find-soundbites

I'd like you to add a skill to Wideframe called "find-soundbites".

What it should do:
Search the transcripts for your topics and return timecoded selects, flagging any topic with no match.

Before you run anything, ask me to confirm:
- The exact list of topics, quotes, or keywords to find.
- How many selects per topic, and any min or max clip length?
- Match spoken words only, or also on-screen and visual content?
- Output as a document, markers, or a selects list?

Then do the work:
1. Ensure every candidate clip has a transcript, running media-transcript-get-batch on any that are missing.
2. For visual matches, run media-analyze-content-batch.
3. Score each segment against each topic and pick the best within the length bounds.
4. Return per-topic selects with the source file and in and out timecodes.
5. Flag any topic with no qualifying match.

Do not grade your own work. When the output is ready, hand it to a separate verification subagent that has its own context window. Give it the output, the choices I confirmed, and the pass criteria below. It reviews the quality, writes specific feedback, and returns PASS or FAIL.

Pass criteria:
- Every returned select contains the topic, checkable against its transcript line.
- Each select has a concrete source file and in/out timecode within the clip duration.
- Every requested topic has at least one select or an explicit no-match flag.
- The transcript search covered every candidate clip.

If it returns FAIL, apply the feedback, redo the affected work, and verify again with a new subagent. Repeat until it returns PASS, then show me the result.

Once it works, save it as a reusable skill named "find-soundbites" so I can run it again by saying /find-soundbites.