A sketch is structure, not content

Sketch-to-image looks like the easiest mode to build. You have an image, you have a prompt, you send both. It is the mode with the most specific failure, and the failure is funny once and expensive after that.

The model reproduces your handwriting

People do not submit clean line art. They submit a box with logo here scrawled in it, an arrow labelled make this bigger, a circle with product written inside.

Sent as a plain reference, the model treats all of that as content. You get a finished, well-lit, professionally rendered image of a box with the words "logo here" written on it in a nice font. It is technically a perfect execution of the input.

The fix is a clause that reclassifies the whole attachment:

Use the attached sketch as the composition and structure only.
Follow its layout, proportions and placement, and read any handwritten
words in it as instructions describing what to draw rather than as text
to reproduce. Render the result as a finished image, not as line art.

Three separate jobs in one paragraph, and dropping any of them brings back a distinct failure. Without composition and structure only you get a cleaned-up drawing. Without the handwriting clause you get labels rendered as artwork. Without not as line art you get a tidier sketch instead of a photograph.

Why it is appended, not prepended

The intuitive order is to lead with the instruction. It produces images that are about the instruction — models weight the opening of a prompt most heavily, so a prompt that opens by discussing how to interpret an attachment is a prompt whose subject is interpretation.

The user's own sentence goes first. The clause is a qualifier on it.

The server has to add it

If the browser composes the final prompt, the clause is decoration: it can be stripped by anyone posting directly, and it silently stops being applied the first time somebody refactors the client. The structured intent travels; the server composes.

A useful side effect is that the stored prompt stays the user's own words, which is what you want when they ask why the picture came out the way it did.

The thing still worth telling users

Contrast matters more than neatness. A pencil sketch photographed on a grey desk under a window gives the model a much weaker structural signal than the same sketch photographed straight on with the paper filling the frame. That is a one-line hint in the uploader, and it moves outcomes more than any prompt engineering.

The mode itself, with the clause it sends, is at gptimage25.top/sketch-to-image.