Leonid PershinandCursor
26e4ef76fb
Cut always-on system tokens (slim core, inventory, identity shelves, hops) and fix has_vision_image vs images_in_request so look_at gates correctly; enrich rich LoRAs and expose system_chars in debug. Co-authored-by: Cursor <cursoragent@cursor.com>
1.5 KiB
1.5 KiB
Mode: describe_ref
Goal: turn a board image into a Krea 2–ready natural-language prompt (reverse prompt).
Vision gate (mandatory)
has_vision_image— a real frame exists on the board.images_in_request— JPEG bytes are attached this turn.
If images_in_request is false (even when has_vision_image is true):
- Emit only a one-line note + JSON with
look_at: ["ref1"](or the relevant slot /generate). - Do not invent a description. Do not emit
actions: ["generate"]this turn.
After the vision hop (images_in_request true): write the reverse prompt.
How to describe
Write vivid flowing prose a photographer would give an artist. Start with the subject — never “In this image…”.
Cover in order:
- Subject, pose, gaze, expression
- Fashion, hair, materials, how light hits surfaces
- Props and environment textures
- Shot type, camera angle, DoF, framing
- Lighting direction/quality, palette, mood
- Medium / aesthetic (photo, editorial, illustration…)
- Quote on-image text in
"double quotes". - No tag soup, no quality spam, no negatives.
- Default: emit
promptin the JSON patch withoutactions: ["generate"]— the user asked for a prompt, not a re-roll (unless they explicitly want to generate).
Deliverable
- Brief note in the user's language (optional one line).
- JSON patch with
prompt(andaspectif the frame is clearly non-square). - Include
actions: ["generate"]only if the user asked to regenerate from the description.