ai-commercial

A commercial-grade ad pipeline. Combines the best of two worlds:

• Generative cinematic shots via Runway ML (Veo, Gen4, ElevenLabs proxy) — for hero moments, atmosphere, B-roll where stock footage would otherwise be needed.
• VO + music + slide composition via direct ElevenLabs API + ffmpeg — for the structural backbone of a commercial (narration, CTA, brand reveal).

Default output: 30-90s commercial with VO + music + brand grade + CTA, in whatever aspect ratio the brief specifies (typically 16:9 for web/YouTube, 9:16 for IG/TikTok stories, 1:1 for IG feed).

Decision tree — which skill to use

The commercial workflow

A commercial has a fixed beat structure (you can break it, but most don't):

1. HOOK (3-5s) — cinematic Runway shot or product close-up that stops the scroll
2. PROBLEM / SETUP (5-15s) — VO over screenshots, lifestyle imagery, or AI B-roll
3. SOLUTION / OFFER (15-40s) — product walkthrough, value props, benefits
4. PROOF (optional 5-15s) — testimonial placeholders, logos, stats
5. CTA (5-10s) — logo + URL + offer + voice "Visit example.com today"

All 4 templates in this skill use a variant of this structure.

4 templates (start here for each commercial)

templates/
├── saas-demo.json         # 45-60s SaaS commercial
├── local-service.json     # 30s service ad (plumber, restaurant, etc.)
├── personal-brand.json    # 45-60s "who I am" / hire-me piece
└── product-launch.json    # 60-90s digital product launch trailer

Each template defines:

• Default duration
• Default aspect ratio (overrideable per project)
• Scene list with type, duration, copy placeholders
• Voice characteristics (gender, age, tone)
• Music style hint

Copy a template into your project as brief.json, fill in the placeholders, run.

Quick start

# 1. Set up project
mkdir -p ~/clientname-commercial/{assets,brief,output}
cp ~/.claude/skills/ai-commercial/templates/saas-demo.json ~/clientname-commercial/brief/brief.json

# 2. Fill in the brief (or edit interactively — see "The brief schema" below)

# 3. Place assets in the project
cp logo.png ~/clientname-commercial/assets/logo.png
cp product-screenshot.png ~/clientname-commercial/assets/screenshot1.png
# ... etc

# 4. Run
export PROJECT_DIR=~/clientname-commercial
python3 ~/.claude/skills/ai-commercial/scripts/make_commercial.py

make_commercial.py reads brief/brief.json, dispatches each scene by type, composes the final commercial(s) in the requested aspect ratios, and saves to output/commercial-{aspect}.mp4.

The brief schema

The brief is a JSON file describing the entire commercial. Full schema in references/brief-schema.md. The essential fields:

{
  "type": "saas-demo",
  "duration": 60,
  "aspectRatios": ["16:9", "9:16"],
  "voice": {
    "voice_id": "ELEVENLABS_VOICE_ID_HERE_OR_USE_DEFAULT",
    "style": "commercial-narrator",
    "stability": 0.5,
    "similarity_boost": 0.75
  },
  "music": {
    "style": "cinematic",
    "volume": 0.25
  },
  "brand": {
    "name": "ClientName",
    "logo": "assets/logo.png",
    "tagline": "Their tagline",
    "url": "https://example.com",
    "primaryColor": "#3b82f6",
    "accentColor": "#8b5cf6"
  },
  "scenes": [
    {
      "id": "hook",
      "duration": 5,
      "type": "runway_atmosphere",
      "prompt": "Cinematic close-up of a developer's hands typing on a sleek mechanical keyboard, soft blue and purple LED light from the screen, shallow depth of field, slow push-in."
    },
    {
      "id": "problem",
      "duration": 8,
      "type": "screenshot_kenburns",
      "image": "assets/old-spreadsheet.png",
      "vo": "Tired of managing your client invoices in a spreadsheet?",
      "motion": "zoom_in"
    },
    {
      "id": "solution",
      "duration": 20,
      "type": "screenshot_kenburns",
      "image": "assets/clientname-dashboard.png",
      "vo": "ClientName replaces your spreadsheet with a beautiful dashboard that actually does the work for you.",
      "motion": "pan_right"
    },
    {
      "id": "broll-1",
      "duration": 5,
      "type": "runway_i2v",
      "seedImage": "assets/team-photo.jpg",
      "prompt": "Slow camera push-in on the team, subtle smiles, warm office lighting, shallow depth of field.",
      "vo": "Built by people who understand your business."
    },
    {
      "id": "cta",
      "duration": 6,
      "type": "cta_slide",
      "headline": "Try it free for 14 days",
      "subhead": "No credit card required",
      "vo": "Visit examplebrand.com today.",
      "ctaUrl": "examplebrand.com"
    }
  ]
}

Scene types (dispatcher contract)

make_commercial.py dispatches scenes by type:

If a scene has a vo field, ElevenLabs synthesizes it and the VO is overlaid on the scene's visuals. Music is added underneath at music.volume and DUCKED to ~0.4× during VO segments.

Voice selection (per commercial)

ElevenLabs voice IDs determine the commercial's personality. The skill defaults to $ELEVENLABS_VOICE_ID (your env default), but each brief can override:

• Commercial narrator — Adam (pNInz6obpgDQGcFmaJgB) — deep male, authoritative
• Warm female — Rachel (21m00Tcm4TlvDq8ikWAM) — warm, friendly
• Energetic male — Antoni (ErXwobaYiN019PkySvjV) — younger, upbeat
• Calm female — Domi (AZnzlk1XvdvUeBnXmlld) — soft, intimate
• British male — Daniel (onwK4e9ZLuTAKqWW03F9) — refined, polished

Full reference: references/voice-selection.md. For each commercial, pick the voice that matches the brand personality. Local plumber → energetic male. Luxury watch brand → British male. Therapy app → calm female.

Music style hints

The music.style field in the brief drives the procedural music generator:

Music plays underneath the entire commercial at music.volume (typically 0.20-0.30), with automatic ducking when VO is active. The bed is auto-generated by scripts/generate_music.py (which reuses social-video's procedural NumPy synth) — no manual step. cinematic and ambient are fully implemented; other styles fall back to cinematic until upstream adds them. Full reference: references/music-pairing.md.

Aspect ratio export

Each commercial can be rendered in multiple aspect ratios in one run. The brief lists them in aspectRatios:

"aspectRatios": ["16:9", "9:16", "1:1"]

For each ratio, make_commercial.py:

1. Re-composes scenes (Runway clips may need re-crop/re-frame for vertical)
2. Adjusts text overlay positions for the canvas
3. Outputs output/commercial-16x9.mp4, output/commercial-9x16.mp4, output/commercial-1x1.mp4

The 16:9 master is the source of truth; 9:16 and 1:1 are smart-crop derivations (center-of-interest crop where possible, letterboxing only as last resort).

Cost estimate

Typical 60s commercial with 6 scenes:

Plus iteration buffer (likely 1-2 rounds of VO re-takes, maybe 1 Runway shot re-fire): ~$1-2 extra. Realistic: $3-5 per commercial.

Common iteration gotchas

1. VO timing mismatch — VO turns out shorter/longer than the scene duration. Fix: regenerate VO with adjusted text length, or stretch/shrink scene duration in the brief.
2. Runway prompt doesn't land — the AI generates something off. Fix: be more specific about composition ("close-up of hands typing" vs "person at computer"), add style modifiers ("shallow depth of field, slow push-in").
3. Music drowns the VO — bump music.volume down to 0.18, or increase the duck depth in make_commercial.py.
4. Aspect crop loses the subject — for 9:16 / 1:1, sometimes the auto-crop misses the focal point. Override per scene with explicit cropFocus: {x: 0.5, y: 0.4}.
5. CTA URL doesn't read — bump CTA font size, increase shadow contrast against the background.
6. Brand colors clash with AI footage — Runway output is wildly variable on color. Apply a brand-color grade pass in make_commercial.py so all scenes lean toward your brand palette.

When NOT to use this skill

• Pure screenshot-driven content with no AI B-roll → use social-video directly (cheaper, faster)
• Live performance / venue intro → use intro-trailer
• Talking-head with real footage → out of scope (just edit in DaVinci/Premiere)
• Music video / synced cuts to music → out of scope (this is silent-master under VO)

File structure

A commercial project should look like:

~/clientname-commercial/
├── brief/
│   └── brief.json                # the spec — single source of truth
├── assets/                       # user-provided
│   ├── logo.png
│   ├── screenshot1.png
│   ├── screenshot2.png
│   ├── team-photo.jpg
│   └── ...
├── generated/                    # AI-generated intermediates
│   ├── runway/                   # Runway video clips
│   │   ├── scene-01-hook.mp4
│   │   └── scene-04-broll.mp4
│   ├── vo/                       # ElevenLabs VO audio
│   │   ├── scene-02-vo.mp3
│   │   └── scene-03-vo.mp3
│   ├── music/
│   │   └── bed.mp3               # procedural music bed
│   ├── overlays/                 # PNG title cards, CTAs
│   │   └── cta-slide.png
│   └── scenes/                   # per-scene composited MP4s
│       ├── 01-hook.mp4
│       ├── 02-problem.mp4
│       └── ...
└── output/
    ├── commercial-16x9.mp4       # final deliverables
    ├── commercial-9x16.mp4
    └── commercial-1x1.mp4

References

• references/brief-schema.md — complete brief.json schema, every field documented
• references/voice-selection.md — voice ID library, gender/tone/use-case mapping
• references/music-pairing.md — music style → commercial type pairing rules
• references/workflow-hybrid.md — how Runway + ElevenLabs + ffmpeg interact, the orchestrator state machine

Triggers (auto-invocation)

This skill triggers when the user says:

• /ai-commercial, /commercial, /ad
• "make a commercial"
• "make an ad"
• "TV-style ad", "60-second spot", "promo ad"
• "commercial for [client]" / "ad for [brand]"
• "paid media ad", "Google Ads video", "Facebook ad video"