Edit Greek Reel

Step 2: Transcribe

1. Install openai-whisper if needed (pip3 install openai-whisper)
2. Transcribe with Whisper medium model, Greek language, word-level timestamps:

model = whisper.load_model("medium")
result = model.transcribe(audio_path, language="el", word_timestamps=True, condition_on_previous_text=True)

3. Save transcript to transcript.json in the same directory
4. Print the full transcript and word timestamps for review

Step 3: Proofread the Transcription

CRITICAL: Whisper makes mistakes, especially with:

• English tool/brand names (e.g., "Cloud Code" → "Claude Code", "CacheSource" → "Cursor")
• Greek spelling errors (e.g., "ευτοματά" → "αυτόματα", "φιτιτικού" → "φοιτητικού")
• Merged or split words

Review the transcript yourself and fix obvious errors. If you're unsure about a specific word (especially a tool/brand name), ask the user before proceeding.

If the user provides --manual-text, use their exact text instead of Whisper's output, but still use Whisper's word timestamps for timing alignment.

Step 4: Build Segments & Timed Words

Based on the silence detection and word timestamps:

1. Define KEEP_SEGMENTS — list of (start, end) tuples of audio to keep

   • Cut silence gaps > 0.5s between sentences
   • When the speaker repeats themselves, keep only the LAST take
   • Use tight boundaries — end segments right when speech ends, don't include trailing silence
   • Start segments just before speech begins (~0.05s padding)

2. Define TIMED_WORDS — list of (word, start, end) with the CORRECTED text mapped to Whisper timestamps

3. Recalculate all timestamps relative to the trimmed output

Step 5: Configure Effects

Subtitles (Karaoke Style)

• Font: Manrope Bold (search for Manrope-Bold.otf or Manrope-Bold.ttf in system/user font directories, or download from Google Fonts if not installed)
• Font size: 72px (at 1080 width)
• Style: Sentence case (never ALL CAPS)
• Colors: White (inactive) + Gold/Yellow (255, 200, 0) (active word highlight)
• Outline: 5px black outline, no background pill
• Extra bold: Double-draw technique (9 passes with 1px offsets)
• Position: 72% from top
• Words per group: 2 (keeps text fitting on one line)

Zoom Effects (Subtle)

• Maximum 5 zoom triggers per video
• Zoom factor: 1.08–1.10x (never more than 1.12x — avoid making viewer dizzy)
• Duration: 0.35–0.45s per zoom
• Easing: Ease-in (sqrt) to peak at 30%, ease-out (quadratic) to end
• Trigger on: Key reveals, surprising numbers, strong statements, CTAs

Sound Effects

• NEVER repeat the same SFX file twice in one video
• This skill ships with pre-trimmed SFX in its audios/ directory (relative to this skill.md file):
  • trimmed_whoosh.mp3 — transitions, reveals
  • trimmed_cash.mp3 — money/price mentions
  • trimmed_fah.mp3 — emphasis, strong statements
  • trimmed_click.mp3 — tool mentions
  • trimmed_bubble_pop.mp3 — light reveals
  • trimmed_riser.mp3 — builds, anticipation
• The skill's base directory is provided at invocation as Base directory for this skill: <path>. Use that path to locate the bundled audios/ folder.
• Also check the video's parent directory for an audios/ folder — the user may have added custom SFX there
• If new untrimmed audio files exist, trim leading silence first:

  ffmpeg -i input.mp3 -ss <silence_end> -acodec libmp3lame -q:a 2 trimmed_output.mp3
  

• Volume: 0.15–0.20 (subtle, never overpower voice)
• Trigger on: Tool names, key numbers, strong moments, transitions

Image Overlays

• Check images/ directory for available logos, screenshots, memes
• Display above the speaker's head area (centered, ~15% from top)
• Logo size: 200px max
• Meme/screenshot size: 500px max
• Animation: Pop-in (ease-out over first 15%) and pop-out (over last 15%)
• Duration: 1.8–2.5s per image
• Trigger on: When the speaker mentions the tool/concept the image represents
• Each image triggers only once
• Convert SVGs to PNG first if needed (use cairosvg)

Step 6: Video Processing

Crop (Object-Cover, Never Stretch)

• Target: 1080x1920 (9:16)
• If --crop-top N is specified, remove N% from the top before fitting
• Always crop to fit the target ratio (like CSS object-fit: cover), never scale-to-fit (which would stretch/distort)
• Center the crop horizontally; for vertical, bias toward bottom-center (keep the speaker's face)

Processing Pipeline (Python + ffmpeg + Pillow)

Pass 1: Trim + Crop + Scale (ffmpeg)

• Build a complex filter: trim each segment, concat, crop to 9:16, scale to 1080x1920
• Concat uses interleaved stream ordering: [v0][a0][v1][a1]...concat=n=N:v=1:a=1
• Output: temp_trimmed.mp4 (libx264, crf 18, aac 192k, 30fps)

Pass 2: Subtitles + Zoom + Images (Pillow frame-by-frame)

• Decode trimmed video to raw RGBA frames via ffmpeg pipe
• For each frame:
  1. Apply zoom effect if active (center-crop + resize)
  2. Composite image overlay if active (with pop animation)
  3. Composite subtitle overlay
• Encode back to mp4 via ffmpeg pipe

Pass 3: Mix SFX (ffmpeg)

• Overlay all SFX using adelay + amix filter
• Use normalize=0 to prevent volume pumping
• Copy video stream, re-encode audio only

Output

• Save as final_<name>.mp4 in the same directory as the input
• Print summary: original duration → final duration, number of effects applied
• Clean up temp files

Important Rules

1. Never stretch video — always crop to fit (object-cover behavior)
2. Proofread before burning subtitles — Whisper WILL get tool names wrong
3. Ask the user if unsure about a word, especially brand/tool names
4. Sentence case only — never ALL CAPS subtitles
5. No background pill behind subtitles — outline only
6. Unique SFX — never use the same sound file twice in one video
7. Subtle zooms — 1.08-1.10x max, 5 per video max
8. Tight cuts — trim silence aggressively, the reel should feel fast-paced
9. Cache transcript — if transcript.json exists, reuse it (skip re-transcription)
10. Keep the last take — when the speaker repeats, always keep the final version