The four-part Halloween prompt
Part one is the identity lock: who is in the photo and what must not change. Part two is the costume: garment, hair, accessories and silhouette. Part three is the scene: location, time of day, practical lights and weather. Part four is the camera and negative rules: lens, grade, anatomy, text and watermark exclusions.
If you are prompting a model without image input, replace the identity lock with a character description. Keep the rest of the structure. The model still needs a subject, an outfit, a place and a camera.
Prompts for Gemini and ChatGPT
Gemini and ChatGPT can rewrite a prompt, but they do not always know which details matter for an image model. Give them the four-part structure and ask them to preserve the identity sentence. Do not ask for a “viral” prompt; ask for a specific scene with a clear light source.
When you paste a prompt into a chat model, ask it to return only the final prompt. That makes it easier to copy into the image workspace without commentary. If the model adds extra characters or props, remove them before generating.
How to iterate without starting over
Change one variable per run. If the face is wrong, fix the identity sentence. If the costume is wrong, change the garment sentence. If the mood is wrong, change the light or grade. Keeping a short prompt log helps you see which change actually helped.
Save the prompt that worked. Seasonal pages are reused every year, and a tested prompt is more valuable than a long list of untested ones.