Qwen output: keep the chain in the thread

What works

• The model shows an art-direction brief before pixels. Users can edit scope before spend.
• Mode controls stay armed through the reply. The job does not reset mid-thread.

What we would push on

• A long single block is hard to scan. Call out what changed from the upload.
• Send readiness should be obvious when the expanded prompt is the thing to approve.

Product bet

Showing prompt expansion builds trust for image spend. Users feel the model understood the reference before GPU time runs.

Tradeoff

Takeaway

For reference-image gen, show the rewritten prompt in-thread and let users edit it before render.

What works

• Chain actions live on the image. Next jobs stay on the artifact, not buried in +.
• Message-level actions cover feedback and export without leaving the thread.

What we would push on

• Aspect label and rendered result must agree. Lying chips break format trust.
• Accuracy disclaimers that sit under the bar lose to confident output above.

Tradeoff

Takeaway

Put chain actions on the image itself. Match aspect label to what actually rendered.

What works

• Source stays pinned in the bar. Context is visible before the next spend.
• Mode follows intent. Short language can swap Create Image to Create Video without re-upload.

What we would push on

• Silent chip swaps teach less than a one-line mode change notice.
• Motion controls belong at prompt time if video is a first-class job.

Product bet

Image-to-video is the upsell path. Reference plus auto mode swap keeps users in one thread instead of a separate video app.

Tradeoff

Takeaway

When a reply references prior media, pin the thumbnail and swap mode chips to match the next job.

What works

• A progress card makes render feel like a job, not a stuck chat.
• Mode settings survive the wait so people can prepare another pass.

What we would push on

• Percent alone is thin for long renders. Stage copy or ETA reduces anxiety.
• Blocking the next send until render finishes kills queueing.

Takeaway

Show percent for video gen, but add stage copy or ETA when renders run longer than a few seconds.

What works

• Preview with duration treats video as a deliverable, not a chat attachment.
• Download and publish sit as equal primaries. Export is the goal.
• Mode bar stays available for another pass without losing context.

What we would push on

• Publish needs a named destination. Trust requires knowing where it goes.
• Aspect mismatch between setting and preview is the same format lie as images.

Tradeoff

Takeaway

Treat video like a deliverable: preview, duration, download, and name the publish target on the button.

What works

• A retrieval signal and inline domains sell "this was researched."
• Structured product lists scan better than a wall of prose for shopper queries.

What we would push on

• Citation chrome can costume a hallucination. Authoritative layout on a fictional product is worse than a plain miss.
• Domains without claim-level links make audit hard. Numbers beat orphan chips.
• Tiny disclaimers lose to confident lists above them.

Product bet

Search mode competes on Perplexity-shaped answers. Retrieval chrome sells freshness even when the model fills gaps with fiction.

Tradeoff

Takeaway

If you show domain chips, tie each claim to a numbered source and surface when retrieval failed or dates look wrong.

What works

• A dedicated audit drawer separates reading from verification.
• Titles, snippets, and overflow count show breadth at a glance.

What we would push on

• An optional panel does not fix a bad main answer. Most people never open it.
• If snippets repeat the same fiction, transparency is not a cure.

Tradeoff

Takeaway

A source drawer helps only if snippets are real. Pair it with date checks and a low-confidence state in the main answer.

What works

• Selection refinement works on research answers, not only plain chat.
• Translate depth matches a multilingual product bet.

What we would push on

• Overlapping ask vs explain actions force a coin flip.
• Search threads need source-check actions on selection, not only rewrite verbs.

Tradeoff

Takeaway

Extend selection menus for search threads with source-check actions, not only translate and explain.

What works

• Follow-ups keep the research session going without a blank bar.
• Armed search mode persists so the next turn stays in the same job.

What we would push on

• Follow-ups that repeat a shaky premise amplify the error.
• Source counts that do not open a list impress more than they inform.

Tradeoff

Takeaway

Follow-ups should deepen verification, not repeat the same shaky premise. Make source badges open the drawer.