Every format below was rendered with gpt-image-2 off a single product cutout. Open any card to copy the exact prompt that produced it, swap in your product, and get the same structure back.
Images came from gpt-image-2 via /v1/images/edits at 1024x1280, which is native 4:5, so no cropping step. The real product cutout is passed as the image reference, which is why the label never gets reinvented.
One angle, many formats. Do not write 40 different messages. Find the one thing your buyer needs to believe, then rebuild that single message as a Reddit thread, a text message, a tier list and a warning label. Same belief, four costumes.
The warning that matters more than the prompts. Leave a factual slot vague and the model fills it in confidently and wrongly. Ours invented a customer review claiming reduced illness, a star rating that contradicted the real one, and search results attributed to real publishers. Every number and every quoted line in these prompts is pinned to something checkable. Do the same with yours.