---
name: video-cut-assistant
description: "Produce edit decision lists, trim plans, and subtitle cues from raw footage transcripts. Ingests a transcript, identifies the beats, writes cut instructions for DaVinci / Premiere, and outputs SRT captions."
license: MIT
---

# Video Cut Assistant

Video Cut Assistant turns a raw footage transcript into an editor-ready plan: it reads the timestamped transcript, finds the natural beats of the conversation, and emits cut instructions plus burned-in-ready subtitles. It's built for solo editors, YouTubers, and podcast-to-clip teams who shoot long-form takes and want a first-pass rough cut decision before they ever open the timeline.

## When to use it

- You recorded a 40-minute talking-head take and need to strip filler, false starts, and dead air before editing.
- You have a podcast or interview transcript and want short-form clip candidates pulled by topic beat.
- You need an SRT caption file aligned to a cut you're about to make, not the raw source.
- You want a paper edit you can hand to a collaborator working in DaVinci Resolve or Premiere Pro.

## How it works

Feed it a transcript with timecodes (WebVTT, SRT, or a plain `[hh:mm:ss] speaker: text` log work best — Whisper/`whisperx` JSON output is ideal because word-level timings survive). The skill then:

1. Segments the transcript into beats — topic shifts, Q/A boundaries, and natural pauses.
2. Marks removable spans: filler words, repeated takes, long silences.
3. Produces a trim plan as an ordered list of keep/cut ranges with source in/out timecodes.
4. Exports two artifacts: an **EDL** (CMX3600 `.edl`) you can import into Resolve or Premiere, and an **SRT** caption file re-timed to the cut sequence.

Because it reasons over text, accuracy depends entirely on the timecode fidelity of your transcript. It does not touch your media files — it only writes the plan and captions.

## Usage

```
curl -fsSL --create-dirs -o ~/.claude/skills/video-cut-assistant/SKILL.md https://claudeskil.com/skills/video-cut-assistant/SKILL.md

# point it at a transcript with timecodes
"Cut my interview.vtt down to the 3 strongest topic beats,
 remove filler and pauses over 1.5s, and give me an EDL + SRT."
```

Output you'll get back: `interview.edl` (importable cut list), `interview.srt` (re-timed captions), and a human-readable beat summary you can sanity-check before importing.

## Tips & gotchas

- **Word-level timestamps matter.** Sentence-only timing produces loose cut points; run `whisperx` (not vanilla Whisper) for tight word alignment.
- EDLs reference source timecode, so your transcript timecodes must match the clip's timeline timecode — offset a transcript that started at `00:00:00` against media that starts at `01:00:00:00` and every cut will be wrong.
- The trim plan is a *paper edit*, not a render. Always review beats before importing; the model can misjudge where a thought actually ends.
- For Premiere, import the `.edl` rather than pasting cuts manually — frame rate mismatches silently drift otherwise.

## FAQ

**Does it edit my video file?** No. It generates the cut list and captions; you apply them in your NLE.

**What if my transcript has no timecodes?** It can still find beats and draft captions, but cut ranges won't be frame-accurate — supply timed input for usable EDLs.

**Resolve or Premiere?** Both. The CMX3600 EDL imports into either; SRT works anywhere.