elite-video-production

name	elite-video-production
description	Complete automated video production system for CC's content pipeline. Use whenever CC provides raw video footage and says "make this a post", "edit this", or asks for a cinematic output. Replaces a human video editor entirely. Covers hook engineering, pacing, caption styling, color grading, audio mastering, sound design, motion graphics, and multi-platform export.
triggers	["video","edit","post","caption","hook","grade","audio","export","reel","short","tiktok","b-roll","lower third","film grain","remotion","ffmpeg","whisper","cinematic"]
tier	full
dependencies	["content-engine"]
canon_references	["sutherland-signalling","dunford-positioning"]
canon_source	brain/MARKETING_CANON.md

Elite Video Production

Raw iPhone footage → cinematic, captioned, platform-ready output. No questions, no half-measures.

When CC Says "Make This a Post"

Run the full 15-step pipeline (Section 14) automatically. The only input needed is the raw video file path. Everything else is decided by this skill.

---

Section 1: The 3-Second Hook (Non-Negotiable)

The data:

• 65% of viewers who pass 3 seconds continue past 10 seconds (Facebook internal)
• TikTok gates For You Page distribution on 3-second hold rate — 60%+ triggers algorithmic push
• First frame is the single highest-leverage edit decision in the entire video

The 3 Hook Rule — all three must fire simultaneously:

Hook Type	What It Is	Rule
Visual	Movement or high contrast in frame 0	Never open on a static, low-contrast shot
Text	1-7 bold words overlaid immediately	No delay — on screen by frame 1
Verbal	Payload first, no intro	Never start with "Hey guys" or "So today"

Hook copy patterns ranked by hold rate:

1. Mistake/Loss Aversion: "Stop doing X" — fear of missing out drives pause
2. Secret/Curiosity Gap: "What nobody tells you about..." — information asymmetry
3. Bold Contradiction: State the opposite of conventional wisdom, no qualifier
4. Mid-Action Open: Video starts in the middle of something interesting, no setup

Curiosity gap templates (use verbatim or adapt):

• "I was wrong about [X]"
• "Here's what $[amount] actually looks like"
• "Nobody talks about this part"
• "I almost didn't share this"
• "The part they cut from every tutorial"

First frame engineering checklist:

• Tight crop: shoulders-up or tighter
• Strong emotional expression on face (not neutral)
• High contrast between subject and background
• Direct eye contact with lens
• Text overlay in upper-middle zone (not bottom 25%, not top 15%)
• Text: max 7 words, bold, high contrast font

---

Section 2: Pacing and Cut Frequency

2024 data: MrBeast reduced from 38 cuts/60s to 23 cuts/60s — viewership increased. Overstimulation editing is dead. Purposeful cuts win.

Cut rate by video length:

Duration	Target Cut Rate	Notes
10-20s viral clip	Every 1-2 seconds	Remove all silence and breath
30-60s educational	Every 3-5 seconds	Jump cuts + zoom punches
60-90s story	Every 5-8 seconds	Burst at key moments (contrast pattern)

Hard rules:

• Never more than 15 seconds without a visual or audio change
• Beat-synced cutting: cuts land on downbeat or snare of background track
• Pattern interrupts (zoom punch, B-roll insert, SFX) increase watch time 85% and conversion 32%

Optimal publish lengths by platform:

Platform	Optimal	Algorithmic sweet spot
TikTok	21-34s	Under 30s gets full re-watch loop credit
Instagram Reels	15-45s	15-30s for discovery, 45s for community
YouTube Shorts	15-35s	Loop completion is the key signal

---

Section 3: The Retention Edit Toolkit

Zoom Punch

Cut (not a slow zoom) from medium shot to 10-15% tighter on the same angle. Use at emphasis words, punchlines, and logic pivots. Implementation:

# FFmpeg scale + crop zoom punch (cut-based, not smooth):
# At timestamp of emphasis word, cut to:
-vf "scale=iw*1.12:-1,crop=iw/1.12:ih/1.12"

J-Cut

Audio from the next clip begins 0.5-2 seconds before the video cuts. Use for topic transitions and cutting into B-roll. Forces the viewer's brain to follow the audio before the image catches up.

L-Cut

Video cuts to new footage but audio continues from the previous clip. Use for screen recordings and demos while narration continues.

Jump Cut + Silence Removal

Cut every breath, filler word, and pause greater than 0.3 seconds. Non-negotiable for all talking-head content.

# auto-editor (pip install auto-editor):
auto-editor input.mp4 --margin 0.2sec --output silent-removed.mp4

Speed Ramp

Requires 120fps source footage. Ramp from 1x to 0.3x at the key visual, then accelerate to 2-4x through the transition. The slowest point lands on the beat.

# FFmpeg variable speed (requires setpts filter):
-vf "setpts=0.3*PTS"   # 3.3x speed
-vf "setpts=3.33*PTS"  # 0.3x slowmo

Flash Frame

1-3 white or black frames on a cut. Use for high-energy cuts, section transitions, and before major statements.

# Insert 2 white frames at cut point:
color=c=white:s=1080x1920:d=2/30,format=yuv420p

Camera Shake

2-4 frame oscillation on the X/Y axis. Duration 0.5-1 second. Coincides with bass hits or impact SFX.

# FFmpeg shake (apply at impact moment):
-vf "crop=iw-20:ih-20:(sin(n/2)*10)+10:(cos(n/2)*10)+10"

Freeze Frame

Pause video 0.5-1.5 seconds while audio continues. Use for reactions, data points, and joke landings.

# FFmpeg freeze at timestamp T:
-vf "setpts=if(between(t,T,T+1),T/TB,PTS)"

Pattern Interrupt Frequency

Duration	Interrupt Cadence	Method
10-30s	Every 2-4 seconds	Zoom punch, B-roll, SFX hit
30-60s	Every 3-6 seconds	Mixed toolkit
60-90s	Contrast Pattern: steady baseline + burst every 25-35s	5-10 rapid cuts in the burst

---

Section 4: Caption Engineering

The data: Captions increase watch time 12-40%. 92% of social video is watched without sound.

Word-by-word highlight beats full-sentence display — guides the eye at speech pace, never ahead or behind.

OASIS / Conaugh McKenna Brand Standard (Default)

Property	Value
Font	Roboto Bold or SF Pro Bold
Base color	White (#FFFFFF)
Keyword highlight	OASIS Blue (#0A84FF)
Animation	Smooth fade/slide-up — never a jarring pop
Case	Mixed-case — NOT all-caps
Chunk size	3-7 words per display, 1-3 seconds each
Position	Center-middle (avoid bottom 25%, top 15%)
Drop shadow	0 2px 8px rgba(0,0,0,0.6) — subtle, no thick outline

Classic Hormozi (High-Energy Option)

Property	Value
Font	Montserrat Bold
Case	ALL CAPS
Primary highlight	Yellow #f7c204
Secondary accent	Green #02fb23
Base	White
Animation	Word-by-word pop/scale on active word

Technical Implementation (ASS Subtitle Format)

WhisperX forced alignment produces word-level JSON timestamps. Convert to ASS with karaoke tags:

# Each word is a separate ASS Dialogue event
# Active word: brand color + scaleX 1.10
# Remaining words: dimmed white at 70% opacity

[Events]
Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
Dialogue: 0,0:00:01.24,0:00:01.58,Default,,0,0,0,,{\c&H0A84FF&\fscx110}STOP{\r} {\c&HFFFFFF&\alpha&H46&}doing this wrong

# \c&H0A84FF& = brand color (ASS uses BGR hex, reversed from RGB)
# \fscx110 = scale X to 110% for active word
# \alpha&H46& = ~70% opacity on inactive words

---

Section 5: Color Grading

iPhone Talking-Head Starting Point (FFmpeg)

-vf "curves=r='0/0 0.5/0.56 1/0.95':g='0/0 0.5/0.48 1/1':b='0/0.22 0.5/0.44 1/0.80',\
eq=saturation=0.80:contrast=1.05,\
vignette=PI/4,\
unsharp=5:5:0.5:3:3:0"

Named Grade Presets (use with --grade flag in pipeline script)

Preset	FFmpeg Filter	When to Use
teal_orange	Shadows→cyan curves, skin→warm orange	Most talking-head, high energy
editorial	eq=saturation=0.6:contrast=1.2, crushed blacks	Serious/professional content
warm	+200K WB feel, saturation boost, gentle vignette	Personal brand, lifestyle
clean	Neutral REC709, micro-contrast, vignette only	Screen recordings, tutorials

Full teal_orange implementation:

-vf "curves=r='0/0 0.25/0.30 0.75/0.82 1/0.95':g='0/0 0.5/0.50 1/1':b='0/0.18 0.25/0.35 0.75/0.60 1/0.75',\
eq=saturation=0.85:contrast=1.08,\
vignette=PI/4"

Grading Rules (Non-Negotiable)

• LUTs at 50-70% blend intensity, never 100%
• Crush blacks slightly — lift off true black for mobile viewing (blacks at luma ~16)
• Background desaturated relative to subject — creates subject separation without a key light
• Skin tones must land on vectorscope skin line (approximately 123° hue, 0.3-0.6 saturation)
• Export: Rec.709 color space, H.264 or H.265, 30fps minimum

---

Section 6: Audio Mastering

Complete FFmpeg Voice Mastering Chain

-af "agate=threshold=0.01:attack=80:release=840:makeup=1:ratio=3:knee=8,\
highpass=f=100:width_type=q:width=0.5,\
lowpass=f=10000,\
anequalizer=c0 f=250 w=100 g=2 t=1|c0 f=700 w=500 g=-5 t=1|c0 f=2000 w=1000 g=2 t=1,\
compand=attacks=0:points=-80/-900|-45/-15|-27/-9|0/-7|20/-7:gain=5,\
loudnorm=I=-14:TP=-1.0:LRA=11"

Chain Breakdown

Stage	Tool	What It Does
Gate	agate threshold=0.01	Kills breath noise and background hiss between words
Highpass	highpass f=100	Removes rumble, HVAC, handling noise below 100Hz
Lowpass	lowpass f=10000	Tames harshness, de-esses upper frequencies
EQ	anequalizer	+2dB at 250Hz (warmth), -5dB at 700Hz (mud), +2dB at 2kHz (presence/clarity)
Compression	compand	3:1 ratio, -7dB hard ceiling, consistent perceived loudness
Loudnorm	loudnorm I=-14	Final normalization to platform standard

Platform LUFS Targets

Platform	Target LUFS	True Peak
YouTube, TikTok, Reels	-14 LUFS	-1.0 dBTP
Podcasts / spoken audio	-16 LUFS	-1.0 dBTP
Broadcast (EBU R128)	-23 LUFS	-1.0 dBTP

Python Pre-Processing (Before FFmpeg)

# pip install noisereduce pedalboard soundfile numpy
import soundfile as sf
import noisereduce as nr
from pedalboard import Pedalboard, Compressor, HighpassFilter, LowShelfFilter

# Step 1: Spectral noise reduction
audio, sr = sf.read("raw_audio.wav")
reduced = nr.reduce_noise(y=audio, sr=sr, stationary=False, prop_decrease=0.75)

# Step 2: Pedalboard studio chain (300x faster than pySoX)
board = Pedalboard([
    HighpassFilter(cutoff_frequency_hz=80),
    Compressor(threshold_db=-18, ratio=3.0, attack_ms=5.0, release_ms=100.0),
    LowShelfFilter(cutoff_frequency_hz=200, gain_db=2.0),
])
processed = board(reduced, sr)
sf.write("processed_audio.wav", processed, sr)
# Hand off processed_audio.wav to FFmpeg mastering chain

---

Section 7: Sound Design

Five Core SFX — Exact Timing

SFX	When to Use	Timing Relative to Cut
Whoosh	Transitions, text reveals, caption slide-ins	Start 2-4 frames BEFORE cut, peak on cut frame
Hit / Impact	Jump cuts, title drops, zoom punches	Exactly on cut frame
Riser	Building to reveal, end of section	1-3 seconds before payoff moment
Bass Drop / Boom	Big scene change, dramatic statement	First frame of the new scene
Pop / Click	Caption word pop, UI element reveals	Exactly synced to text appearance frame

Volume Mixing (Voice Is Reference at 0dB)

Layer	Level Below Voice
Background music (calm)	-20 to -25 dB
Background music (high energy)	-8 to -12 dB
SFX	Just below music (felt, not heard)

Hormozi SFX Formula

• Zoom punch → impact hit SFX exactly on the cut frame
• Caption word pop → click at -20 dB exactly on text appearance
• Section transition → whoosh starting 3 frames before cut
• Major claim → 0.5-1 second of silence before delivery, then hit SFX on the first word

Remotion Free SFX Assets

https://remotion.media/whoosh.wav
https://remotion.media/whip.wav
https://remotion.media/page-turn.wav
https://remotion.media/switch.wav
https://remotion.media/mouse-click.wav
https://remotion.media/shutter-modern.wav
https://remotion.media/shutter-old.wav

---

Section 8: B-Roll and Overlays

B-Roll Duration Rules

Content Type	B-Roll Clip Length
Fast social (TikTok/Reels)	1-3 seconds
Educational / explainer	3-7 seconds
Hard maximum (short-form)	10 seconds

Timing rule: B-roll appears the MOMENT you say the relevant word — not after the sentence ends.

Meme/reaction inserts: 1-2 seconds maximum, placed after the punchline only. Maximum 2 per video.

Screen recordings: Use L-cut (voice continues). Wrap in a phone or laptop frame mockup. Duration 3-10 seconds.

B-Roll Sourcing Hierarchy (in order of impact)

1. Real footage CC shot — strongest authenticity signal, highest retention
2. Screen recordings of real results (Stripe dashboard, Skool analytics, live demos)
3. Higgsfield-generated cinematic B-roll (LIVE as of 2026-05-24) — the primary engine for any B-roll CC didn't shoot. CLI bound to [e-mail], see brain/integrations/higgsfield.md.
   • Seedance 2.0 (seedance_2_0) — short 1-5s cinematic VIDEO cutaways. Use skills/cinema-worldbuilder/SKILL.md to write the prompt (picks cinema mode, locks camera grammar, gates characters).
   • Nano Banana Pro (nano_banana_2) — production-grade stills. Use skills/banana-pro-director/SKILL.md to write the prompt.
   • Soul Cinematic (soul_cinematic) — atmosphere / mood / brand-register stills.
   • Marketing Studio Image/Video (marketing_studio_image / marketing_studio_video) — product-style commerce B-roll.
4. Fal.ai Flux Schnell — fallback only. Use when (a) CC's Higgsfield credits are exhausted, (b) the still is throwaway and < 1 second on screen so Fal's lower quality is invisible, (c) speed matters more than fidelity. $0.003/image vs Higgsfield's credit cost.
5. Text overlays as visual substitute when no relevant footage exists.

Higgsfield CLI for B-Roll (Primary)

Always cost-check before submit. Credits cost real money the second the subscription kicks in. Submission goes through scripts/higgsfield_gen.py — it cost-checks first, enforces a per-job credit cap (HIGGSFIELD_MAX_CREDITS_PER_JOB, default 2), refuses jobs that would overrun the account balance, and is idempotent (same model+prompt+duration returns the cached artifact instead of resubmitting).

# Step 1: Write the prompt via cinema-worldbuilder (for video) or banana-pro-director (for stills).
#         These skills handle cinema-mode picking, camera locks, character gates.

# Step 2 (optional -- generate does this again internally): eyeball the cost.
python scripts/higgsfield_gen.py cost --model seedance_2_0 --duration 3 --json

# Step 3: Generate. Cost-check -> budget guard -> submit --wait -> download, one command.
python scripts/higgsfield_gen.py generate --model seedance_2_0 \
    --prompt-file media/prompts/<slug>.txt --duration 3 \
    --out data/higgsfield/<brand>/ --json
# Artifact lands at data/higgsfield/<brand>/<job_id>.mp4 with a <job_id>.meta.json
# sidecar (cost, credits before/after, prompt hash). Image models (nano_banana_2,
# soul_cinematic, gpt_image_2) need no --duration -- the wrapper omits it.
# HIGGSFIELD_MOCK=1 dry-runs the whole chain with a real placeholder MP4.

Reference-image jobs: the wrapper doesn't pass --image yet. For a job that needs a character-lock reference (--image ./media/refs/character_lock.png, local path auto-uploaded or upload_id), use the raw higgsfield generate create ... --image ... --wait --json surface documented in brain/integrations/higgsfield.md — cost-check manually first.

Model picker for B-roll specifically:

• Talking-head support cutaways → seedance_2_0 with M2 Studio/Editorial mode (white-void brand register)
• Atmosphere / manifesto plates → seedance_2_0 with M5 Atmospheric mode (OASIS palette baked into prompt)
• Product demos → marketing_studio_video (commerce grammar built in)
• Cheap fast stills (when video is overkill) → nano_banana_2 or soul_cinematic

Fal.ai (Legacy Fallback)

# pip install fal-client
import fal_client

result = fal_client.subscribe(
    "fal-ai/flux/schnell",
    arguments={
        "prompt": "Professional iPhone-style photo, [description matching video topic]",
        "image_size": "portrait_4_3",
        "num_inference_steps": 4,
    },
)
# result["images"][0]["url"] → download and overlay at timestamp

---

Section 9: Cinematic Effects (FFmpeg)

# Vignette (standard):
vignette=PI/4

# Lens distortion (subtle barrel/pincushion):
lenscorrection=k1=-0.15:k2=-0.05

# Chromatic aberration (split RGB channels):
rgbashift=rh=-3:gh=0:bh=3

# Motion blur (frame blending):
tblend=all_mode=average

# Dynamic zoom push-in (subtle, Ken Burns-style):
scale=iw*1.05:-1,crop=iw/1.05:ih/1.05

# Ken Burns slow zoom (for still image B-roll, 250 frames):
zoompan=z='min(zoom+0.0015,1.5)':d=250:s=1080x1920:fps=25

# Film grain (luminance-masked, 35mm look):
# Layer 1: Generate grain noise
# Layer 2: Mask by luminance (grain less visible in blacks)
-filter_complex "[0:v]noise=alls=25:allf=t+u[grain];[0:v][grain]blend=c0_mode=overlay:c0_opacity=0.08[out]"

# Light leak overlay (requires light_leak.mp4 asset):
-filter_complex "[1:v]colorchannelmixer=aa=0.3[leak];[0:v][leak]blend=c0_mode=screen[out]"

Morph Cut (Jump Cut Smoothing)

For jump cuts on talking-head footage where the background doesn't change:

# FFmpeg interpolation approximation:
minterpolate=fps=60:mi_mode=mci:mc_mode=aobmc

# Better: RIFE Python model (Real-Time Intermediate Flow Estimation)
# pip install rife-ncnn-vulkan
# Produces true motion-interpolated frames between cuts

---

Section 10: Auto-Reframing (Landscape to Portrait)

Landscape 16:9 → Portrait 9:16 (1080x1920) with tracked crop.

Face Detection Pipeline

# pip install mediapipe ultralytics deep-sort-realtime opencv-python numpy

import mediapipe as mp
import cv2
import numpy as np
from deep_sort_realtime.deepsort_tracker import DeepSort

# MediaPipe: production-grade, better than Haar cascades on iPhone footage
mp_face = mp.solutions.face_detection
tracker = DeepSort(max_age=30)

# Rule of thirds positioning: eyes at y=640px in 1080x1920
# Headroom: 15-20% of frame height above head top
TARGET_EYE_Y = 640
HEADROOM_RATIO = 0.175

def smooth_crop_coordinates(history: list, new_x: int, new_y: int, weight: float = 0.85) -> tuple:
    """Weighted moving average prevents jitter."""
    if not history:
        return new_x, new_y
    prev_x, prev_y = history[-1]
    return int(prev_x * weight + new_x * (1 - weight)), int(prev_y * weight + new_y * (1 - weight))

---

Section 11: AI-Powered Pipeline Tools

Stage	Tool	Install	Notes
Transcription	WhisperX + large-v3-turbo	pip install whisperx	6x faster than base Whisper, forced word alignment
Silence removal	auto-editor	pip install auto-editor	Pairs with WhisperX word gaps for filler detection
Filler detection	WhisperX word list scan	—	Scan for: "um", "uh", "like", "you know", "so", "basically", "literally"
Scene detection	PySceneDetect	pip install scenedetect	ContentDetector + librosa energy peaks
Highlight scoring	Claude API	—	Score transcript chunks for hook strength and virality potential
Caption styling	ASS subtitle format	—	Word-level with brand color highlights (see Section 4)
Audio pre-processing	noisereduce + pedalboard	pip install noisereduce pedalboard	Run before FFmpeg chain
B-roll generation	Fal.ai Flux Schnell	pip install fal-client	$0.003/image, <1 second
Face tracking	ultralytics + DeepSORT	pip install ultralytics deep-sort-realtime	Smooth portrait crop

WhisperX Transcription

import whisperx
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
model = whisperx.load_model("large-v3-turbo", device, compute_type="float16")

audio = whisperx.load_audio("processed_audio.wav")
result = model.transcribe(audio, batch_size=16)

# Forced alignment — word-level timestamps
align_model, metadata = whisperx.load_align_model(
    language_code=result["language"], device=device
)
result = whisperx.align(result["segments"], align_model, metadata, audio, device)
# result["word_segments"] → [{word, start, end, score}, ...]

---

Section 12: Lower Thirds

Specification for 1080x1920 portrait safe zone.

Property	Value
Y position	1704px (120px above safe zone bottom)
Name font size	52-60px Bold Inter or Helvetica Neue
Title font size	36-40px Regular or Medium weight
Name color	#FFFFFF
Title color	#CCCCCC
Text shadow	0 2px 8px rgba(0,0,0,0.6)
Background	rgba(0,0,0,0.6) or brand color at 80% opacity
Max lines	2
Max characters per line	28

Animation spec (Remotion spring):

• Animate in: 18-24 frames, spring {stiffness: 120, damping: 60}
• Hold duration: minimum 3 seconds (90 frames at 30fps)
• Animate out: 12-18 frames ease-in

---

Section 13: Motion Graphics (Remotion)

Spring Physics Presets

import { spring, useCurrentFrame, useVideoConfig } from "remotion";

const frame = useCurrentFrame();
const { fps } = useVideoConfig();

// Cinematic title entrance
const cinematicTitle = spring({ frame, fps, config: { mass: 1, damping: 80, stiffness: 100 } });

// Premium slide-in (card, panel)
const premiumSlide = spring({ frame, fps, config: { mass: 1, damping: 30, stiffness: 120 } });

// Lower third
const lowerThird = spring({ frame, fps, config: { mass: 0.8, damping: 60, stiffness: 150 } });

// Word pop (captions)
const wordPop = spring({ frame, fps, from: 0.8, to: 1.0,
  config: { mass: 0.5, damping: 40, stiffness: 200 } });

// Staggered children (index = item index in list)
const staggeredFrame = frame - (index * 4);

12 Animation Principles Encoded

Principle	Implementation
Anticipation	Move 3-5px in the opposite direction for 4-6 frames before main movement
Follow-through	overshootClamping: false, allow 2-8% overshoot past target
Overlap	Stagger child elements by 4-8 frames each
Squash/stretch	Start scaleX: 1.05, scaleY: 0.95, spring to 1, 1
Never over-animate	Maximum 3 properties animated simultaneously per element

---

Section 14: Production Workflow (15-Step Pipeline)

This runs in order, every time. No steps are optional.

1.  ENHANCE AUDIO       noisereduce + pedalboard pre-processing → processed_audio.wav
2.  TRANSCRIBE          WhisperX large-v3-turbo with forced alignment → word_segments.json
3.  DETECT FILLERS      Scan word_segments.json for: um, uh, like, you know, so, basically, literally
4.  REMOVE SILENCE      auto-editor --margin 0.2sec + manual filler word cuts → clean_cut.mp4
5.  GENERATE CAPTIONS   Build ASS file from word_segments (brand style or Hormozi)
6.  ADD ZOOM PUNCHES    librosa volume spike detection → insert zoom punch at top 5-8 energy peaks
7.  INSERT B-ROLL       Claude API scores transcript → Fal.ai generates images → overlay at mentions
8.  COLOR GRADE         Apply --grade preset via FFmpeg curves + eq + vignette
9.  MASTER AUDIO        Full FFmpeg chain: gate → highpass → lowpass → EQ → compand → loudnorm -14 LUFS
10. ADD SFX LAYER       Whoosh on transitions, pop on caption words, hit on zoom punches
11. RENDER LOWER THIRDS ASS event or Remotion overlay if lower third needed
12. ADD FILM GRAIN      Luminance-masked grain at 8% opacity + vignette PI/4
13. CHECK FIRST FRAME   Verify: face tight crop, high contrast, text overlay present, no dead eyes
14. EXPORT PER PLATFORM Trim duration, adjust aspect, set caption length per spec (table below)
15. SCHEDULE            Zernio API (scripts/late_tool.py) across all 6 platforms

Platform Export Specifications

Platform	Max Duration	Aspect	Resolution	Caption Max
Instagram Reels	90s	9:16	1080x1920	2,200 chars
TikTok	180s	9:16	1080x1920	4,000 chars
YouTube Shorts	60s	9:16	1080x1920	5,000 chars
LinkedIn	600s	9:16	1080x1920	3,000 chars
Facebook	240s	9:16	1080x1920	63,206 chars
X/Twitter	140s	9:16	1080x1920	280 chars

FFmpeg Export Command (Base)

ffmpeg -i clean_cut.mp4 -i processed_audio.wav \
  -vf "[video_filters_from_steps_above]" \
  -af "[audio_filters_from_step_9]" \
  -c:v libx264 -crf 18 -preset slow \
  -c:a aac -b:a 192k -ar 48000 \
  -color_primaries bt709 -color_trc bt709 -colorspace bt709 \
  -movflags +faststart \
  output_1080x1920.mp4

---

Section 15: Quality Gate

Before delivering any output, verify all of the following:

Hook:

• First frame: tight crop, expression, eye contact, high contrast
• Text overlay present on frame 1 (max 7 words)
• First spoken word is the payload, not an intro

Pacing:

• No gap exceeding 15 seconds without visual or audio change
• Silence and filler words removed
• Cut rate matches video length tier (Section 2)

Captions:

• Word-by-word highlight active (not full sentence)
• Brand style applied (OASIS Blue or Hormozi depending on content)
• No captions in bottom 25% or top 15% of frame

Audio:

• Measures at -14 LUFS (or target per platform)
• True peak at or below -1.0 dBTP
• No clipping, no audible room noise, no breath artifacts

Color:

• Skin tones on vectorscope skin line
• Blacks slightly lifted for mobile viewing
• Export is Rec.709

Output:

• Duration within platform maximum
• Resolution 1080x1920
• File size under platform upload limit

---

Section 16: Iteration & Taste Profile Evolution

The 15-step pipeline above defines the system. The taste profile defines CC's specific preferences inside that system. Read the taste profile BEFORE every edit. Update it AFTER every approve/reject.

The loop

1. CC drops raw footage.
2. Read brain/video-style/taste_profile.md FIRST — before transcripts, before B-roll planning, before anything.
3. Plan the edit against (taste_profile.md) ∩ (this skill's 15-step pipeline).
4. Run scripts/video_editor.py edit (or compose manually for the long-form custom case).
5. Telegram review → CC approves, asks for changes, or rejects.
6. Log the outcome:
   - approved + clean → brain/video-style/approved_edits/YYYY-MM-DD_<slug>.md, note what landed
   - approved with note ("good but next time X") → log to approved_edits AND propagate X to taste_profile.md
   - rejected → brain/video-style/rejected_edits/YYYY-MM-DD_<slug>.md, propagate CC's verbatim reason to taste_profile.md
7. Next edit starts at step 2. Taste profile is now sharper.

What "propagate to taste_profile.md" means

Structured edit, not vague note:

• Find the relevant section (Hook / Pacing / Captions / Grade / Audio / SFX / B-roll / Lower thirds / Motion / Length / OASIS-overrides / Things-CC-does-NOT-want)
• Add a dated bullet: - 2026-MM-DD: <preference>. Source: rejected_edits/<slug>.md. Why: <CC's actual words>.
• If the new entry contradicts an older one, strikethrough the older with ~~text~~ and reference the newer. Don't delete history — show evolution.
• Update the last_updated: frontmatter field.

Promotion (preference → structural rule)

When the same preference appears 3+ times across edits with consistent CC language, promote it from a dated bullet to the structural copy above the bullets in that section. Mark the source bullets with ✓. This converts "CC's recurring feedback" into "default behavior" — the system gets opinionated.

Quarterly consolidation

Every ~90 days OR every ~10 edits (whichever comes first), do a consolidation pass:

• Read every entry in taste_profile.md
• Merge duplicates
• Promote stable preferences (3+ confirmations) to structural copy
• Update the last_consolidation: field
• Note the consolidation in the Iteration Log table at the bottom of taste_profile.md

When the taste profile and the 15-step pipeline disagree

Taste profile wins. The pipeline is the universal best-practice; the taste profile is CC. If pipeline default says "teal_orange grade" and taste profile says "CC consistently picks editorial for personal-brand pieces," editorial wins for personal-brand pieces.

Where this connects

• brain/video-style/README.md — full explanation of the system
• brain/video-style/taste_profile.md — the living document
• brain/video-style/approved_edits/ + rejected_edits/ — the source corpus
• skills/video-editing/SKILL.md Step 8 (Telegram review) — the capture point for approve/reject feedback

---

Obsidian Links

• [[skills/content-engine/SKILL]] | [[brain/CAPABILITIES]] | [[memory/content-strategy]]
• memory/content_pipeline_vision.md | [[skills/INDEX]]
• [[brain/playbooks/youtube_video_pipeline]] | [[brain/integrations/claude-video]]
• [[brain/video-style/README]] | [[brain/video-style/taste_profile]] | [[brain/integrations/higgsfield]]