A major live-streaming platform · AI highlight clips · 2026

AI cuts the highlights. Creators need to trust them, and not miss them.

The platform wanted AI to cut every stream into highlight clips. When AI gets it wrong, the creator looks bad. I led the flows and specs. With the PM, I decided when to tell creators, what to show them, and who presses publish.

Role

Product Designer (freelance): flows, wireframes, specs, decision docs

Team

PM, in-house designer (design review)

Duration

2026

700M+

Yearly visits (platform scale)

3

Handoff principles

A / B

Options handed to the PM

One card, two states: AI highlights in progress, and AI highlights ready
One card, two states. Right after the stream it says clips are in progress, then how many are ready. Screens are de-identified under NDA.

The brief

“After each stream, use AI to cut highlights for creators to publish.”

The earlier wireframe came back from review with seven comments. Before changing anything, I sorted them.

Do as asked

Confirm once more before publishing

Do it, and go one step further

Keep the notice on the profile, even with push off
Notify a day before expiry, and on expiry

Turn into an A/B choice

Not sure creators care which stream a clip came from
From one clip, see the others from the same stream

Ask first

Let creators hide the prompt once they know it
“Don't show again” (duplicate of the above)

Design decisions

Don't let AI clips get missed or published by mistake

Visible even with push off

Even with push off, the notification center and profile still show what AI made. Right after a stream, a notice says highlights are in progress, so nobody thinks the feature broke.

Left: the notification center. Right: a close-up of the AI clip notification toggle, and of the in-progress card shown after a stream
Left: notification center. Top right: settings, where AI clip alerts can be turned off on their own. Bottom right: the notice after a stream.

Drafts flag what's expiring. Publish and delete ask first.

Drafts flag clips expiring within three days, and publish opens a check of the AI's title, price and tags. AI can pick the wrong moment or title, so a person takes the last step.

Three close-ups: a draft card flagged as expiring soon, the publish confirmation, the delete confirmation
Left to right: my drafts, confirm publish, confirm delete.

The key step

I turned the open question into two options the PM could compare

Whether clips should show which stream they came from shaped the whole drafts page. I built both versions end to end and wrote down who each one suits.

How often does your creator stream?

Drag the slider. The same 11 clips switch to the layout that fits. Redrawn from the wireframes.

Now and thenAlmost daily
Version B: grouped by stream

One card per stream with days left and clip count; tap to expand. For daily streamers, it stays findable as clips pile up.

The trade-off I wrote for the PMA: expandedB: grouped by stream
FitsOccasional streamersFrequent, long-term streamers
UpsideLots of clips at a glance, feels rewardingStays organized as clips pile up, and shows the source
CostEndless scrolling as clips growOne more tap to learn
Pick it ifthe goal is “AI did so much for me”the goal is clips that are easy to find and manage

Status and reflection

The hard part of this AI feature was the handoff to people

Wireframes and decision docs are delivered. The requirements kept changing after that, but the handoff rules stayed: don't let clips be missed, don't let them go out by mistake, and always show their status. After launch I'd track draft opens, draft-to-publish rate, drop-off at the publish step, and how many creators turn notifications off.

If I did it again, I'd show both versions to a few creators and let their reactions shape the trade-off. This time I relied on reasoning alone.