GPT Image 2.5 menu board prompt: quote every line, then proofread

To make a menu board with GPT Image 2.5, put every line of the menu in the prompt inside quotation marks, say how many items and sections there are, give the layout in words, and request a portrait or landscape size that fits the screen. Then proofread the output letter by letter, because a menu is the kind of image where one wrong digit in a price becomes a real problem.

This post gives a prompt template, a request with a valid custom size, and a checklist. Parameter facts come from the Image API docs, read 2026-10-03. Nothing here claims the model never misspells; it says how to catch it.

How do I write the prompt?

Describe the board first ("chalkboard menu, white hand-lettered text, three sections"), then list each section with its exact text. Put the real text in straight quotes and keep each price attached to its item on the same line. State the count: "exactly 6 items, 2 per section". Close with one constraint line: "No other text, no logo, no extra items."

Keep the copy short. Fewer, shorter lines give the model less to get wrong. If you have 30 dishes, make three boards.

Chalkboard cafe menu, portrait, dark slate background, white hand-lettered chalk text, three sections with underlined headings.
Section 1 heading: "COFFEE". Items: "Espresso  3.00", "Flat white  4.00".
Section 2 heading: "TOAST". Items: "Avocado  7.50", "Ricotta & honey  6.50".
Section 3 heading: "SWEETS". Items: "Almond cake  4.50", "Brownie  3.50".
Exactly 3 headings and 6 items. No other text, no logo, no extra items.

What size and quality should I ask for?

Sume's custom image_size needs both edges to be multiples of 16, a maximum edge of 3840, an aspect ratio of at most 3:1, and 655,360 to 8,294,400 pixels. A 1080x1920 screen is not valid as typed, since 1080 is not a multiple of 16; ask for 1088x1920 and crop, or use 1024x1792 and letterbox. Quality defaults to high when omitted, and small text is the reason to keep it there for the final.

Try the layout at medium, then rerun the approved wording at high.

How do I proofread the result?

Do it every time; model text is plausible, not guaranteed.

• Read each line aloud against your source list, character by character, prices included.
• Count headings and items. If the model added or dropped one, say the count again in the prompt.
• Zoom to 100 percent; small letters often fail where large ones pass.
• Check duplicates: a heading rendered twice is a known failure, so state how many times each appears.
• Have a second person read it before it goes to a printer.

How do I fix one wrong line?

Do not regenerate the board. Use an edit: pass the board as an input_references URL, add a mask_url covering just that line, and say "change only the masked text to \"Flat white 4.00\"; keep everything else identical". Sume lists mask_url for GPT Image 2.5 edits. Set aspect_ratio: "auto" so the size matches, and check that the rest of the board did not shift.

Each completed edit is billed, and a failed one is not, so a one-line fix costs one generation rather than the whole board again.

When should I not use an image for the menu?

If prices change weekly, a generated image is the wrong source of truth. Generate the background art with GPT Image 2.5 and set the text in your own design tool, where a price edit takes seconds. Use the model for the look, and a text layer for the words that must be right.

What about other languages?

The same method applies to a menu in another language: quote every line in the target language and proofread with a native reader. Sume's docs make no per-language accuracy claim for GPT Image 2.5 text, so treat accents, non-Latin scripts and long words as higher risk and zoom in on them during review. If you translate an existing board, ask the model to translate the text and change nothing else, then verify every line again.

Print the final board only after the proof is signed off, and keep the exact quoted copy in a file so the next edit starts from the true text instead of from what the image appears to say.