Prompt with images and documents

When a request includes images or PDFs, a few habits make answers markedly more reliable.

Put the material first

Place images and documents before your question in the message. The model reads the material, then the instruction, and the question stays fresh at the end.

Name what you send

When there are several files, label them: “Image 1 is the shelf photo; image 2 is the planogram.” For documents, use tags with names (<document name="2025-msa.pdf">) so answers can point to the right one.

Ask for what you can check

• “List each line item you can read on the receipt, then the total” is checkable; “is this receipt valid?” is not.
• Ask the model to say when something is unreadable rather than guess: “If a value is blurred or cut off, write unreadable.”
• For documents, ask for a page number or a quote with each claim, or turn on citations.

Image quality

• Crop to the part that matters; a phone photo of a whole desk wastes tokens on the desk.
• Keep text in images at least 12 pixels high after any resizing.
• Rotate images upright before sending.

Charts and tables

Ask for the data first, then the interpretation: “Write the values from the chart as a table, then describe the trend.” Errors in reading are then visible instead of hidden inside a conclusion.

Cost

Images and PDF pages are billed as input tokens; a full-page scan costs far more than its extracted text. If you only need the text of a born-digital PDF, extract it yourself and send text. See Vision and PDF support.