Acim24/motion-video-audio

Motion Video with Audio (launch-film style)

Turns a brief into a finished, mastered MP4: launch-film visuals (warm paper, big serif headlines, one terracotta accent, chat demo typing itself, bars revealing the winner, quiet logo outro) plus a complete soundtrack:

• Voice-over — one optional vo line per scene, synthesized with edge-tts. Scenes size themselves to the narration.
• Music — an original ambient score synthesized to the exact timeline (pad, sub, piano arpeggios, soft pulse, riser, and a resolving hit on the logo frame). Royalty-free because it's generated from code. Or use the user's own track.
• SFX — whooshes on cuts, ticks on words/bars, typing while the reply types, a chime on the winning bar, counter ticks, impact + shimmer on the logo. Placed inside the scene components, so they stay in sync automatically.
• Mix — music ducks under narration; final files are normalized to −14 LUFS / −1.5 dBTP (social/YouTube target).

All content lives in one file, src/storyboard.ts.

Workflow

1. Understand the brief. Get: brand/product name, the one message to remember, 2–5 supporting points, real numbers (if any), target length, platform, narration language and voice gender. If the user gave enough, don't interrogate; state your assumptions. Never present invented benchmark numbers as real — use visibly labelled placeholders (footnote) or skip the benchmark.

2. Scaffold.

   bash <skill-dir>/scripts/new_project.sh <target-dir>
   

   Needs Node 18+, ffmpeg, and edge-tts on PATH (pipx install edge-tts) with internet for narration. Check with which edge-tts ffmpeg first; if edge-tts is missing and can't be installed, build without vo and tell the user.

3. Write the storyboard. Read references/style-guide.md (visual pacing, copywriting, recipes) and references/audio.md (narration writing, voices, music/SFX/mix settings). Scene types (fields in src/types.ts):

   type	use for
   headline	kinetic type: opener, section titles, closing line. *word* = accent italic
   chat	UI demo: prompt → thinking dots → typed reply (with typing sound)
   benchmark	horizontal bars, one highlight: true bar (chime when it lands)
   bigNumber	a big stat counting up (counter ticks + impact)
   features	2–4 numbered capability cards (pop per card)
   logo	outro: drawn mark or logoSrc, name, tagline, CTA (music hit lands here)

   Each scene: optional vo, optional durationInFrames (30 fps; omit it on narrated scenes to auto-fit), optional tone: 'light' | 'dark'. Leave the logo scene without vo — let the music hit breathe.

4. Build audio — this also fixes the final timeline:

   cd <target-dir> && npm run audio
   

   Read the printed timeline: per-scene start, length, narration length, and "(stretched to fit vo)" warnings. If a scene got stretched a lot, shorten its vo instead — long narration makes slow, sleepy scenes.

5. Visual QA (stills are fast, renders aren't):

   npx tsc --noEmit
   bash <skill-dir>/scripts/contact_sheet.sh <target-dir> Main-16x9 /path/sheet-wide.png
   bash <skill-dir>/scripts/contact_sheet.sh <target-dir> Main-9x16 /path/sheet-tall.png
   

   Read the PNGs; fix overflow, crowding, more than one accent per scene.

6. Audio QA without ears — after rendering, check loudness and look at the mix:

   ffmpeg -i out/video-16x9.mp4 -af ebur128=peak=true -f null - 2>&1 | grep -E "I:|Peak:"
   ffmpeg -y -i out/video-16x9.mp4 -filter_complex "showwavespic=s=1800x300" -frames:v 1 /path/wave.png
   

   Expect I ≈ −14 LUFS, peak ≤ −1 dBFS, narration bursts clearly standing above the music bed, and the biggest swell at the logo.

7. Render (runs audio → both formats → mastering):

   npm run render     # → out/video-16x9.mp4, out/video-9x16.mp4
   

   ~30 s video ≈ 1–2 min. Report paths, duration, voice used, and mention npm run studio for live preview with sound.

Customizing

• Brand colors/fonts: src/theme.ts tokens only.
• Logo: file in public/, logoSrc: 'logo.svg'.
• Own music: file in public/, audio.music: 'track.mp3' (only tracks the user has rights to). Own recorded narration: see references/audio.md.
• New scene types / formats / codecs: references/extending.md.

Brand and honesty guardrails

Emulate the style, not the brand: don't reproduce Anthropic's (or anyone's) logo or present the video as an official release of a company the user doesn't represent. Synthetic voices must not imitate a real, identifiable person. Benchmarks: real user data or visibly marked illustrative.