video-post-production（视频后期制作）

Video Post-Production

Use this skill when the user gives you one raw video and wants a finished short video with:

• readable subtitles
• highlighted key phrases
• basic SFX emphasis
• optional BGM
• a rendered export

The default workflow is single-input automation: the user provides the video, and you handle the rest.

Inputs

Required:

• one raw video file with spoken audio

Optional:

• preferred subtitle color/style
• preferred BGM mood
• preferred SFX intensity
• existing BGM file
• existing SFX directory

If optional inputs are missing, proceed with defaults.

Outputs

Create a working directory next to the input video:

• <video_name>_output/

Expected output files:

• alignment.json
• production_plan.json
• subtitles.ass
• final.mp4

Default behavior

If the user does not specify a style:

• subtitles: bold, high-contrast, mobile-readable
• highlights: only on high-value phrases
• SFX: light to medium density
• BGM: optional, low-volume, ducked under speech

Do not block on missing BGM or SFX. A valid delivery can still be:

• subtitles only
• subtitles + generated simple SFX
• subtitles + BGM

Workflow

1. Check prerequisites

Confirm these are available:

• ffmpeg
• ffprobe
• python3
• faster-whisper if transcription is needed
• a Chinese-capable font for ASS rendering

Use:

ffmpeg -version | head -1
ffprobe -version | head -1
python3 -c "from faster_whisper import WhisperModel; print('faster_whisper OK')" 2>/dev/null || echo "Need: pip3 install faster-whisper"

If faster-whisper is missing:

pip3 install faster-whisper

2. Transcribe the video

Run:

python3 <skill-path>/scripts/align_speech.py \
  --video "<input_video>" \
  --output "<workdir>/alignment.json" \
  --model "medium" \
  --language "zh"

This produces word-level timing and segment-level timing.

3. Build a post-production plan

Read alignment.json and produce production_plan.json.

The plan must contain:

• subtitle_groups
• sfx
• bgm

Minimum shape:

{
  "subtitle_groups": [],
  "sfx": [],
  "bgm": {
    "mood": "inspirational",
    "tempo": "medium",
    "volume": -18
  }
}

Planning rules

Subtitle grouping:

• group into natural spoken phrases
• do not split a phrase awkwardly
• prefer 3-8 Chinese characters per phrase
• prefer 0.5s or longer readable display windows
• keep within mobile reading comfort

Keyword highlighting:

• highlight emotional, contrastive, numeric, warning, CTA, and payoff phrases
• do not over-highlight
• default target: 1-3 highlighted phrases per 10 seconds

SFX placement:

• opening transitions
• key claims
• warnings
• CTA moments
• major topic pivots

Do not place SFX on every subtitle. Keep it sparse enough to feel intentional.

BGM choice:

• educational / explainer -> uplifting / neutral
• warning / risk -> tense / restrained
• positive / benefit-driven -> hopeful / energetic

4. Generate ASS subtitles

Run:

python3 <skill-path>/scripts/generate_subtitles.py \
  --plan "<workdir>/production_plan.json" \
  --output "<workdir>/subtitles.ass" \
  --video-width 720 \
  --video-height 1280 \
  --font "Heiti SC" \
  --font-size 48

If you need deeper ASS styling guidance, read:

• <skill-path>/references/ass_effects.md

5. Prepare audio

BGM

Preferred order:

1. user-provided BGM
2. existing local BGM in workspace
3. proceed without BGM

If BGM exists, prepare it:

ffmpeg -stream_loop -1 -i "<bgm_file>" -t <video_duration> \
  -af "volume=-18dB,afade=t=in:d=2,afade=t=out:st=<duration-3>:d=3" \
  -y "<workdir>/bgm_prepared.wav"

SFX

Preferred order:

1. user-provided SFX directory
2. existing local SFX directory
3. generate simple fallback tones/noise

If no SFX files are available, read:

• <skill-path>/references/audio_resources.md

6. Render final video

Run:

python3 <skill-path>/scripts/render_video.py \
  --input "<input_video>" \
  --subtitles "<workdir>/subtitles.ass" \
  --plan "<workdir>/production_plan.json" \
  --bgm "<workdir>/bgm_prepared.wav" \
  --sfx-dir "<workdir>/sfx" \
  --output "<workdir>/final.mp4" \
  --resolution "720x1280"

If no BGM exists, omit --bgm. If no SFX directory exists, omit --sfx-dir.

Quality bar

The result should feel like a polished short-form knowledge video:

• subtitles are easy to read on mobile
• highlighted words are meaningful, not noisy
• SFX emphasize real rhetorical beats
• BGM stays below speech and never masks intelligibility
• final export is usable without extra manual cleanup

Fallback rules

If the raw video is usable but supporting resources are missing:

• still deliver subtitles and a rendered video
• note what was skipped
• do not block the whole workflow waiting for perfect assets

If transcription quality is poor:

• keep subtitle grouping conservative
• prefer fewer highlights
• avoid aggressive SFX timing

Bundled resources

Scripts:

• scripts/align_speech.py
• scripts/generate_subtitles.py
• scripts/render_video.py

References:

• references/ass_effects.md
• references/audio_resources.md

Final delivery checklist

Before finishing, ensure:

• alignment.json exists
• production_plan.json exists
• subtitles.ass exists
• final video renders successfully
• subtitles are visible
• speech remains clear
• SFX and BGM are not overpowering