Blogger Video Slicer

Use this skill for local video files only. Do not download from links; ask the user to provide a local file path first.

The goal is content-first slicing: one exported clip should contain one clear content point. Aim for 15-20 seconds, but preserve a complete thought when the duration target conflicts with meaning.

Required Tools

• Run the bundled script: scripts/slice_blogger_video.py.
• Require ffmpeg/ffprobe on PATH, or install Python package imageio-ffmpeg.
• Require one local speech-to-text package:
• Prefer faster-whisper.
• Fall back to openai-whisper.
• If dependencies are missing, stop cleanly and tell the user what to install. Never invent transcripts or output empty clips.

Command

Run from the skill folder:

python scripts/slice_blogger_video.py --input "video.mp4" --output "output"

Useful options:

python scripts/slice_blogger_video.py --input "video.mp4" --output "output" --min-duration 15 --max-duration 20 --language zh --review-only --keep-temp --flat-output

Default Behavior

• Create the next sequential output folder: output/no.1, output/no.2, output/no.3, etc.
• Never overwrite earlier video results unless the user explicitly says to replace or overwrite.
• Export only the core clip by default. Do not add intro/outro branding videos unless the user explicitly asks for branding.
• Generate three cover candidates per clip with short filenames: clip_001-1.jpg, clip_001-2.jpg, clip_001-3.jpg.
• Prefer cover frames showing clear emotion or action: cute expressions, smiles, effort, hardship, high-difficulty movement, surprising reactions, or other click-worthy moments.

Workflow

1. Confirm the input path is a local video file.
2. Extract audio.
3. Transcribe audio with local speech recognition and timestamps.
4. Build candidate clips from transcript segments.
5. Prefer cuts at punctuation, pauses, sentence endings, and natural transition phrases.
6. Trim irrelevant edges:
   - Remove greetings, throat-clearing, filler, or setup that does not match the middle topic.
   - Remove topic changes, calls to action, repeated filler, or the start of the next point at the end.
   - Preserve material needed to understand the main content.
7. Export final clips unless --review-only is used.
8. Generate per-second preview frames, storyboard images, cover images, and reports.
9. After running, verify counts for clips, previews, storyboards, covers, report.csv, and report.md.

Output Structure

output/
  no.1/
    clips/
      clip_001.mp4
    previews/
      clip_001/
        second_000.jpg
        second_001.jpg
    storyboards/
      clip_001.jpg
    covers/
      clip_001-1.jpg
      clip_001-2.jpg
      clip_001-3.jpg
    report.csv
    report.md

report.csv columns:

• clip_id
• start_time
• end_time
• duration
• topic
• reason
• trimmed_intro
• trimmed_outro
• clip_path
• storyboard_path
• cover_1_path
• cover_2_path
• cover_3_path

Quality Bar

• Split or reject clips that cover two unrelated points.
• Trim openings and endings that do not match the clip's central topic.
• Accept clips shorter than 15 seconds or longer than 20 seconds only when needed for a complete thought.
• Use second-level previews and stitched storyboards to review whether the first and last seconds still match the middle content.
• Report why each clip was selected and what intro/outro content was removed.