Combine text, image and audio for one verifiable task
Choose only the media needed, preserve timestamps and source references and check the answer against each input rather than an invented story.
In brief
A multimodal task can combine a screenshot, spoken explanation and text question. Use those inputs to answer a defined need, such as locating a visible error, rather than asking for a broad interpretation of everything.
Keeping each input identifiable lets you check which piece supports a conclusion.
Choose a small complete task
For a setup problem, a readable error screenshot and your text description may be enough. A whole video is unnecessary if it adds private material without helping the question. For a video procedure, include the relevant interval and any audio needed to understand it.
Keep observations separate
Ask the model what is visible, what is heard and what is inferred. A spoken claim that a button was clicked is not the same evidence as a frame showing its resulting state. Preserve the origin of each observation when reconciling conflicting inputs.
Verify the decisive moment
Open the cited frame or interval and check labels, numbers and sequence. If a brief step is missing from the analysis, provide a more focused permitted input or inspect it yourself. Do not let a fluent combined summary conceal an unreadable screen or unheard condition.
Before you act
Supported media, length and sampling depend on the model and API. Video analysis can miss a brief event, and audio interpretation can mishear a crucial word.
What to do next
- Define the question and select the relevant permitted media.
- Label sources and preserve the needed page, frame or timestamp context.
- Ask for evidence references and verify important conclusions against the original inputs.
