Claude's output teardown

Updated June 15, 2026

Most Claude replies never leave the message stream. A factual answer like “why is the sky blue?” renders as prose with per-message actions, copy, read aloud, feedback, Try again, or quote a passage to refine.

When the work is substantial enough to ship, Claude graduates it to an artifact. See the artifacts teardown for split pane, preview, code, and publish.

Short answers stay in the thread

Factual Q&A with copy, read-aloud, feedback, and Try again. No side pane needed.

What works

Not every answer becomes an artifact. Simple replies stay in the stream so reading flow stays uninterrupted.
Per-message actions (copy, read aloud, thumbs, Try again) sit directly under the answer. Refinement targets one turn, not the whole chat.
Edit on the user bubble supports thread branching when someone revisits an earlier prompt.

What we would push on

Try again and edit sit beside each other with equal weight. First-time users may not know edit forks while Try again keeps the same branch.

Takeaway

Default to in-thread output. Artifacts should feel like a promotion, not the norm.

Pattern: Response Refinement

Pattern: Thread Branching

Try again without losing the first draft

After Try again, 2/2 navigation lets users compare versions of the same answer.

What works

Try again generates an alternate answer instead of overwriting the first one.
1/N navigation with Previous version tooltip makes comparison explicit. Users can pick the better draft or keep exploring.
Versioning applies to in-thread text answers before any artifact opens.

What we would push on

Version controls are small and easy to miss under long answers. High-stakes edits may need a clearer “you are viewing version 2” banner.

Takeaway

Non-destructive Try again lowers the cost of asking for a redo. Pair it with visible version navigation.

Pattern: Response Refinement

Thumbs up opens structured feedback

Optional details field asks what was satisfying; submit sends the conversation for model improvement.

What works

Thumbs up sits in the per-message action row beside copy, read aloud, and Try again. Feedback targets one reply.
The modal is lightweight: one optional text field (“What was satisfying about this response?”) plus Cancel and Submit.
Disclaimer states the full conversation may be sent to Anthropic, explicit consent before submit.

What we would push on

Modal interrupts flow for a simple thumbs up. One-click satisfaction with details optional on expand might feel faster.

Takeaway

Pair low-friction thumbs with an optional qualitative capture when users want to explain what worked.

Pattern: Feedback

Thumbs down adds issue taxonomy

Negative feedback adds an issue-type dropdown plus optional details before submit.

What works

Thumbs down opens a richer form than thumbs up, issue type dropdown helps route reports without free-text guessing.
Optional details field still available for specifics (“What was unsatisfying about this response?”).
Same conversation-disclosure copy as positive feedback keeps trust symmetric.

What we would push on

Asymmetric depth (dropdown on down, not up) makes sense for triage but may feel punitive. Keep the down path fast if users skip the form.

Takeaway

Use structured issue types on negative feedback where you need to categorize failures; keep positive path lighter.

Pattern: Feedback

Reply on a selection

Highlight part of the answer; Reply tooltip targets refinement to that passage.

What works

Users can select prose inside an assistant message, not just react to the whole bubble.
Reply tooltip appears on the selection so the follow-up is scoped (“fix this bullet”) instead of vague (“make it better”).
Works on list items and paragraphs alike; selection stays highlighted until the user acts.

What we would push on

Reply on selection is discoverable only after highlighting. No persistent affordance for users who do not try select-first.

Takeaway

Let users anchor refinement to a span of text when the issue is local, not global.

Pattern: Response Refinement

Quote lands in the composer

Selected passage appears as a quote chip above the prompt; user adds the fix on top.

What works

Quoted text carries into the composer as a removable chip, Claude sees exactly which lines the user is talking about.
Send stays available once the user types a follow-up; quote + new instruction ship as one message.
Alternative to Try again when the user wants a surgical edit, not a full rewrite.

What we would push on

Quote chip truncates long selections. Verify the model still receives full context on multi-paragraph quotes.

Takeaway

Quote-to-reply is refinement without regenerating the whole answer, pair it with Try again for full rewrites.

Pattern: Response Refinement

How output fits with artifacts

Claude’s output model starts in the thread: quick answers, Try again with version history, thumbs up/down with optional detail modals, quote-to-reply on selections, and edit-to-branch when the user changes direction.

When the deliverable is code, a document, or an interactive tool, Claude promotes it to an artifact, card in the reply, then split pane beside the chat. That handoff is covered in the artifacts teardown.

Compared with ChatGPT’s per-message refinement menu, Claude spreads refinement across Try again, version navigation, and follow-ups while an artifact is open.