The four looks everyone is copying, the real prompts behind them, and the craft words that decide whether your image looks expensive or looks like AI.
Saving a prompt gets you one image. Learning why it works gets you every image after that. So this is the craft layer underneath the library, taken from the prompts with real numbers behind them and from OpenAI's own prompting documentation.
If you want the five business jobs to point it at instead, from sketch-to-mockup through merch and flyers, I covered those in ChatGPT Images 2.5 for your business. This guide is how to make any of them look good.
Two facts that change how you work. There is no seed, so the same prompt twice gives you two different images, and consistency has to come from reference photos instead. And 2.5's real leap is editing, not generating. It holds an image steady while you change one thing, which means you stop re-rolling and start directing.
Every prompt worth copying has the same bones. OpenAI's guidance is to write it as labeled sections rather than one long paragraph, and the viral ones all do exactly that. Type, format, subject, identity lock, materials, pose, environment, lighting, camera, grade, text, constraints.
Four of those do most of the work:
The full fill-in-the-blank version is in the pack at the bottom.
Camera language is the fastest lever you have. One caveat from OpenAI first: these are cues for appearance, not physics. "85mm f/1.8" is a style word, not a setting. It biases the look, it does not simulate a lens.
Delete these five words
8K, ultra-realistic, masterpiece, hyper-detailed, award-winning. They push the image toward waxy skin and the glossy AI sheen. Replace them with "realistic skin texture, visible pores, subtle natural imperfections." Golden hour has gone the same way, so reach for flat overcast light or window light instead.
The camera-roll look is the one people cannot believe, and it works by inversion. You stop describing beauty and start describing a camera failing. Every lever is a flaw.
The line I would not cross
These are brilliant for content and concepts. They are designed to look like documentary evidence, so use your own face, or someone who gave you written permission, or an invented person. Never use a real person who has not agreed to it, and never dress something up as though it actually happened. OpenAI's own policy bans generating a real person's likeness without their consent.
When an image is 90% right, almost everyone throws it away and rolls again. That is the expensive mistake, because there is no seed and you will never get back to the version you liked.
Instead, use the formula. Tell it to change only X, then hand it a long explicit list of everything to preserve, and finish with make no other changes. The longer and more specific that preserve list, the less the rest of the image drifts. One change per turn, never five.
Two honest limits. Repeated edits can still move details you asked to keep, so check the whole image every time rather than just the bit you changed. And if something has to be pixel-perfect, like your actual logo, paste it over the top in Canva afterwards rather than trusting the prompt. On text generally, 2.5 handles short bold copy well and still fumbles small dense type, prices and dates, so typeset the final numbers yourself.
The skeleton, one full working prompt for each of the four looks, the identity-lock lines that keep a face or a product the same across images, and the edit formulas. Swap anything in brackets for your own details.
How to set it up. Fastest way. Copy the pack, pick the look you want, fill in the brackets, paste it into ChatGPT with your reference photo attached. To keep it, download the file and save it as a Claude Skill or a ChatGPT Project so you can say "write me a Look 2 prompt for this product" and get the finished prompt back.