name: studio-video-director description: Studio-grade AI video and Shorts production pipeline. Use when planning, scripting, generating, assembling, or quality-gating any video: hackathon demos (16:9), Shorts/Reels (9:16), product commercials, movie breakdowns, viral clipping, or any multi-scene narrated video. Enforces true stereo audio, -14 to -16 LUFS loudness, a visual cut every 3-4 seconds, zero static slideshows, and a strict aspect-ratio lock. Free tier is $0 (local FFmpeg + Whisper only); Pro adds multi-scene assembly, auto-ducking, caption burning, niche templates, and batch gating.
Studio Video Director
A battle-tested, measurement-driven pipeline for producing studio-grade videos with free local tools. Every rule below was paid for with real failed builds: mono mixes, over-compressed audio, slideshow Shorts, clipped captions, and dead air. A build that violates any gate is a failed build -- fix it, don't ship it.
When to use this skill
- Any video deliverable: hackathon demo, product demo, Short/Reel, movie breakdown, commercial, tutorial, trailer, podcast clip.
- Before rendering: planning, script, storyboard, shot prompts.
- After rendering: the 10-gate pre-delivery check on the FINAL encode.
- When a video "looks fine but feels amateur": run the gate, fix the failing gates, re-check. The gate names the exact defect.
The 8-step workflow
- Brief -- one paragraph: audience, one idea, one CTA, length,
aspect (pick ONE and lock it:
16:9demo/long-form or9:16Shorts/Reels -- NEVER mix ratios in a project). - Crew plan (6 roles) -- Director (vision + cut list), Writer (script, <=30 words per shot, ~150 wpm), DP (camera + framing), Art Director (palette + set), VFX (motion + transitions), Sound Designer (VO + music + SFX). Each role signs off before generation.
- Storyboard + cut list -- written BEFORE animating. A visual or camera cut at least every 3-4 s (no gap > 4 s, hard fail). Hook in the first 3 s: title/question + motion. Never a static cover opener.
- Prompts (six-part) -- per segment, <=10 s each, using
references/prompt-template.md: duration/format, subject, scene, beats with timestamps, camera language, sound/dialogue, constraints. Reuse the subject description VERBATIM across segments. - Generate segments -- real motion only: screen recording (Playwright, deviceScaleFactor 2, 1920x1080@30fps for demos), B-roll, or animation. Ken Burns on stills: at most ONE segment of <=4 s per Short, never the whole video.
- Assemble -- normalize every scene to the locked frame
(scale + pad, never stretch), join with the concat demuxer, master
the audio (Pro:
pro-bundle/scripts/assemble_multiscene.py). - Sound -- neural VO only (edge-tts / studio TTS; gTTS banned). Music: a real licensed track (Mixkit / Pixabay / YouTube Audio Library), ducked -20 dB (Shorts) or -22 dB (demos) under the VO. Real designed SFX synced to cuts. Synthesized beds/SFX banned.
- Gate + deliver -- run
scripts/quality_gate.pyon the FINAL encode. All 10 gates must PASS. Save the master + .SRT inrenders/and never trash them.
Quality gates (all 10, no waivers)
| # | Gate | Rule |
|---|---|---|
| 1 | Specs | exact duration/resolution/fps; audio channels == 2 |
| 2 | Loudness | integrated -14..-16 LUFS, TP <= -1.5 dBTP, LRA >= 6 LU |
| 3 | Dead air | no silence >= 0.5 s except a trailing tail <= 1 s |
| 4 | Cut cadence | no gap between cuts > 4 s |
| 5 | VO coverage | faster-whisper transcript covers >= 80% of script words |
| 6 | Frames | 00:02 / mid / end frames pass the named frame checklist |
| 7 | Stereo | ffprobe channels == 2 on the final mix |
| 8 | Real music/SFX | human attests real licensed music + real SFX |
| 9 | No slideshow | no frozen segment > 4 s |
| 10 | Master protection | .SRT sidecar present, no *-broken / *-v1 / *-v2 siblings |
Run it:
python3 scripts/quality_gate.py final.mp4 --aspect 16:9 \
--srt final.srt --script vo-script.txt --frames-dir checks/
Individual gates:
python3 scripts/audio_lufs_check.py final.mp4
python3 scripts/extract_frames_check.py final.mp4 --aspect 16:9 --frames-dir checks/
Captions rule
- 16:9 demos: separate
.SRTsidecar ONLY. Zero burned-in narration captions. On-screen UI callouts are allowed -- they are not subtitle captions. - 9:16 Shorts/Reels:
.SRTsidecar AND styled lower-third safe-zone captions (the Shorts player ignores sidecars; muted viewers need on-screen text).
Named frame checklist (apply to 00:02, mid, end frames)
(a) nothing clipped at any edge, (b) no dead band/empty area > 25% of frame, (c) every on-screen word readable at phone size, (d) captions visible where the script says they should be, (e) no placeholder/fake UI text, (f) subject fills the frame.
$0 tooling (free tier)
- ffmpeg / ffprobe 8.x, Python 3.10+, faster-whisper (local, CPU).
- Optional external free credits only (OpusClip / Vizard / Quso / 2short free tiers) -- never invoked by any script in this repo, never a paid API, never a signup, never a card.
- This repo contains NO credentials, NO API keys, NO personal data.
Pro Edition
pro-bundle/ adds: multi-scene FFmpeg assembly with two-pass
loudnorm mastering, VO + music auto-ducking, Whisper .SRT + burned
caption builder, 18 niche templates, and a one-command batch gate.
See README.md (Pro Edition section) and pro-bundle/README.md.