A sticker pack succeeds when each image answers a recurring moment. Begin with eight message jobs you already use: yes, no, thanks, wait, on it, wow, done, and laugh. Then give every job a different silhouette. A raised thumb can carry yes, a flat palm can carry wait, and a small bow can carry thanks. At message size, the body shape should communicate before anyone reads the caption.

Keep each caption to one to three words and make the language sound like the person or brand. A creator might choose LET'S GO, ONE SEC, and POSTED. A shop might use PACKED, THANK YOU, and ON THE WAY. Write the reaction map first so ChatGPT is solving a set with eight distinct functions. Without that map, image generation often gives you several versions of happy and no useful answer for the awkward moments.

Use a portrait only when you have permission to use the likeness. An original mascot works just as well and is often easier to repeat. Describe the features that must survive every image: face shape, hair or mascot geometry, one outfit, two main colors, one accent, outline weight, crop, and finish. Ask for one coherent pack in a single request so the model can see the reactions as related parts of the same visual system.

The identity lock should be short enough to repeat verbatim in every repair. For example: same rounded face, black bob, plum sweatshirt, warm cream highlights, thick charcoal outline, chest-up crop, transparent background. Avoid vague labels such as cute, premium, or viral unless you translate them into visible decisions. Consistency is not sameness. The expression, hand pose, and caption should change, while the face, palette, line quality, scale, and edge treatment stay fixed.

Do not regenerate all eight because one hand, word, or expression missed. Save the strongest pack, isolate the failed tile, and name the smallest visible correction. Attach the approved pack as the reference and repeat the identity lock before the repair: same face, same outfit, same palette, same outline, same crop. Then request only the correction, such as change the right hand to a flat wait gesture and render WAIT exactly once.

OpenAI notes that image-edit selections are not always precise and that an edit can extend beyond the highlighted region. Treat the selection as guidance, not a perfect fence. After every repair, compare the face, outline thickness, background transparency, and scale with the other seven stickers. If a nearby feature changed, return to the approved tile and narrow the instruction. The goal is not a prettier isolated image. It is a repaired member of the same set.

A large preview can hide the failures that matter in a chat. Reduce every sticker to roughly the size it will occupy beside a message. Can you identify the reaction without zooming? Is the face still readable? Does the caption survive? Does the outline separate from both a white and near-black background? Look for pale halos, clipped hands, stray pixels, duplicate features, misspelled words, uneven margins, and one sticker that appears much larger than the rest.

OpenAI says the new mobile experience can download a pack or add it to iMessage or WhatsApp on supported devices, and that generated images remain available under Images for later download or sharing. Save the approved original before installing anything. Keep a simple pack sheet with the eight jobs, final captions, identity lock, and approval date. When you add a ninth reaction later, generate against that sheet instead of asking the model to remember an old conversation.