/video-repurposing

name: video-repurposing
description: Turn long-form video into platform-ready clips - LinkedIn posts, TikTok/Reels, teasers, highlight reels - with word-level captions, using FFmpeg and Remotion. USE THIS SKILL WHEN the user wants to repurpose a YouTube video or any long video into social clips, mentions video, clip, footage, mp4, mov, or any video format, wants to cut, stitch, combine, or edit video files, needs captions, subtitles, or TikTok-style word highlighting, wants a teaser, trailer, or highlight reel, asks about video transcription or whisper, mentions FFmpeg, ffprobe, or Remotion errors, or has .mp4/.captions.json files or video folders. Runs a mandatory intake, proposes an edit plan with exact timestamps, previews in Remotion Studio, and renders only on approval.
disable-model-invocation: true

Video Repurposing

Turns long-form video into platform-ready clips with captions. FFmpeg for fast CLI operations, Remotion for styled compositions you can preview before rendering. Battle-tested on a real LinkedIn publishing pipeline; the rules below exist because breaking them produced visibly worse clips.

---

Rules (Always Enforce)

**Opening Clip (CRITICAL):**

• The first clip is the hook - it must grab attention immediately
• Must start at the start of a complete sentence or idea with an explicit subject
• NEVER start with joining words ("But", "And", "So", "Now", "Then", "However")
• NEVER start with pronouns without context ("It", "This", "That", "They", "These", "There")
• NEVER start with filler, hedging, or weak openers ("I think", "You know", "Basically")
• Review the full transcript and deliberately select the strongest possible opening - a hook, an insight, a bold statement, or a compelling question

**Closing Clip (CRITICAL):**

• Must end at the natural conclusion of a thought - never trailing off, never a hard stop mid-idea
• Should leave the viewer with a takeaway, a call-to-action, or a sense of closure
• If the last word sounds like it leads into something else, keep going until it resolves
• Good endings: a summary statement, a strong opinion, a forward-looking insight
• Bad endings: trailing off, abrupt cuts, filler, unrelated tangents

**Clean Cut Points (every cut, no exceptions):**

• Cuts happen at a natural pause in the content - never mid-sentence, never mid-word
• Respect the idea, not just the words: a cut is clean when the speaker just finished a complete thought and there is a natural breath after the last word
• Every clip must be a complete, standalone thought with an explicit subject
• The stitched video must tell a coherent story, not feel like random fragments

**Editing modes - pick ONE in intake:**

1. **Hard-cut mode (DEFAULT for social clips: LinkedIn, TikTok, Reels).** No transitions. Ever. No fades on video or audio, no crossfades, no dissolves. Clip ends on one frame, the next begins on the very next frame. Audio starts and stops at full volume. Clean hard cuts read as confident and native on social feeds; fades read as template-ware. Build with plain `<Sequence>` blocks + `premountFor={60}` (see Composition Rules below). **Do not use `TransitionSeries`** - it has a null-props bug that crashes Studio in current Remotion versions. 2. **Crossfade mode (optional, for trailers/teasers/montages when the user explicitly wants blended transitions).** NEVER allow overlapping audio noise: cosine curves only (audio fade-out 1.5s, fade-in 1.2s, visual crossfade 1s), no fade-in on the first clip, no fade-out on the last. See [references/transitions.md](references/transitions.md) for the `createVolumeFn` pattern and the TransitionSeries bug workaround.

**Transcription-first timestamps (CRITICAL - the single source of truth):**

• All clip boundaries come from whisper word-level timestamps, never from scrubbing or guessing
• `startMs`/`endMs` = whisper timestamps of the first and last spoken words → these define the captions
• Audio/video cut points DERIVE from those: `audioEndMs = endMs + audioBufferMs` per clip. Whisper timestamps run ~200-700ms ahead of actual speech, so without the buffer the last word gets clipped
• Punctuation tokens (".", ",") have timestamps but are NOT spoken - never use them as `endMs`
• Captions are filtered by the raw transcript boundaries, NOT the buffered audio end, so words from the next sentence never flash on screen during the buffer zone
• Always verify by ear in Studio: the last word resolves fully, the next sentence doesn't bleed in
• Full pattern with code in [references/transcription.md](references/transcription.md)

**Source video prep (before anything else):**

• Check the codec with ffprobe. AV1 is too slow to decode in Remotion Studio (YouTube often serves AV1 by default) - always transcode to H.264
• `-movflags +faststart` is mandatory: without it the moov atom sits at the end of the file and Remotion reads the whole file before it can seek, making Studio crawl. Already-H.264 files remux losslessly: `ffmpeg -i in.mp4 -c copy -movflags +faststart out.mp4`
• Videos go DIRECTLY into `public/` with no spaces in filenames

**Workflow:**

• ALWAYS complete Phase 2 (Intake) before making any edits - never skip clarification questions
• ALWAYS propose an edit plan with exact timestamps and get user approval before executing
• ALWAYS preview in Remotion Studio before rendering
• ALWAYS run QA checks before showing the preview (see [references/qa-testing.md](references/qa-testing.md))
• NEVER render automatically - only when the user explicitly approves

---

Brand kit (one-time, per user)

All styled output (captions, title cards, graphics) uses the user's brand, captured once in intake and reused forever:

| Setting | What to ask | Default if they have no preference | |---|---|---| | Colors | brand dark, background/light, 1-2 accents | dark `#0A0A14`, light `#F5F2EA`, accent `#B8E0B0` | | Fonts | heading font (Google Fonts name) | Inter (via `@remotion/google-fonts`) | | Caption style | frosted pill (default)