# Master prompt with reference-image roles
A long, fixed prompt that tells the image model which inputs it is receiving and what each is for:
The three references are plugged into the Nano Banana node in that order: the 3D node first, product detail second, extra details third. The labels come from Compositor annotation layers (labelled references).
The whole thing is built to be reusable. The long prompt, the annotated reference images and the labelled node layout are not touched between runs. Per image you supply only two things:
The prompt in the middle updates itself from those inputs. The length is not a barrier, because the user never edits it: it is "not so scary, it's just long."
Without roles, the model may copy the ugly 3D look from image A, or take pose from a beauty photo. Naming each role removes those freedoms, in the spirit of Prompt layers exist to remove model freedoms. This is the prompt side of Split placement from appearance: a stand-in sets geometry, the image model sets looks.
Each reference image is labelled with a text layer in its corner (see Compositor annotation layers (labelled references)). The prompt can then refer to "image A/B/C" and give each one a single, distinct role.
Giving each image a role, and telling the layout image what to ignore, stops the model from blending pose, material and lighting cues across references.