Automating a two-track class edit from a transcript.

I was editing a recorded class in Premiere Pro. The speaker and the slides were recorded on separate, synchronized video tracks, with the slides covering the speaker. I needed to choose which one appeared, then remove the discussion sections and the short requests to advance a slide.

I had already cut the first 10 minutes by hand. I noticed the speaker often said "so" when he started reading a slide, and asked Claude to find those moments in the transcript.

Claude compared its cue sheet with the decisions already saved in my project. 19 of the 21 show/hide decisions matched. The two differences helped explain my choices: I opened on the speaker, and I kept the speaker on screen through a brief slide title instead of flashing the slide up for a few seconds.

We used those cuts to guide the rest: hold a slide for at least 6 seconds and the speaker for at least 8, with a shorter allowance during a slide request. Those were starting rules for this recording. I still needed to watch the transitions.

We considered exporting the sequence, changing it elsewhere, and importing it again. Reading the project exposed audio processing and proxy links that the proposed rebuild wouldn't preserve. We chose an in-place script to protect the already-finished audio work.

I ran a dry run first, reviewed its report, then ran the script in a duplicate project. It cut the slide track and disabled the sections where the speaker should be visible. Reading the saved file confirmed all 61 segments in the planned region matched the cue sheet. The earlier hand cuts were unchanged.

The middle discussion was just over 16 minutes. Together with 18 short asides, that made 19 removal ranges, about 17 minutes in total. Keeping the gaps open until they were checked meant the original timestamps stayed useful throughout the pass.

Watching the cut, I heard a slide request that should have been removed. The transcript had written it as "Fine." A search for "slide" couldn't find it.

We used the saved removal map to work backward from 27:36 in the edited video to the original transcript. Claude then looked for short, isolated utterances followed by a pause and a return to slide reading. That found a second request with the same mistranscription.

The keyword search and the new structural check together found 20 requests in this recording. The structural check reported no false positives in this pass. That result came from one class. It still needed checking on another.

I removed the two late requests by hand. The saved project confirmed their gaps matched across all four tracks. The removal map made the missed moment easier to locate; listening was what found it.

After the first Close Gap, the slides and both audio tracks ended at 34:13. The speaker track included a subscribe outro and ended at 34:28. All four tracks had zero gaps, their clip edges stayed on frame boundaries, and the title card at the beginning hadn't moved.

The amount of content was unchanged across that gap-close operation. It had removed the empty spaces left by the delete pass.

The two late corrections were then verified as matching gaps, totaling about 7 seconds. Closing those gaps was predicted to leave teaching through about 34:06 and the outro through 34:21. The session record ends before that last gap-close was verified.

The timing rules and the request detector came from this class. Another recording needs its own review before those rules can be trusted there.

We also didn't test running the removal script twice on the same project. The dry run and the saved-file checks let us inspect each planned pass; they don't establish that repeating it is safe.

I still have to watch the edit. A file can match the cue sheet exactly and contain a cut I'd want to change.