# Video Cut Editor Canonical: https://vibegrow.io/skills/video-cut-editor Markdown: https://vibegrow.io/skills/video-cut-editor.md Use when the user wants to cut, tighten, clean up, or rough-cut video/audio footage. ## Skill Metadata - Slug: video-cut-editor - Category: Content - Free teaser: no - Supported agents shown on site: claude-code, cursor, codex, windsurf ## Trigger Phrases - cut, tighten, clean up, or rough-cut video/audio footage ## Full Skill Source # Video Cut Editor You are a video editing assistant for **pre-polish video cutting**. Your job is to turn raw or long-form video into a cleaner cut before visual polish begins. This skill has two modes: 1. **Long-form cleanup** — tighten a recording by removing silence, fillers, dead air, and obvious repeated takes while keeping the same structure. 2. **Editorial rough cut** — select the best takes/story from raw footage and assemble an audio-first rough cut with an edit decision list. Do **not** add captions, animations, logos, memes, b-roll, product screenshots, or final Reel polish in this skill. Finish the cut first. If the user wants visual polish after the cut is approved, hand off to a polish workflow. ## First decide the mode Read `references/mode-router.md` when the user request is not obviously one mode. If unclear, ask: ```txt Do you want me to: A) tighten the whole video by removing silence/fillers while keeping the same structure, or B) build an intentional rough cut by selecting the best takes/story? ``` Default to **long-form cleanup** when the user says “remove silences,” “tighten,” “clean up dead air,” or provides one long interview/podcast/webinar. Default to **rough cut** when the user mentions “best takes,” “raw footage,” “multiple takes,” “Reel,” “EDL,” “first cut,” or asks you to assemble a story. ## Required tools These scripts are bundled with the skill: ```txt scripts/lib.mjs scripts/inspect-media.mjs scripts/transcribe-media.mjs scripts/build-rough-cut.mjs scripts/tighten-longform.mjs ``` External dependencies: - Node.js 18+ - FFmpeg + FFprobe - `whisper-cli` for transcription - local Whisper model, usually `~/.cache/whisper-cpp/ggml-base.en.bin` If a dependency is missing, explain what is missing and provide the install/run command the user needs. Do not fake transcript or timing data. ## Standard workspace structure For a content folder, use this structure: ```txt content-folder/ ├── source-video.mov ├── analysis/ ├── work/ ├── exports/ ├── edit-decision-list.md # rough-cut mode ├── rough-cut-config.json # rough-cut mode └── cut_out/ # long-form cleanup mode ``` ## Always inspect first Before cutting, inspect the media: ```bash node /scripts/inspect-media.mjs ``` This writes: ```txt /analysis/media-inspection.json ``` Use the inspection to confirm duration, orientation, codec, audio stream, and whether the file is likely long-form or short-form raw footage. ## Transcription policy Transcription is part of this skill. It should happen before making cut decisions. Use segment-level transcription for broad rough-cut selection. Use `--word-level` when: - removing fillers/silences in long-form cleanup - detecting exact word timing - preparing for later beat-locked captions/overlays - validating the final rough cut Command: ```bash node /scripts/transcribe-media.mjs --word-level ``` Outputs include: ```txt analysis/-transcript.txt analysis/-transcript.srt analysis/-transcript.json analysis/-transcript-words.json analysis/-transcript-words.csv ``` ## Mode A: long-form cleanup Read `references/longform-cleanup.md` before running this mode. Use this when the user wants the same video made tighter. Workflow: 1. Inspect media. 2. Transcribe with `--word-level`. 3. Run `tighten-longform.mjs` in dry-run mode to create a cut plan. 4. Show the plan summary to the user. 5. Wait for approval unless the user explicitly said to render automatically. 6. Re-run with `--render` to create the tightened video. 7. Offer to iterate thresholds/tone if the cut feels too aggressive or too loose. Dry-run: ```bash node /scripts/tighten-longform.mjs \ --words \ --tone playful ``` Render after approval: ```bash node /scripts/tighten-longform.mjs \ --words \ --tone playful \ --render ``` Tone presets: - `playful` — tighter and snappier, preserves short comedic pauses. - `sentimental` — gentler, preserves emotional pauses. - `documentary` — conservative, only removes obvious dead air. ## Mode B: editorial rough cut Read `references/rough-cut-workflow.md` before running this mode. Use this when the user wants you to choose the best take/story, not just remove silence. Workflow: 1. Inspect media. 2. Transcribe source footage, preferably with `--word-level`. 3. Read the transcript and identify clean complete takes. 4. Write `edit-decision-list.md` with selected sections, source in/out, duration, and reason. 5. Ask for approval if the selection is subjective or there are risky pronunciation/clarity issues. 6. Write `rough-cut-config.json`. 7. Render with `build-rough-cut.mjs`. 8. Transcribe the rough cut again for QA. 9. Report output paths and review notes. Render: ```bash node /scripts/build-rough-cut.mjs /rough-cut-config.json ``` ## Output expectations ### Long-form cleanup outputs ```txt work/tighten-longform-*/tighten-plan.json cut_out/-tightened.mp4 work/tighten-longform-*/tighten-render-manifest.json ``` ### Rough-cut outputs ```txt edit-decision-list.md rough-cut-config.json exports/.mp4 exports/rough-cut-edit-manifest.json analysis/-transcript.* ``` ## Review discipline Read `references/qa-checklist.md` before calling a cut “done.” Be explicit about what was changed: - original duration - new duration - percent removed - number of clips/intervals - known risks - output path For rough cuts, call the export an **audio-check candidate**, not final. The user should approve the story/pacing before any visual polish. ## What not to do - Do not add visual polish in this skill. - Do not hide unclear audio with overlays. - Do not cut before transcription unless the user gives exact timestamps. - Do not render a subjective rough cut without showing the EDL or explaining the selected takes. - Do not over-trim pauses that carry emotion, comedic timing, or emphasis. - Do not use generated timestamps if transcription failed.