ChatGPT's output teardown

What works

• Long answers use clear hierarchy , title, intro line, numbered list with bold lead-ins , without opening a side pane.
• Thought for 5s sits above the answer as a lightweight signal that reasoning happened, without cluttering the reply.
• Per-message actions stay below the bubble once the answer completes.

What we would push on

• The Thought label is easy to miss on fast scroll. Users who want transparency may not discover Activity without trying the chevron.

Takeaway

Default to readable in-thread formatting. Surface reasoning as an optional expand, not inline noise.

What works

• Activity opens beside the chat so the main answer stays clean while curious users inspect planning bullets.
• Panel header shows duration (Activity · 5s) and a Done state when reasoning finishes.
• Copy and Edit icons on the Thought row offer quick export or correction without opening the full panel.

What we would push on

• Activity is a second surface to learn. Mobile or narrow layouts need a clear back path so users do not lose the thread.

Takeaway

Separate reasoning trace from deliverable answer. Let power users opt in via one obvious affordance.

What works

• Regenerate creates alternate answers instead of overwriting the first draft.
• < N/N > controls sit in the per-message action row, version history is one click away from copy and feedback.
• Each version keeps the same action set (copy, thumbs, share, regenerate, more).

What we would push on

• Pagination is subtle on long answers. High-stakes edits may need a stronger “viewing version 4 of 4” cue.

Takeaway

Non-destructive regenerate lowers the cost of trying again. Pair it with visible version navigation.

What works

• Positive feedback completes in one tap with an inline tooltip, no form interrupting flow.
• Thumbs up sits beside copy, share, and regenerate so feedback targets one reply.
• Asymmetric to thumbs down: satisfaction is fast; dissatisfaction gets structure when users opt in.

What we would push on

• One-click up means no qualitative capture on praise. Product teams may want an optional “tell us more” on long press.

Takeaway

Keep positive feedback frictionless. Reserve forms for when you need categorized failure signal.

What works

• Thumbs down opens Share feedback with six chips, Incorrect or incomplete, Not what I asked for, Slow or buggy, Style or tone, Safety or legal concern, Other.
• Optional details textarea captures specifics without forcing prose.
• Disclaimer states the conversation may be included, explicit consent before submit.

What we would push on

• Submit stays disabled until a chip is selected. Users who only want to vent in free text may bounce.

Takeaway

Use chip taxonomy on negative feedback for triage; keep the positive path lighter.

What works

• Share targets a single assistant message, not the whole chat. Useful for checklists or one-off answers.
• Preview card shows title and opening lines so sharers know what recipients will see.
• Copy link plus X, LinkedIn, and Reddit cover common distribution paths from one modal.

What we would push on

• Public links need clear expiry or revocation UX elsewhere in settings, the modal alone does not show access scope.

Takeaway

Message-level share turns a good answer into a portable artifact without exporting the full thread.

What works

• Overflow keeps the action row clean, secondary flows live behind the … menu.
• Branch in new chat forks exploration without losing the original thread.
• View sources and Read aloud extend the same message into research or accessibility modes.

What we would push on

• Branch and sources are hidden until users open overflow. First-time users may never find them.

Takeaway

Put regenerate and feedback in the primary row; tuck fork, sources, and read-aloud one level deeper.

What works

• Users can select prose inside an assistant message. Refinement scopes to one list item or paragraph.
• Ask ChatGPT keeps the follow-up in-thread; Start writing promotes the selection into an editable block.
• Floating menu appears on the selection so intent is explicit before the user types.

What we would push on

• Selection menu is discoverable only after highlighting. No persistent hint 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.

What works

• Start writing turns a passage into a bordered block with its own Edit, copy, and Expand affordances.
• Block stays in the thread, context from prior turns remains visible above.
• Expand likely opens a fuller canvas; Edit signals the user can co-author without leaving chat.

What we would push on

• Two promotion paths , Ask ChatGPT vs Start writing , may confuse users who only want a quick rewrite.

Takeaway

Use writing blocks when output graduates from a reply to something worth editing and sharing.

ChatGPT’s output model starts in the thread: formatted prose, optional Activity for reasoning, regenerate with version history, lightweight thumbs up, structured thumbs down, message share, and overflow for branch, sources, and read aloud.

When a passage needs co-editing, selection → Start writing promotes it to a writing block with Edit and Expand. Image generation is a composer tool , see the composer teardown , with results landing inline in the same thread.

Compared with Claude’s Try again plus quote-to-reply, ChatGPT spreads refinement across regenerate, selection actions, and writing blocks while keeping most answers in-stream.