AI answer

Tab

Moves through New thread, Share, the reasoning and sources rows, the source cards, each citation, the table and code regions, code Copy, the action bar, the related questions and the composer.

Enter, Space

On the reasoning or sources row, opens or closes it. On a citation, opens the source in a new tab; focusing it already shows the source card and lights its sentence and card.

Enter

In the composer, sends the question. Shift+Enter adds a line.

Arrow uporArrow down

In the Regenerate menu, moves between options; Arrow right opens Switch model and Arrow left closes it.

Escape

Closes the Regenerate menu, the reason picker or a citation card and returns focus to its trigger.

• The block is a labelled region named by the question, with aria-busy while an answer streams. A polite status region announces Thinking, Writing the answer, Answer ready, Answer stopped or the error.
• Action results (copied, link copied, feedback sent, version shown) go to a second polite live region; the share links say when the preview copies a sample link.
• The reasoning and sources rows are buttons with aria-expanded and aria-controls; closed panels are inert, so their content leaves the tab order.
• Citations are links named "Source n: title, site"; the source card opens on keyboard focus as well as hover, and a lit sentence is a visual aid only, never the only cue.
• The comparison table keeps header cells with scope, and the table and code scroll inside focusable, labelled regions on narrow screens.
• Status is never color alone: copied shows a drawn check, errors an icon and text, feedback a filled glyph with aria-pressed, and stopped or failed answers a note with a Try again button.
• Asking a related question moves focus to the new question heading; New thread moves focus to the composer.
• Pointer focus draws no ring; keyboard focus shows Arc's shared focus ring.

• "Thinking" uses the Arc text shimmer for about a second and settles to solid text as "Thought for 9s" when the first words arrive.
• The answer streams in chunks on a requestAnimationFrame clock: about every 34ms, three to eight words commit at once and fade in over 160ms on a plain ease-out, with no blur, travel or stagger. A steady pace of about a hundred words a second plus catch-up drains bursts, so a full answer lands in two to four seconds. Settled text never re-renders or reflows.
• Headings, list items, table rows and code lines appear as their text arrives; a line or row waits for its newline, so containers never pop in half built.
• The reasoning and sources panels open on the smooth spring; source cards fade in 20ms apart, at most six steps. Hovering the sources row fans the mark chips apart on the CSS spring.
• Regenerating or switching versions crossfades the answer while its height springs to the new length, and the version number rolls up or down.
• Copy and Share morph their icon (a drawn check) and label in place without changing width; Send turns into Stop in the same spot.
• Reduced motion keeps the chunked arrival but drops every fade, roll, fan and spring, and keeps all states and announcements.

• The reading column caps at about 704px with even gutters that shrink to 16px on phones; the question scales between 22px and 28px with the container.
• Source cards form a two column grid in a wide column and a snap scrolling row on phones, faded at the trailing edge with a mask; a citation that lights an off screen card scrolls it into view.
• The table and code scroll horizontally inside their frames; the action bar wraps its model line under the actions.
• The top bar truncates the title and keeps Share; the composer stays docked at the bottom and grows to six lines.

• Streamed markdown is parsed into numbered units that stay stable as text arrives, so each commit only appends new spans and earlier text never re-renders.
• One requestAnimationFrame loop per answer paces the reveal and commits at most every 34ms; it stops as soon as the answer is on screen.
• Finished versions render whole without the reveal, and the answer body is memoized against hover changes outside its sentences.

• Choose it for a single answer page with sources: help centers, internal search, docs assistants, answer engines. It shows one question at a time; asking another replaces it.
• Stream markdown in `text` events: paragraphs, ### headings, 1. and - lists, pipe tables and fenced code (```js filename.js). Inline **bold**, `code` and [n] citations work everywhere; [n] points at sources[n - 1].
• Send `sources` early, during reasoning, so the stack fills while the model thinks. Send `thought` with the thinking duration when the first text is ready; without it the time to first text is used.
• Brand marks are sample content. Pass your own `icon` per source; keep color brands in their colors and mono brands on currentColor.
• Regenerate keeps every version in memory for the session and calls `respond` with mode and the previous answer as markdown, so your prompt can ask for a shorter, longer or different answer.
• The demo compresses time: thinking plays in about a second and the reported thinking time and duration are what the simulated model would have taken.

The full library index for assistants is at /llms.txt.