Transcript
Merge Mode
English
Problem: we couldn't name it yet
People were slow in the editor, and nobody could say why. They hovered over the toolbar, flipped Play English / Play Japanese to work out what it did, and looked for speed in two places. None of it was a bug anyone could file.
Nobody asked for more features. Every session went the same way: open the editor, stare at it, work out what it wanted. The AI was doing its job. The interface wasn't showing anyone what it had done.
T]the Room _ Simon Sinek
Find & ReplaceMerge Cards
Change RateOptimize
Generate SubtitlesGenerate Audio
Add
Hire Professional
Export現実には、私はばかです、そして私はそれについてひっくり返っていません、例えば、私は理解していません
非常に複雑なことなので、私は多くの質問をするので、話している人は
私は何か複雑なこと、私がそれらを理解しているかのように言葉で話している...
神経科学、私はあなたが何について話しているのかわかりません、そして私は...
私はそれを理解しています。
The reality is,
very complicated things, and so I ask a lot of questions so that the person who’...
me something complicated, who’s talking in terms as if I understand them, fina...
neuroscience, I don’t know what you’re talking about, and I ask enough questio...
that I understand it.
00:00.000
00:07.480
00:07.480
00:13.260
00:13.260
00:13.260
00:18.380
00:22.100
00:22.100
00:22.940
Generate
17 Words
1 Word
Haruto
Haruto
Haruto
Haruto
Haruto
Audio Rate
1.01
Video Rate: 1.01x
Respeech
RecordUpload
1
Edit Audio
Dashboard
Teams
Process
Transcript
Subtitle
Elements
Audio
Text
Upload
SettingsPlay English
Play Japanese
00:01:42 / 00:08:23
The argument
Leadership wanted to ship more: AI tools, export options, a command palette. I played them recordings of users hesitating over the toolbar instead. We agreed to restructure alongside feature work, not instead of it.
Two dead ends
Fewer buttons. Cleaner, but “I still don't know where things are.” Hiding actions doesn't give anyone a mental model.
Step-by-step tabs. Tested well as a concept and failed in use: “But I need to go back and forth.” Localization isn't a line.
Find & Replace
Merge CardsChange RateOptimizeGenerate SubtitlesGenerate AudioTranscriptrailTranslatecardGenerate SubtitlestoolbarExporttop barThe problem wasn't how many actions there were, or their order. Every action was asking for attention at the same time.
Solution: Four tiers, four zones
I mapped every action on two axes: how often it's used, and how much deliberate intent it needs. Frequent actions stay on the editing surface. Infrequent ones sit in the top bar. Deliberate ones live in context panels, and actions that cut across the editor are a search away in the command menu.
The editor now has four zones: context panels, transcript cards, video preview and a live timeline. The active card follows the playhead, and languages are always visible.
The screens above are from my draft, made before the editor moved onto Aero, the design system I built at Vitra. The shipped editor is built from Aero’s tokens and components, so the colours, buttons and spacing the audit flagged come from one system instead of being drawn by hand.
The button nobody saw
In testing, people edited the source text, then couldn't find how to re-translate it. The Translate button sat between the two columns and looked like a divider. We moved it to the bottom of the card and made it activate only when the source had changed. In the second round, translation mistakes dropped noticeably.
reads as a divider
What changed
A core task went from about 15 minutes to about 5. I timed it by hand, so read it as “roughly a third of the time,” not a lab result.
When software does most of the work, the interface has a different job. It isn't where the work happens anymore. It's where people understand the work.