The Gap
“Video is visual content, but finding a moment in it means reading transcripts and scrubbing timelines. The thing that makes video easy to watch is exactly what makes it hard to navigate.”
Long-form video has no spatial address. A document has paragraphs, headings, a table of contents. You can navigate it by structure.
Video only has time.
Prompted Find makes the timeline navigable by meaning. You ask what you want to find.
The interface segments the video semantically, highlights relevant moments, and jumps you there.
The query returns a location, not a transcript, not a summary. A timestamp you can trust.
Decisions
Each decision was a choice about what kind of interface this is: tool, search engine, or navigation system.
01
The answer is a timestamp, not text
“Search as navigation.”
The query moves you inside the video. The interface doesn't summarize or transcribe. It relocates you to the moment where the answer lives.
02
Segments, not a search results list
“The timeline is the interface.”
The video timeline is divided into labeled semantic segments. Relevant ones are highlighted. You see the shape of the answer before you click it.
03
The query stays visible while you watch
“Context persists.”
The question panel stays open alongside the video. You can ask follow-ups without losing your place. The interface is a co-pilot, not a one-shot lookup.
Why this pattern matters beyond video. The same logic applies to any long-form content without navigable structure: podcasts, recorded meetings, legal documents, research papers. Prompted Find is a prototype of a pattern: ask a question, get a location. The medium changes. The interaction stays the same.
The Interaction
The prototype lives inside a YouTube watch page. The interaction has three steps, and none of them take the user out of the video.
What I'd Change
- Confidence indicators. Not all matches are equally strong. The timeline should visually encode match confidence so the user can prioritize where to jump.
- Multi-video support. The current prototype works with a single video. The real value is cross-video search: "find the part about heat shields" across an entire lecture series.
- Real embeddings, not keyword matching. The prototype uses concept aliasing and keyword scoring. A production version would use semantic embeddings to handle queries the alias table doesn't cover.