Generate a social caption from images you own

Produces a platform-shaped caption - and on headline platforms a headline and a description - from 1 to 10 images.

SCOPES: this route needs images:write to CALL, but its OUTPUT reads as text (D-15-23). The result is stored as a caption_generation, which classifies into the text container, so finding it later through /v1/gallery, /v1/library or /v1/jobs/{id} needs text:read. A token holding only images:write can run this and be charged for it, and then not be able to read the result back. The response body returns the caption directly, so a single call still works.

Give the images as library file_uuids (upload with POST /v1/files/upload) or inline as base64 in images; exactly one of the two. Limits: 10485760 bytes per image, 26214400 bytes across the call, context at most 2000 characters.

The prompt is built SERVER-SIDE. There is deliberately no raw prompt field: it would turn the caption credit path into a general-purpose vision prompt. context and the blueprint fields are untrusted text spliced into a model instruction and the builder emits a guard sentence about them - a mitigation, not a guarantee.