Roughcut

Eight clips, in shooting order

EP003 came in as eight clips from one afternoon, 41 minutes 48 seconds of footage. probe reads each file’s frame rate, length and start timecode, and puts the clips in shooting order.

That order matters more than anything else this stage does. The rule for picking a take is “the last clean one wins”, which only works if the tool knows which clip was recorded later. So the order is written into sources.json as an explicit clip_index. I can edit that file when a pickup lands in the wrong place, and every later stage reads it exactly as written.

Every word, with a time

transcribe runs Whisper over each clip and keeps every word with its start, its end and how sure Whisper was: 3,142 words for this episode.

It’s the only stage that needs the GPU, and it runs once. Everything after it reads the cached transcript, so changing how lines are matched or cut is a re-run of a few seconds, never a re-transcribe. The highlighted half-minute is one line of the script being read again and again.

Every place a line was read

align matches the script against the transcript. Aligning the whole script in one pass would find each line once, usually at its first read, which is the one that went wrong. So the tool first pins the script to the transcript with phrases that only occur once, then runs a local alignment (Smith-Waterman) inside each pinned stretch to find every separate read of every line.

Each read is scored on coverage, accuracy, Whisper’s confidence, stumbles and speaking rate, then sorted into clean, usable or flubbed. This line was read eight times; the four the tool kept are a stop halfway and three clean reads. The last clean read wins.

The channel is a client’s, so the line and its reads in the panel are stand-ins of the same length. The takes, their timings and their scores are the episode’s own.

One path through all of them

Choosing each line’s take on its own goes wrong in two ways: neighbouring lines can claim the same words, and a cut can jump back in time. So the tool chooses every line’s take at once, with a Viterbi pass over the whole script. It penalises cuts that run backwards or overlap, and rewards takes that sit next to each other in one clip, so a paragraph read cleanly in one go stays one clip instead of eight jump cuts.

The last-clean-take rule stays the default. The pass only overrules it to fix the order, and flags every line where it did. Switch the panel between the two: on this episode they differ on 15 lines. Line by line, one cut runs backwards; the whole-script path has none, and fewer cuts.

Cutting in the silence

Whisper’s word timings are inferred, not measured, and they’re often wrong by more than the pause between two words. Its word ends tend to fall inside the word, which clips the last consonant. So align measures the audio instead: one decode per clip gives a loudness curve in 10 ms steps, and each cut snaps to the nearest real silence on the safe side of the word.

In the panel, Whisper ends the line’s last word at −18 dB, while it’s still sounding. The measured cut lands 300 ms later, at −80 dB. Across 193 cuts in one clip of an earlier episode, measured out-points landed at −78.9 dB instead of −45.3, and not one landed inside speech.

What it couldn’t decide

Nothing the tool decides is hidden. proof renders the cut list straight to sound and picture with ffmpeg, so I can listen before Resolve is involved, and it cross-checks the cut list’s seconds against its frames. The report lists every line the tool wasn’t sure of, each with every take it found, so I check its choice instead of redoing it.

For EP003 that was 61 of 147 lines. Most are small: 37 cuts with no silence to land in, and lines whose takes scored too close to call. Two lines were never found in the footage, and 24 stretches of speech matched no line of the script, 3 minutes 29 in all. Each one is listed with a command that plays it.

The rough cut

plan turns the cut list into the exact calls Resolve’s scripting API needs, without calling Resolve itself, so the plan can be read before anything is built. Claude then builds the timeline in Resolve from that plan, following a written procedure.

EP003’s cut list is 157 clips, 18 minutes 48 seconds out of 41:48. In Resolve, the 60 alternate reads went onto their clips as native takes, so switching a read is a click and its sound comes with it. The notes became 69 markers on the clips themselves, so they move when I cut. From there the edit is mine: pacing, graphics, the grade and the sound.