The four-part Halloween prompt
Part one is the reference lock: use the uploaded photo and preserve face, skin, pose, and the number of people. Part two is the costume: one outfit, one silhouette, two or three accessories. Part three is the scene: one location and one lighting direction. Part four is the negative constraint: no changed face, no extra limbs, no text, no watermark.
This structure gives the model a clear hierarchy. It knows what cannot change before it knows what should change.
Prompts for image editors versus text-only models
An image editor needs a reference sentence and a preservation sentence. A text-only model has no reference, so the prompt should describe the person more fully and should not claim to preserve a face it has never seen.
If you are copying a recipe into Gemini, ChatGPT, or another assistant, tell it to use the attached image as the reference. A prompt without the image attached is just a scene description.
How to adapt a recipe without breaking it
Change one bracket at a time. Replace the costume, then run it. Replace the location, then run it again. If you change every bracket at once, the result tells you nothing about which choice worked.
Keep the avoidance line at the end even when you shorten the rest. It is the cheapest protection against text artifacts, extra fingers, and watermarks.
What a good Halloween prompt should not contain
Do not ask for a celebrity likeness, a copyrighted character, or a real person you do not have permission to use. Do not ask for gore, hate imagery, or content involving minors in a sexualized context.
Do not stack ten style references. One photographic direction, one costume, and one location is enough. A long prompt with contradictory light directions usually produces a flat, synthetic image.
How to test a prompt without wasting credits
Test the identity lock before you test the costume. Run the prompt with a plain background and a simple outfit, and check whether the face still looks like the source photo. If the identity is already drifting, a more elaborate costume will not fix it. Once the face holds, add the costume, then the scene, then the lighting.
Keep a small prompt log with four columns: the source photo, the prompt version, the model lane, and the result you kept. After three or four runs you will see which sentence is doing the work. Most people end up with a short prompt of five or six sentences rather than a page of style keywords, because every extra instruction competes for attention.