A ramen prompt returned an editable recipe artifact: ingredient list, numbered steps, and a serving-size input that recomputes the quantities. Change "4" to "6" and the ingredients update.

A Three.js fragment shader was written and rendered in an interactive 3D viewer in the same turn. The viewer is stateful — camera angles persist, the scene keeps running, and the user can rotate around to see the disk from any angle.

Two artifacts that work together: a full-month itinerary with week-by-week cards, and a recommendation card listing curated stops. Both are anchored to a Google Map with pinned locations. Reading the plan is itself an act of navigation.

The user sees the answer being built. A streaming preview of "Create the month-long autumn Japan trip plan as a markdown artifact" lands before the final UI does, with an explicit loading state in the chat.

An NBA recap prompt returned a sports analytics surface — team logos, final score, an MVP card, a quarter-by-quarter chart, comparison bars across categories. None of the visual structure was in the prompt. Claude inferred the right shape of the answer.

Asked for a poem about a rainy day, Claude produced a typeset HTML artifact with custom typography and atmospheric styling. A second variant added a CSS-driven rain animation. The same content rendered with two very different intents.

A "what is a qubit" prompt turned into a comparison artifact: a coin lying flat for the classical bit, a coin spinning in the air for the qubit. A second view added a "Flip the bit" toggle and inline code. The UI is doing the teaching, not the paragraphs.

When Claude needs information from the user to continue, it asks with a UI, not a question. Numbered options, a free-text fallback, keyboard navigation hints, a progress indicator. The MCQ is keyboard-first and lives inside the chat.

A weather query produces a city card with current conditions, a five-day forecast strip, and an atmospheric background that reflects the condition. Numbers are anchors. The visual is the story.

A prompt about chat-management strategy turned into a flowchart with branching decision logic. The diagram does what a paragraph cannot: it makes the structure visible at a glance. The flowchart is not a supplement to the answer; it is the answer.

An email-drafting prompt returned a tabbed composer with two tonal variants — "Direct ask" and "Invite discussion" — each a complete email with a subject line, body, and an "Open in Mail" handoff. A small but telling pattern: Claude generates options, not the answer.

When an artifact needs to run code, Claude does not silently execute. The permission card surfaces the request inline in the conversation, shows the actual code that would run, and offers explicit "Always allow" and "Deny" controls. Consent is part of the UI grammar, not a hidden setting.