Multimodal systems are only useful when users can predict what the model understands. The interface must translate messy input into clear intent.
Start from user confidence
Users need answers to three questions:
- What did the system ingest?
- What does it think the task is?
- How can I correct it quickly?
Interaction principles
- show source previews for every modality
- surface confidence states in plain language
- keep correction actions one click away
Designing with friction in the right places
A model can process ten inputs, but that does not mean the user should provide ten inputs every time. Use progressive disclosure and ask for more evidence only when confidence drops.
Example prompt assembly pipeline
const buildTaskContext = (inputs: string[]): string => {
const normalizedInputs = inputs.map((input) => input.trim()).filter(Boolean);
return normalizedInputs.join('\n\n---\n\n');
};Product takeaway
Great multimodal UX is not about exposing model power. It is about making intent legible from first interaction to final output.
For teams running design reviews, publish shared patterns in W3C WAI guidance language so accessibility intent is explicit.
