An assistant answers word by word while tool-call chips spin, resolve to a check and settle under a blinking caret
This is the format it was composed for. Switch above to see it lay itself out for another: same animation, re-arranged for the frame.
Everything below can be changed without touching the animation. Hand the list to an agent, or edit the values at the bottom of the component file.
The user's message, shown as a chip at the top right of the transcript. One sentence, the answer is the subject here, not the question.
now: Turn the launch notes into a 30-second film.
The assistant's first reply, revealed word by word before the tools run. 20–30 words is the sweet spot for the 8s budget; longer is streamed faster to fit.
now: Reading the notes now. I'll pull the headline, three feat…
The closing reply, revealed after the last call resolves. This is the line that sits on the held frame, so make it the payoff, 12–18 words.
now: Six scenes, twenty-nine point eight seconds. Preview is r…
Tool-call chips, appearing between the two replies and overlapping by half a second. Two is the honest number for an 8s scene; three is the ceiling and squeezes the prose.
now: 2 items
Name in the header, beside the accent mark. A product name works better here than a model name.
now: Assistant
Mono model tag in the metadata rail. Empty string drops the row.
now: registry-1
Streaming speed multiplier. 1 is ~13 words a second, which is what a fast stream looks like; 0.7 reads more deliberate. Long copy is sped up automatically to make the budget regardless.
now: 1
Show the metadata rail, model, token count, tool progress, elapsed clock. Ignored on portrait and square frames, where the transcript takes the full width instead.
now: true
Seeds the per-word jitter that keeps the stream off the metronome. Change it for a differently-paced take of identical duration; the same seed always renders identical frames.
now: stream-01
Brand token id, one of the keys of BRANDS.
now: slate
No Tailwind, no CSS, no asset files. Inline styles only.
An AI chat panel where the assistant's answer streams in word by word, a couple of tool calls appear inline as small chips with a spinner that resolves into a check and a muted result line, and the caret is still blinking on the last frame. Precise and quiet, not flashy.