Cut pauses, ums and restarts

When to use

• Someone hands over a talking video and wants it tighter: "cut the pauses", "take out the ums", "remove the bits where I start over", "make it flow".
• The video is one take or several takes in order (a phone recording, a Loom, an interview, a listing walkthrough).

Prerequisites

• The cuttingfloor MCP server is connected.
• The video is reachable: a URL for media.import {source: {url}}, or a file on the user's machine sent with media.upload then media.import {source: {upload_id}}.

Steps

1. The assistant should import every take with media.import, and wait with job.get {job_id, wait_seconds: 50} until each is ready. Several takes in order go on one timeline as they are; there is no need to join them first.
2. The assistant should transcribe each take with media.analyze {kind: "transcript"}. The words are right but their times are not exact: a word's end often stretches over the pause after it, and short "um"s and "uh"s are often missing from the text.
3. The assistant should measure where the speech really is with one code.run script over all takes (inputs: their media_ids). Read the words from cf.media.get({media_id, include: ["transcript"]})["transcript"]["words"] (the analyses.transcript entry is only a summary). FFmpeg is in the sandbox: run silencedetect (start at n=-32dB:d=0.12) on each take's audio, turn the silences into speech bursts, and print each take's words beside its bursts. Check the first burst against the first word: a soft opening ("okay, so") can sit under the threshold, and a threshold near the room noise counts breaths as speech, so adjust per take. Cut on burst edges, never on transcript times.
4. The assistant should have Jev mark the filler in the same code.run script, with cf.judge.ask (it uses cuttingfloor's own key; no key goes in the script). For each take, ask two yes/no questions per word, with the word marked [[like this]] in the take's text: "Is the marked word the first word of filler?" and "Is it the last word of filler?", where filler is "words the speaker can leave out so that what is left says exactly the same thing and still reads as a whole sentence: hesitations, verbal tics, a repeated word, the false start of a phrase they then say again, and throwaway asides", with a few short examples of each. Pair firsts (0.3 or more) with lasts up to 12 words on, ask "Are the marked words filler?" of each pair with the sentence as it would read without them, and keep the best spans first (0.6 or more), never overlapping. Up to 300 questions go in one call; a 10-second take needs about 7 calls in all and answers in under a second.
5. The assistant should review every span Jev marks before cutting; Jev finds the problems, the assistant decides the exact cut:
   • read the sentence without the span: if it lost meaning (a restart span that swallows the word the speaker kept, like "nobody"), choose the cut that keeps the clean last try instead;
   • move each cut to the nearest speech-burst edge, so no cut lands inside a continuous phrase (cut a whole repeated line, not "line. Just");
   • add the pauses from step 3 that are longer than about 0.6 s, trimmed to about 0.15 s each side of the speech; Jev reads text and does not see pauses;
   • keep an opening "okay, so" when it is the speaker's voice; cut it when the next words make a stronger start.
6. The assistant should build the cut on one timeline: timeline.create at the takes' size and frame rate, clip.add each take in order, then one edit.batch of range.extract ranges, latest first, so earlier frame numbers do not move. Or, often simpler: one clip.add per kept range with source: {in, out} in frames.
7. The assistant should smooth every cut: transition.add {kind: "crossfade", duration_frames: 4, mode: "handles"} on the clip before each cut (the video clip or its linked audio clip; either way only the sound crossfades and the picture cuts hard, and changed lists each crossfade added), and clip.set_transform {scale: 1.1} on every other section so a cut between two moments of the same shot reads as a cut-in rather than a jump.
8. The assistant should even out the sound if it needs it: audio.set_master {gain_db, limit_db: -1}, with the gain worked out from the loudness qa.check measures.
9. The assistant should look at its work before handing it over: render.frames on both sides of every cut, and qa.check.
10. The assistant should export with export.video (social_9x16_h264 for vertical, web_1080p_h264 otherwise) and hand back the file, saying how much was cut and that the order and every intended line are kept.

Decision table

Required behavior

• Never change the order of what was said, and never cut a line the speaker meant to keep.
• Cut on measured speech bursts, not transcript times.
• Say plainly what was cut: how many pauses, restarts and repeats, and the running time before and after.