LeadGrow Video Pipeline v2

Storyboard first,
ship with confidence

Seven gated stages from raw Tella footage to verified final video. Every stage has a hard gate — no skipping the storyboard, no shipping without QA.

✓ 4 skills · 80 eval queries · 26-point QA
7
Gated stages
4
Skills
26
QA checks
23
Taste decisions
The Pipeline

Seven stages. Each has a gate that must pass before advancing. The storyboard is mandatory — you cannot build overlays without one.

01
Intake
Source material
energy level
target length
Gate
02
Storyboard
Visual treatment
per phase
user approval
Hard Gate
03
B-Roll
Capture plan
agent-browser
screen records
Soft Gate
04
Overlay Specs
Content + tokens
animation + timing
transcript-verified
Gate
05
EDL
Edit decision list
sync rule verified
segment map
Gate
06
Render
HyperFrames HTML
ffmpeg composite
frame verification
Gate
07
QA Review
26-point checklist
pass/fail scoring
fix → re-render loop
Gate
No overlay work without an approved storyboard. The pipeline enforces this.
Skill System

Four skills with clear responsibilities. The orchestrator loads dependencies in the right order and delegates to the right skill at each stage.

Conductor

video-pipeline

Single entry point for all video editing. Gates each stage, loads dependencies in order (brand.md → DESIGN.md → motion-editing → video-use), tracks pipeline state per project. Prevents skipping storyboard.

20 eval queries out of scope defined
Creative Director

motion-editing

Defines HOW to edit — cut cadence, zoom behavior, B-roll strategy, overlay choreography. Runs the 5-checkpoint storyboard flow. Prescribes creative direction.

20 eval queries 13 taste decisions out of scope
Quality Auditor

video-qa

MEASURES and scores rendered video against 26 quality checks. Produces structured pass/fail reports with timestamps and fix recommendations. Audits, not prescribes.

20 eval queries 26 checks out of scope
Brand Reference

design-system

Token LOOKUP for exact hex codes, fonts, type scale, spacing, easing, shadows. The reference book — you ask it what color ember is, not whether a composition passes QA.

20 eval queries 10 taste decisions out of scope
Stage Breakdown

What each stage does, what gates it, and who it delegates to.

Stage 01

Intake

Identify source material, set energy level, check prerequisites. Creates the pipeline state file.

Gate Transcript or raw footage exists. brand.md loaded. Energy level and target length set.
Delegates to Self-contained — video-pipeline handles intake directly.
Stage 02

Storyboard

Analyze transcript, propose visual treatment per phase — hook, sections, transitions, payoff. User approves each phase.

Hard Gate Storyboard produced and every phase approved by user. No overlay work until this passes.
Delegates to motion-editing CHECKPOINT 1 (Storyboard)
Stage 03

B-Roll Capture

Extract every B-roll reference from storyboard. Classify as captured, needs agent-browser, needs screen record, or needs design work.

Soft Gate Capture plan produced. Can proceed with pending items noted.
Delegates to motion-editing CHECKPOINT 2 + agent-browser for captures
Stage 04

Overlay Specifications

Design each overlay with exact brand tokens — content, style, animation type, timing verified against SRT timestamps.

Gate Every storyboard overlay has a spec. All timing verified against transcript. User approved.
Delegates to motion-editing CHECKPOINT 3 + design-system for tokens
Stage 05

EDL Assembly

Assemble approved storyboard + B-roll + overlay specs into a complete Edit Decision List with sync-verified timestamps.

Gate EDL JSON produced. Sync rule applied (src_start = AUDIO_START + out_start). User reviewed segment map.
Delegates to motion-editing CHECKPOINT 4 (EDL Assembly)
Stage 06

Render

Build HyperFrames compositions, render to MP4/PNG, composite onto base video with ffmpeg, mix audio.

HyperFrames ffmpeg PNG-sequence for alpha
Gate Each composition previewed. Frames extracted and visually verified at transitions. Audio volume checked.
  • PNG sequence for alpha overlays (never WebM)
  • amix normalize=0 — voice 100%, music 6%, SFX 40%
  • Self-contained CSS tokens in every composition
  • Three visual systems in every fullscreen card
Stage 07

QA Review

Run the full 26-point quality checklist. Flag failures with timestamps. Fix → re-render → re-check until all pass.

Gate All 26 checks run. User approves (pass all, or pass with acknowledged issues).
Delegates to video-qa skill (standalone 26-point audit)
  • Pacing — 5 checks (hook cadence, content cadence, TH duration, transitions, mode shift)
  • Cuts — 4 checks (types, J/L-cuts, fades, word boundaries)
  • Zooms — 3 checks (Ken Burns, frequency, direction)
  • B-Roll — 6 checks (real content, duration, zoom, density, escalation, demo footage)
  • Overlays — 5 checks (motion tokens, ember rule, settle time, hold duration, lower third)
  • Audio — 3 checks (filler words, fades, music placement)
  • Brand — 4 checks (palette, shadows, three visual systems, self-contained tokens)
Taste Decisions

23 production-validated editorial rules. Each came from a real correction or confirmed approach.

Never

Floating overlays on screenshare

Motion graphics are fullscreen transitions that replace the frame. Never floating cards on active demos.

Always

Three visual systems

Every fullscreen uses pixel emblems + whiteboard sketches + eyebrow pills. Missing one = incomplete.

Always

One ember per frame

Exactly one ember element. The stat, OR the CTA, OR the highlight word. Never multiple.

Never

Phantom SFX

SFX fires only on motion graphic transitions. No clicks on static assets, no SFX on TH cuts.

Always

PNG sequence for alpha

HyperFrames WebM has no alpha channel. Render to PNG sequence, feed to ffmpeg.

Always

Verify frames after build

Extract and check frames at transitions. Exit code 0 is not verification.

Never

Text-only overlay cards

Every card gets SVG illustration areas — pixel art, whiteboard sketches. Text-only = rejected.

Always

Self-contained tokens

Inline :root CSS vars in every composition. External tokens.css can fail to load.

Always

Timestamp from transcript

Every composition start matches word-level SRT timestamp. Never guess timing.

Critical Gotchas

Hard-won lessons from production. Ignore at your peril.

Storyboard is mandatory The pipeline will not allow overlay or motion graphic work without an approved storyboard. If you ask to "build an overlay" and no storyboard exists, the pipeline redirects to storyboarding first. This is by design — guessing at timing and content produces throwaway work.
HyperFrames WebM has no alpha --format=webm outputs yuv420p despite docs claiming transparency. Use --format=png-sequence for overlays, then feed PNGs directly into ffmpeg.
amix normalization crushes audio ffmpeg amix divides volume by input count. Always use normalize=0. Set levels explicitly: voice 100%, music 6%, SFX 40%.
EDL sync rule is the #1 desync source Every talking head src_start must equal AUDIO_START + out_start. Using raw SRT timestamps diverges as soon as B-roll shifts the output timeline.
Scribe transcripts are gitignored JSON cache costs real money to regenerate. Never delete raw/ until project is archived.