Монтаж в Claude Code

Edit my talking-head video into a vertical "before / after" showcase reel with motion graphics. Source: [PATH TO VIDEO]. In the video I speak instructions to the editor ("move me left", "show a views counter", …): the edit must literally execute every instruction at the moment I say it.

## 1. Analysis
- Probe the source (resolution, fps, duration). Keep the output fps equal to the source fps.
- Transcribe with word-level timestamps (Whisper large-v3, Russian). Merge sub-word tokens into words.
- Grab full-res frames to find where my face is and which areas of the frame are free for text: never put text over the face or mouth.
- Split the speech into scenes, one per instruction, and plan a graphic for each, timed to the exact words.

## 2. Layout (1080×1920, 9:16)
- Background: dark navy at the top fading to bright electric blue at the bottom (#02041a → #040a33 → #0a1f8a → #1440ff), plus a soft blue radial glow at the bottom center.
- Top: the title "OPUS 5.5" in white Montserrat SemiBold, ~118px, centered at y≈330.
- Top window (x=48, y=400, 984×554, 16:9, square corners, drop shadow): the EDITED version, built at 1920×1080 and scaled down into the window.
- Label "original" (white, Montserrat SemiBold, ~50px) under the top window.
- Bottom window (same size, y≈1062): the untouched source video, muted and in sync with the top window.
- Audio: voice from the source plus SFX from the edited version.

## 3. Visual style of the edited version
- Kinetic text in Montserrat ExtraBold/Black, white with a soft dark shadow. Key words get a blue gradient (#c3d3ff → #2f5bff); "Opus 5.5" / "Claude" get an orange gradient (#ffd0b0 → #d9573a) and a rotating 12-ray orange spark.
- UI elements are dark glass cards: rgba(12,16,38,0.72), 1px white border at 14% opacity, radius 22–30, backdrop blur. Small chips have a glowing gold bar on the left (like "ТОП-5").
- Motion: ease-out entrances of 0.4–0.6s with blur → sharp, 50–100ms stagger, slight overshoot only on pops. Something always moves (slow push-ins, counters, shimmer). Straight cuts only, no fancy wipes.
- Reframe the source with smooth eased moves (translate + scale) when the speech asks for it ("move me left" / "back to center"). Always scale enough that no empty edges show.
- Short bottom captions (max 3 words, lowercase, active word in the blue gradient) only where no other text duplicates the speech.

## 4. Scenes (adapt to the actual transcript)
1. **Hook.** On the left "можешь и дальше собирать ролики ~~вручную~~", with an animated red strike-through. On the right "а я покажу, что умеет **Opus 5.5**": a big orange 3D-extruded title with an impact sound.
2. **"Move me left, show 5 rules on the right."** The video slides left; a full-height blue gradient panel slides in on the right, with a glowing left edge and an orange glow in the corner. Chip "ТОП-5" → title built word by word → "миллионы просмотров" in the gradient → 5 numbered glass cards that appear one by one, exactly when I say "one by one on separate cards", each with a pop and a brief highlight.
3. **"Back to center, show my most popular reel, run a views counter."** Reframe to center. A Reels-style phone card pops in on the right (thumbnail, heart/comment/share icons, "🔥 ТОП-1" badge). Under it a glass "ПРОСМОТРЫ" counter: it ticks slowly at "run", then rockets up (ease-out-expo) to the final number at "fast up to the final value". Then flash, impact and sparkle.
4. **"+50K subscribers in a month thanks to Claude, use a progress bar instead of a chart."** On the right: "за последний месяц / +50 000 / подписчиков / благодаря Claude ✷". Then a plain line chart appears and gets crossed out with a red X. It is replaced by a wide glass progress bar (milestones 10K–50K, shimmer, live number) that fills from 0 to 50 000 while I say it, with a pop on each milestone and an impact at the end.
5. **"Sound design: first without effects, then add sounds."** Big "саунд-дизайн меняет видео" with an animated equalizer. A pill toggle "ЗВУК: ВЫКЛ" while three demo actions play in TOTAL silence (card slide-in, light-sweep transition, letters popping in). On "add sounds" the toggle flips to "ЗВУК: ВКЛ" with a click. The same three actions replay, each with its SFX (whoosh, cinematic whoosh, pops per letter), while a DAW-like "SFX · ДОРОЖКИ" timeline with a moving playhead lights up its clips.
6. **"Show 20 caption styles on the same phrase."** Chip "ТОП-20 · СТИЛИ ТИТРОВ". The word "монтаж" cycles through 20 styles (~7 frames each), each switch with a soft click: classic stroke, neon, gradient, outline, 3D shadow, yellow box, serif italic, chrome, glitch RGB split, handwriting, pill, pixel, fire, thin minimal, highlighter, glass, comic, brutal, cinema, Opus orange. Show the counter "07 / 20 · Неон" and 20 progress dots.
7. **CTA.** "хочешь тоже монтировать с Opus 5.5?" → "напиши **МОНТАЖ** в комментариях". A glass comment bar types "монтаж" letter by letter (key-press sounds), and the send button pulses. On the left, comment bubbles pop in with pings ("монтаж 🔥", "монтаж!", "уже попробовала — собрала ролик за вечер").

## 5. Sound
- Voice at 1.0, no ducking, no limiter on the voice. SFX quiet relative to the voice: whoosh ~0.11, pop ~0.13, ping ~0.10, sparkle ~0.12, impact ~0.2. Every SFX lands on a motion hit. Target ≈ −20 to −18 LUFS, peak ≤ −1.5 dB.
- No background music unless I provide a licensed track.

## 6. Tech & QA
- Build in Remotion (React): compositions `OpusMontage` (1080×1920) and `Edited16x9` (1920×1080). Keep scene timings and SFX cues in one place, and keep placeholders (view count, reel thumbnail, title) as constants.
- Before the full render, render stills at every scene and check them: no text clipped at the edges, nothing over the face, layers ≥1s, no static stretches ≥3s.
- Render H.264, CRF ~17, AAC 320k. Report which numbers or images are placeholders I need to replace.