Name the region
Name each image, identify the region, and separate what should be observed from what may be interpreted.
Beyond text
Prompts for images, PDFs, audio and video
State which file, page, region, speaker or time range the model should use — and whether it should observe, compare, extract or interpret.
Keep page references
Require a page, table or paragraph for every important finding. Verify OCR and small text.
Separate speakers and confidence
State the language and purpose. Keep timestamps and flag uncertain transcription.
Bound the time range
Ask for events, changes or quotations within a specific sequence, with timestamps.
A four-part multimodal brief
Material:
- Image 1: [what it shows / source]
- Image 2: [what it shows / source]
Task:
Compare only [named region or property].
Working rules:
- First describe what is actually visible.
- Separate observation from interpretation.
- Do not guess text or details that cannot be read.
- State which image and region supports each finding.
Deliver:
A table with observation | possible interpretation | uncertainty | recommended check.Three common failures
- Unclear reference: “What is the difference?” without naming the images or properties to compare.
- False precision: asking the model to identify tiny text, an exact count or spatial position without verification.
- Hidden instructions in files: treating document or image text as commands rather than untrusted content.
Sources
- Anthropic: Vision — images, order and limitations
- Google: Prompt design strategies — multimodal structure and references
- OpenAI: Images and vision — official vision guidance