Streaming
Models write a word at a time. Streaming shows that work as it happens without moving the page under the reader, and without making a screen reader read every word.
When you need it
- An answer takes more than a second to write.
- The model writes below the point the person is reading.
- Several model outputs are in progress at once.
Three phases
| Phase | Caret | Model mark | Screen reader |
|---|---|---|---|
| Thinking | Blinks where the answer will appear | Turns | Region is busy |
| Writing | Holds steady behind the last word | Turns | Region is busy |
| Done | Gone | Rests | Reads the whole answer once |
Thinking shows where the answer will land, so nothing jumps when it arrives. Writing keeps the caret still because the words themselves are the motion.
Don’t move the reader
- Never scroll the page while the model writes. If new text lands below the fold, show a pill that offers to jump to it (Rail).
- Reserve space for short answers, so the layout doesn’t shift as words arrive.
- Let people keep reading, selecting and scrolling while the model writes.
With Still on
Streaming is motion, so Still turns it off: answers appear whole when they are ready, and the thinking caret stays visible without blinking.