EyalZameret/Dudu-Director-Video-kit

Dudu: learning from reference videos

Claude cannot watch a link directly: opening a YouTube URL only shows the title and description. This skill gives Dudu "eyes" with yt-dlp and ffmpeg.

1. Extract

python tools/study_video.py "<url>"            # --threshold 0.2 for soft cuts, 0.4 for busy action

It writes to library/study/<video-id>/:

• meta.json: title, channel, duration
• shots.json: every shot's start, end, length, plus statistics (cuts per 10 s, median shot)
• shots-N.png: one labeled frame per shot, 12 per sheet
• transcript.txt: captions, if any The downloaded video is deleted at the end.

If the download fails with HTTP 403 or "page needs to be reloaded", update yt-dlp (pip install --user -U yt-dlp) and make sure Node is on PATH: the tool passes --js-runtimes node --remote-components ejs:github to solve YouTube's JS challenge.

2. Look (picture)

Read every shots-N.png and shots.json. Study five crafts:

• Directing: beats, where the story slows down and speeds up, what each sequence is for.
• Camera: per shot, size (EWS/WS/MS/MCU/CU/ECU/insert/POV), angle, lens feel (wide/normal/tele), light, movement.
• Acting: expressions, eyes, body, the emotional arc, the moments that sell the story.
• Storyboard: the shot-by-shot grammar (event → reaction → detail).
• Sound: see step 3.

3. Listen (sound)

Claude cannot hear, so use two sources and cross-check them:

1. Measurements: Read audio.json (loudness per 0.5 s and per shot, silences), waveform.png and spectrogram.png (red lines = cuts, white ticks = 10 s). Loud dense bands = action music or effects; a dark top band = highs filtered ("underwater"); clear horizontal lines = tones or voice; gaps = silence.
2. Gemini's ears (optional, ask the owner): Gemini can watch a YouTube link with audio. With dudu-browser, open a new Gemini chat (plain chat, no tool) and ask, in English, for a section-by-section sound breakdown with timestamps (music, sfx, silence, dialogue, hits on cuts) and an acting breakdown, telling it to say "not sure" when unsure. If the owner has no Gemini, rely on the measurements alone and say so in analysis.md. Keep what both sources agree on. Mark anything only Gemini reported (exact dialogue lines especially) as "per Gemini, unverified".

4. Write storyboard.json and analysis.md (same folder)

• storyboard.json: every shot with size, angle, lens, light, move, action, acting, sound, why, grouped into sequences. Schema in tools/storyboard.py. Render with python tools/storyboard.py library/study/<id>/storyboard.json, and screenshot it to check.
• analysis.md: header (source, length, type, the measured numbers) → Structure → What to steal → Camera direction → Acting → Sound design (a time / sound / technique table) → For our 30-60 s ads (the 3 most reusable moves) → What not to copy (real people, brands, length) → Numbers to reuse. Write techniques, not descriptions: "alternate face close-up and wide event shot in action beats", not "she drives a car".

5. Update the library index

Add one row to library/INDEX.md: id, title, type, pace (cuts per 10 s, median), and the 1-2 techniques that make this reference worth reusing.

6. Use it

When writing a brief, script, shot list or cut, check library/INDEX.md for matching references and say which technique you are borrowing and from where. When the owner says "like that ad", open its analysis.md first.

Rules

• Study only, never reuse footage, frames, music or likenesses. The download is for private analysis and is deleted; tell the owner once that downloading may conflict with the site's terms.
• Frame sheets stay local (git-ignored). analysis.md, shots.json and meta.json may be committed.
• Report back to the owner in their language: the 3-5 most useful techniques, with shot numbers as evidence.