweep
and even soft light from the front.
Keep: the product, its proportions, its label, its position in the frame
and the light already falling on it.
Constraints: no text, no props, no reflections, no second product.
After: check the product's outline for stray grey or clipped edges.That last line is the acceptance check. Look along the product’s outline, where a background swap leaves evidence.
Prompts are not the only input
Three other ways to direct the model are documented, and mixing them up with prompting is where most tutorial confusion starts.
- Templates give you a starting point such as a poster or a logo, which you then fill with details. OpenAI’s note is explicit: templates are not yet available in Work mode. See Templates.
- Sketch turns a drawing into the layout. On mobile, type @ in the message box, pick Sketch, draw the outline, confirm it, then add written instructions. The sketch carries placement, not detail. See Sketch.
- Comments put an instruction on the image itself, tighter than a whole-frame edit. See Comment editing.
The prompt you used can also be shared with the image, so someone else can rerun your idea against their own material.
Why a prompt drifts off target
Three causes cover most of it. The prompt describes an effect instead of a thing (“high quality”, “premium feel”); the frame was never stated, so the model picks a ratio and a distance for you; or the exclusion list is doing work the subject description should have done. Adding “very high quality” to a prompt that never named the subject will not fix any of them.
When you edit instead of regenerate, one documented limit applies: “Highlights are not always precise, and edits may extend beyond the area you selected.” After a selected edit, look along the rim of the region you drew; if something outside it moved, delete that turn and resend with a tighter selection.
The rule that keeps a session productive is one variable per turn. Change the light, the ratio or the lettering — not all three — so that when the result improves you know why.
GPT Image 2.5 prompt FAQ
- Does GPT Image 2.5 need a special prompt format? No. There is no official syntax and no schema; the five-part order above is a working convention that keeps prompts editable.
- Should the prompt be in English? English is the safest default, but the model also follows instructions to add text in other scripts — state the exact string and check the result.
- How long should a prompt be? Long enough to fix the five parts and no longer. Specificity is the variable, not length.
- Can I set the size in the prompt? Yes, and that version survives sharing. The aspect ratio picker does the same job interactively.
- Are there official prompt templates? Templates exist as a product feature with preset formats, not as a published library of prompt text. See Templates.
Where to go next
- Getting started covers the first session end to end.
- Templates covers preset formats and the Work mode gap.
- Sketch covers drawing the layout before you prompt.
- Comment editing covers instructions placed on the image.