damngoodprompts
Menu

Image and multimodal prompting

Prompt around what the model can see, what it should inspect, and what uncertainty should look like.

Last reviewed: August 10, 2026GeminiChatGPTClaudeGrok
Technique

Best suited to Gemini - strong fit when the task depends on image, document, audio, or video context in the same workflow

The task

You want the model to inspect an image, screenshot, chart, document scan, or mixed media input. The prompt must say what visual evidence matters and how the answer should handle uncertainty.

Why the old approach is outdated

“Describe this image” is too broad for most work. It encourages a caption, not a decision.

Multimodal prompts work better when they name the inspection target: layout issues, visible text, object presence, chart meaning, document fields, or mismatch between screenshot and requirement.

If you are not sure why image, audio, video, and file inputs change request size, start with tokens and context windows.

The current approach

Inspect this checkout screenshot for usability problems.

Focus only on visible evidence: labels, field order, error messages, contrast, spacing,
and unclear actions. Do not infer analytics, user intent, or backend behavior.

Return blocking issues first, then smaller polish items. If something is hard to read,
say so instead of guessing.

When to use a different model

Use ChatGPT when the visual task is part of a broader writing, coding, or product workflow. Use Claude when the model must compare images against a long written brief. Use Grok only when the image task also depends on current public context and the selected model supports the needed input.

What to avoid

  • Asking the model to identify tiny details without allowing uncertainty.
  • Treating a model’s visual answer as measurement.
  • Mixing visual inspection with unsupported claims about what users did.
  • Omitting the image’s purpose. A design QA prompt is different from an accessibility prompt.

Source set