Draw boxes on a photo, then edit it: Sketch-style markup on Sume

OpenAI's announcement thread says Sketch lets users draw directly in ChatGPT by typing @Sketch, to communicate a design visually (OpenAI Developer Community, read 2026-10-04). That panel lives in ChatGPT. The Sume Image API has no drawing surface, only a prompt and up to 16 reference images on ChatGPT Image 2.5 (Sume Image API). The useful part of Sketch is still reachable: draw on the picture yourself and tell the model what the drawing means.

The markup pattern

Reference order is yours to define in the prompt, so name each image explicitly and do not assume the model infers it.

• Draw numbered boxes on a copy of the photo, one box per change.
• Keep the clean original as a second reference so the model can see what is under the marks.
• In the prompt, say which reference is the marked copy and which is clean, then list each box and its change.
• State that the boxes and numbers must not appear in the result.

A prompt block that works as a template

The text below is a template, not a measured result. Treat the first output as a test and check that no box outlines leaked into the picture.

Image 1 is the clean photo. Image 2 is the same photo with numbered red boxes.
Apply these changes, each only inside its box:
1. Box 1: replace the sofa cushion with a mustard yellow one.
2. Box 2: remove the cable on the floor.
Keep everything outside the boxes identical to Image 1.
Do not draw any box, outline or number in the result.

What to check

Marks sometimes survive into the output. Diff the result against the clean original: if changes appear outside the boxes, tighten the wording and re-run. Failed generations are not billed on Sume, but a finished edit with a leaked box is, so look before you batch.