euchmob/Youtube

---

name: video-editor description: עורך סרטונים — making YouTube Shorts edits from Avatar series (currently Avatar: Seven Havens). Use for ANY request to make, plan, edit or improve a video/short/edit for the user's YouTube channel. Holds the user's taste rules, what they liked and disliked, and the working technical pipeline.

עורך סרטונים (Video Editor)

The user makes YouTube Shorts edits about Avatar (right now: Avatar: Seven Havens). This skill is the memory of their taste. Keep it updated: after every video, add what the user liked to "Liked" and delete/replace anything they disliked. Their feedback always wins over anything written here.

Talk to the user in Hebrew. They follow along from their phone.

Workflow (always in this order)

1. Know the series first. Read series-notes.md (episodes, characters, themes, theories). Refresh it with a web search when new episodes have aired.
2. Pitch before building (rule 8). Before downloading/rendering anything expensive, send the user: the video idea in one line, the clip list (episode + scene, in order), the music choice, and the script word-for-word exactly as it will be heard/captioned. Wait for approval.
3. Find clean source clips, build, verify (captions-vs-speech check), send.
4. Update this skill with the feedback.

The user's rules

1. Understand the series — the hottest theories, the most interesting moments of each episode. Use that to pick ideas fans care about.
2. Clean, good-quality clips. Do NOT use TikTok reposts with burned-in subtitles/meme text and then try to erase them. Search harder for clean HD sources, even if it takes longer.
3. One topic per video, in series order. Every clip must serve the single topic (e.g. "Pavi is afraid of being the Avatar"). Clips must follow the show's chronology — never show episode 3 and then episode 1. 3b. The script must tell a story, not list clips. One through-line (e.g. "the twins remember nothing, and every adult around them hides something"): hook question → clues → twist → payoff, ideally ending on an open question that makes viewers comment. Every scene answers the previous scene's question and raises the next. When pitching, explain how each beat connects. Never state a fan theory as fact — let the edit imply it.
4. Creative, alive transitions. The video must feel energetic and alive — vary transitions (punch-in zooms, whip/blur, flash, shake on impacts, speed ramps, light leaks, match cuts). Don't repeat the same transition every cut.
5. Music is the most important part. Use music from Avatar: The Last Airbender (the original series score). Pick the track that best fits the mood and the timeline (build-ups landing on the emotional beat, drops on cuts). Keep searching for the best-fitting track.
6. Captions appear only while someone is speaking. A caption over silence = viewers leave. A short hold after a line ends is OK. Also: avoid long silent stretches in the video overall.
7. Caption design: glowing, beautiful, feels like it belongs to the Avatar world. Every character gets their own colour chosen from their personality/element; when that character speaks, their captions use that colour. Decide the palette before building.
8. Pitch first (see workflow) — don't burn tokens/time before the user approves the plan.
9. Visuals: the show already looks good — push it a bit further so it pops on a Shorts feed (contrast, saturation, warm/cool split, bloom/glow, subtle grain, vignette). Lively, eye-catching.

Format

• Vertical 1080x1920 for YouTube Shorts, ~30-55 s.
• Loudness ≈ -14 LUFS, true peak -1 dB. Dialogue clearly above the music (music ducks under speech).

Liked

• Glowing word-by-word captions that pop in as each word is spoken (first two edits).
• Warm colour grade with glow.

Disliked (don't do again)

• Just cutting the official trailer — too easy/boring. Use real episode scenes.
• TikTok clips with burned-in text (even after removal, it's not clean).
• Captions that show while nobody is speaking.
• Mixing topics / jumping around the timeline.
• Ideas built on old episodes (pitched eps 1-3 "Pavi's fear" → "not interesting"). Build ideas from the most recent batch of episodes (when asked: eps 4-6) — that's what's relevant to viewers now.

Character caption palette (Seven Havens)

Decide/confirm per video; keep consistent across videos once the user approves. See series-notes.md for personalities. Current proposal:

Character	Colour	Why
Pavi (Avatar, earthbender twin)	gold-amber #FFC93C + white-gold glow	Avatar light, warmth, earth
Nisha (twin, earthbender)	jade green #4CE08A	earthbender, fierce/loyal
Jae (Brigade airbender, narrator)	sky silver-white #E6F3FF + soft blue glow	air, calm, by-the-schedule
Karthik (Brigade commander)	deep orange #FF8A2E	leader, Brigade orange uniform
Korra (past Avatar)	cyan/Avatar-state white #B8F4FF	spirit / Avatar State
Villains / Faceless	blood red #FF3B3B	threat
Others	soft white #F2F2F2	neutral

Technical pipeline (what works in the cloud container)

Scripts live in pipeline/ in this repo.

• Clean clip sources: the official YouTube channel "Avatar Legends" posts clean HD episode clips — best source. Paramount+ is DRM (unusable). Trailer is on Dailymotion but only 288p. Don't try to bypass bot-checks (Invidious/Anubis, Cloudflare) — not allowed.
• Downloads: YouTube blocks the cloud IP ("Sign in to confirm you're not a bot") — needs the user's cookies.txt (yt-dlp --cookies cookies.txt), or run the session on the user's computer, or the user uploads the clips. TikTok works with yt-dlp --impersonate chrome (needs pip install curl_cffi) but official TikToks have burned-in text → avoid (rule 2). Dailymotion only serves 288p anonymously.
• Speech/captions: separate vocals with Demucs (python3 -m demucs --two-stems=vocals), transcribe the vocal stem with faster-whisper large-v3 (word timestamps), then snap every word onto voiced frames of the vocal stem (pipeline/build_short.py: verified_words) and clip caption events to voiced frames. Verify: zero caption frames over silence (±40 ms). Whisper mishears names: Ellora, Jae, Pavi, Nisha, Karthik — fix them.
• Mix: vocal stem + music (sidechain-ducked by dialogue) + SFX (pipeline/sfx.py synthesizes whoosh/impact/riser/subdrop) → loudnorm -14 LUFS.
• Captions: ASS rendered by ffmpeg/libass. Glow = 3 layers (wide blur, tight blur, crisp core). Font: Inter Display Black (installed); consider an Avatar-style display font.
• Output video is ~100+ MB — never commit .mp4 (GitHub 100 MB limit); send via SendUserFile.