@jairodri/huckleberry-cli 1.0.2 → 1.1.0

package/skills/creative/modes/talking-head.md

@@ -0,0 +1,2208 @@
+# Talking Head Mode
+**depends-on: [hook-generator]**
+
+Generate talking head videos, presenter content, UGC-style testimonials, and lip-synced videos. This mode handles AI presenter generation, script-to-video pipelines, lip-sync workflows, and platform-optimized delivery for every format where a person speaks to camera.
+
+**Loaded by:** `creative/SKILL.md` when user selects mode 4 (Talking Head).
+**Model selection:** Handled by `brand/stack.md` — never hardcode model IDs in prompts or workflows.
+**Brand consistency:** Reads from `./brand/creative-kit.md` before every generation.
+**Presenter consistency:** Approved presenters saved to `./brand/creative-kit.md` for reuse across videos.
+**File output:** All deliverables saved to `./campaigns/{campaign}/video/talking-head/`
+
+---
+
+## Why This Mode Exists
+
+**The problem:** Talking head videos are the most persuasive content format but:
+1. Recording yourself is time-consuming and requires confidence
+2. Professional presenters are expensive ($500-5000+ per video)
+3. UGC creators charge $100-500 per post and may not match your brand
+4. Iterating on scripts means re-filming everything
+5. Scaling personalized video is nearly impossible manually
+6. Each video model has a different API interface, making multi-model comparison error-prone
+7. No single model wins every prompt — quality depends on the specific content
+8. FTC and platform rules on AI-generated presenters are evolving rapidly
+
+**The solution:** AI talking heads that:
+- Generate professional presenter videos in minutes
+- Let you iterate on scripts without re-recording
+- Create unlimited variants for A/B testing
+- Maintain consistent brand presenter identity
+- Scale personalized outreach cost-effectively
+- Reference MODEL_REGISTRY.md for all API payloads (never hardcodes model names)
+- Run parallel multi-model generation and let the user pick the winner
+- Include proper AI-generated content disclosures
+- Save approved presenters to the creative kit for cross-video reuse
+
+**The game-changer:** Combining avatar generation + lip-sync + TTS lets you:
+- Create a consistent "brand spokesperson"
+- Update any script without re-filming
+- Test multiple presenter styles quickly
+- Produce video content at 10x the speed
+- Generate voiceover in a consistent brand voice via TTS
+- Localize into multiple languages from a single base video
+
+---
+
+## Model Selection
+
+**Do NOT hardcode model IDs.** Always refer to `brand/stack.md` for the current default video model and its verified API payload.
+
+As of this writing, the models used in this mode are:
+
+### Video Generation Models
+
+| Role | Model | Registry Section | Estimated Cost (5s clip) |
+|------|-------|-----------------|--------------------------|
+| **Default** | Kling 2.5 Turbo Pro | Video Generation > Default Model | ~$0.40 |
+| **Comparison** | Google Veo 3.1 | Video Generation > Comparison Model: Google Veo 3.1 | ~$0.80-1.50 |
+| **Comparison** | OpenAI Sora 2 | Video Generation > Comparison Model: OpenAI Sora 2 | ~$0.60-1.20 |
+
+### Lip-Sync Model
+
+| Role | Model | Registry Section | Estimated Cost |
+|------|-------|-----------------|---------------|
+| **Lip-Sync** | Kling Lip-Sync | Lip-Sync > Model: Kling Lip-Sync | ~$0.30-0.60 |
+
+### How to Call
+
+1. Open `brand/stack.md`
+2. Find the **Video Generation** or **Lip-Sync** section
+3. Copy the verified payload structure for the desired model
+4. Insert your constructed prompt, source image, and parameters
+5. Execute the API call via your configured generation provider
+6. For hero content: run all three video models in parallel (see Parallel Multi-Model Generation below)
+
+### Cross-Model Parameter Cheat Sheet
+
+Every video model uses different parameter names for the same concept. Always consult MODEL_REGISTRY.md before writing any API call.
+
+| Concept | Kling 2.5 | Veo 3.1 | Sora 2 |
+|---------|-----------|---------|--------|
+| **Starting image** | `start_image` | `image` | `input_reference` |
+| **Duration** | `duration` (5, 10) | `duration` (4, 6, 8) | `seconds` (4-12) |
+| **Aspect ratio** | `aspect_ratio` ("16:9") | `aspect_ratio` ("16:9") | `aspect_ratio` ("landscape") |
+| **Prompt adherence** | `guidance_scale` (0-1) | — | — |
+| **Negative prompt** | `negative_prompt` | `negative_prompt` | — |
+| **Audio generation** | Not native | `generate_audio` | Native (always on) |
+| **Ending frame** | `end_image` | `last_frame` | — |
+| **Resolution control** | Fixed 1080p | `resolution` ("720p", "1080p") | Fixed |
+| **Reproducibility** | — | `seed` | — |
+
+### Model Strengths for Talking Head Content
+
+Refer to MODEL_REGISTRY.md for authoritative details. Summary for presenter-specific routing decisions:
+
+**Kling 2.5 Turbo Pro (Default for Talking Heads):**
+- Best for people and natural movement
+- Most realistic human faces
+- Handles casual movements well (ideal for UGC)
+- Best lip-sync ecosystem (Kling Lip-Sync pairs natively)
+- Strong at controlled expressions and gestures
+
+**Google Veo 3.1:**
+- Highest fidelity video output
+- Native audio generation (set `generate_audio: true`)
+- Good for establishing scenes with ambient audio
+- Slower generation, higher cost
+- No square aspect ratio support
+
+**OpenAI Sora 2:**
+- Strong prompt comprehension
+- Good at character-driven content
+- Native audio generation (always on)
+- Most variable generation times
+- No square aspect ratio support
+
+### When to Use Which — Talking Head Specific
+
+```
+FOR MAXIMUM REALISM (people quality):
+    → Kling 2.5 Turbo Pro (best faces, most natural movement)
+
+FOR SPEED + QUALITY BALANCE:
+    → Kling 2.5 Turbo Pro (fastest for people content)
+
+FOR BUILT-IN AUDIO:
+    → Veo 3.1 (generates audio with video)
+
+FOR UGC AUTHENTICITY:
+    → Kling 2.5 (handles casual movements well)
+
+FOR CORPORATE/FORMAL:
+    → Kling 2.5 or Sora 2 (cleaner, more controlled)
+
+FOR HERO/FLAGSHIP:
+    → Run all 3 in parallel, pick winner
+```
+
+---
+
+## Parallel Multi-Model Generation
+
+For hero content and any case where quality matters more than cost, run the same presenter prompt through all three video models simultaneously.
+
+### Why Parallel Beats Sequential
+
+- No single model wins every prompt — the best output varies by content
+- Running in parallel takes the same wall-clock time as the slowest model (~6 min)
+- Running sequentially would take ~15 minutes for all three
+- Total cost for a parallel comparison run: ~$2.20 (see Cost Awareness below)
+- Eliminates guessing — present all three, let the user pick
+
+### Parallel Execution Pattern
+
+Use task agents to fire all three API calls simultaneously:
+
+```
+PARALLEL GENERATION PLAN
+─────────────────────────────────────────────
+Task 1: Kling 2.5     → cost ~$0.40, time ~3min
+Task 2: Veo 3.1       → cost ~$1.00, time ~5min
+Task 3: Sora 2        → cost ~$0.80, time ~6min
+─────────────────────────────────────────────
+Total:                   ~$2.20, ~6min (parallel)
+```
+
+### Translating One Prompt to Three APIs
+
+Given a presenter prompt, the payload differs per model. Example for a 16:9 talking head T2V call:
+
+**Kling 2.5 payload (from MODEL_REGISTRY.md):**
+```json
+{
+  "model": "{{video_model_from_stack}}",
+  "input": {
+    "prompt": "{{presenter_prompt}}",
+    "duration": 5,
+    "aspect_ratio": "16:9",
+    "guidance_scale": 0.5
+  }
+}
+```
+
+**Veo 3.1 payload (from MODEL_REGISTRY.md):**
+```json
+{
+  "model": "google/veo-3.1",
+  "input": {
+    "prompt": "{{presenter_prompt}}",
+    "duration": 8,
+    "aspect_ratio": "16:9",
+    "resolution": "1080p",
+    "generate_audio": true
+  }
+}
+```
+
+**Sora 2 payload (from MODEL_REGISTRY.md):**
+```json
+{
+  "model": "{{comparison_video_model_from_stack}}",
+  "input": {
+    "prompt": "{{presenter_prompt}}",
+    "seconds": 8,
+    "aspect_ratio": "landscape"
+  }
+}
+```
+
+**Critical:** Always verify these payloads against MODEL_REGISTRY.md before execution. Parameter names change when models update.
+
+---
+
+## Cost Awareness
+
+Before generating, always estimate and communicate the cost to the user.
+
+### Per-Generation Estimates
+
+| Model | Duration | Estimated Cost | Typical Time |
+|-------|----------|---------------|--------------|
+| Kling 2.5 | 5s clip | $0.30-0.50 | 2-5min |
+| Kling 2.5 | 10s clip | $0.60-1.00 | 4-8min |
+| Veo 3.1 | 8s clip (720p) | $0.60-0.80 | 3-6min |
+| Veo 3.1 | 8s clip (1080p) | $1.00-1.50 | 5-8min |
+| Sora 2 | 8s clip | $0.60-1.20 | 3-10min |
+| Kling Lip-Sync | 2-10s clip | $0.30-0.60 | 1-3min |
+
+### Common Workflow Cost Estimates
+
+| Workflow | What You Get | Estimated Cost |
+|----------|-------------|---------------|
+| Single model presenter | 1 video | ~$0.40-1.20 |
+| Parallel comparison (3 models) | 3 videos to compare | ~$2.00-2.50 |
+| Style exploration (5 styles x 1 model) | 5 presenter approaches | ~$2.00-6.00 |
+| Presenter + lip-sync | 1 presenter video + 1 synced video | ~$0.70-1.80 |
+| Full pipeline (generate + lip-sync + 3 models) | 3 videos + lip-sync on winner | ~$2.50-3.10 |
+| UGC batch (5 variants, single model) | 5 different UGC videos | ~$2.00-5.00 |
+
+### Cost Communication Template
+
+Before executing, inform the user:
+
+```
+ESTIMATED GENERATION COST
+─────────────────────────
+Models: [list models]
+Clips: [number of clips]
+Duration per clip: [seconds]
+Lip-sync: [yes/no]
+Estimated total: ~$X.XX
+Estimated time: ~Xmin (parallel) / ~Xmin (sequential)
+
+Proceed? [Y/n]
+```
+
+---
+
+## Presenter Style Exploration (Before Generation)
+
+**Critical insight:** Don't generate with one style and hope it works. Explore genuinely DIFFERENT presenter styles first.
+
+### The Style Exploration Process
+
+**STEP 1: GENERATE 4-5 DIFFERENT PRESENTER STYLES**
+
+This is NOT: Same person with different clothes
+This IS: Fundamentally different presenter archetypes that each tell a different story
+
+```
+[YOUR BRAND] - Style Exploration
+
+Generate presenter concepts for these 5 directions:
+
+1. CORPORATE AUTHORITY
+   - Demographic: 35-50, professional appearance
+   - Setting: Modern office, corporate environment
+   - Wardrobe: Business professional, suit/blazer
+   - Energy: Confident, measured, authoritative
+   - Vibe: "Trust the expert"
+
+2. RELATABLE FRIEND
+   - Demographic: 25-40, approachable look
+   - Setting: Home office, kitchen, casual space
+   - Wardrobe: Smart casual, comfortable
+   - Energy: Warm, conversational, genuine
+   - Vibe: "Let me share what worked for me"
+
+3. ENERGETIC CREATOR
+   - Demographic: 22-35, creator aesthetic
+   - Setting: Ring light setup, content studio
+   - Wardrobe: Trendy casual, branded
+   - Energy: High, dynamic, enthusiastic
+   - Vibe: "You HAVE to try this"
+
+4. EXPERT EDUCATOR
+   - Demographic: 30-55, credible appearance
+   - Setting: Study, library, professional backdrop
+   - Wardrobe: Smart casual, glasses optional
+   - Energy: Calm, explanatory, helpful
+   - Vibe: "Let me explain how this works"
+
+5. LIFESTYLE ASPIRATIONAL
+   - Demographic: 28-45, aspirational look
+   - Setting: Beautiful home, travel location, luxury
+   - Wardrobe: Elevated casual, tasteful
+   - Energy: Relaxed confidence, success aura
+   - Vibe: "This is what my life looks like"
+```
+
+**STEP 2: IDENTIFY WINNER**
+
+After generating style exploration:
+```
+REVIEW each presenter style:
+
+Which presenter:
+- Best matches brand voice?
+- Would audience trust most?
+- Fits the content type?
+- Has right energy level?
+- Would work across multiple videos?
+
+WINNER: [Selected style]
+BECAUSE: [Why this style wins for this brand/use case]
+```
+
+**STEP 3: EXTRACT PRESENTER PRINCIPLES**
+
+Once winner identified:
+```
+WINNING STYLE EXTRACTION
+
+Demographics:
+- Age range: [X-X]
+- Gender: [if specific]
+- Ethnicity: [if specific]
+- Overall look: [descriptors]
+
+Environment:
+- Primary setting: [where they present from]
+- Background elements: [what's visible]
+- Lighting style: [natural/studio/mixed]
+
+Wardrobe:
+- Style: [formal/casual/etc.]
+- Colors: [palette]
+- Accessories: [if any]
+
+Delivery:
+- Energy level: [1-10]
+- Speaking pace: [slow/medium/fast]
+- Hand gestures: [minimal/moderate/expressive]
+- Eye contact: [direct to camera always]
+
+Audio:
+- Voice tone: [warm/authoritative/energetic]
+- Pacing: [conversational/punchy/measured]
+```
+
+**STEP 4: SAVE TO CREATIVE KIT**
+
+Once a presenter is approved, save the full specification to `./brand/creative-kit.md` so it can be reused across all future videos. See the Presenter Consistency System section below.
+
+**STEP 5: APPLY ACROSS CONTENT**
+
+Use extracted principles for:
+- All future videos maintain consistency
+- Same presenter = brand recognition
+- Variations in script, not in presenter
+
+---
+
+## Presenter Archetype Deep Dives
+
+### Corporate Authority
+
+**When to use:** B2B, financial services, healthcare, enterprise SaaS, professional services
+
+**Visual Formula:**
+```
+[Man/Woman] in [30s-50s], [silver/dark hair], wearing [tailored blazer/suit],
+in [modern glass office/conference room with city view], [warm professional lighting],
+[confident composed expression], [seated at desk OR standing with slight lean],
+[direct eye contact with camera], [subtle hand gestures], corporate executive style
+```
+
+**Setting Options:**
+- Corner office with city view
+- Modern conference room
+- Executive desk with minimal decor
+- Standing at presentation screen
+- Seated in designer chair
+
+**Wardrobe Options:**
+- Tailored navy blazer over white shirt
+- Grey suit, no tie (modern)
+- Classic suit with subtle tie
+- Blazer over turtleneck (thought leader)
+- Professional dress (solid colors)
+
+**Energy Markers:**
+- Measured pace
+- Deliberate movements
+- Confident pauses
+- Minimal but purposeful gestures
+- Assured vocal tone
+
+**TTS Voice Selection for Corporate Authority:**
+- Voice qualities: Deep, measured, confident, warm baritone or clear alto
+- ElevenLabs recommended voices: Professional male/female, mature tone
+- Pace: 0.9x-1.0x speed (slightly slower conveys authority)
+- Avoid: Overly energetic, breathy, or youthful voices
+
+---
+
+### Relatable Friend (UGC Style)
+
+**When to use:** DTC brands, consumer products, wellness, beauty, lifestyle
+
+**Visual Formula:**
+```
+Friendly [demographic] sitting in [casual setting], natural window light,
+holding/showing [product], genuine excited expression, talking directly to
+camera like filming a selfie video, authentic UGC testimonial style, casual
+comfortable body language, 5 seconds
+```
+
+**Setting Options:**
+- Bright kitchen counter
+- Cozy living room couch
+- Home office with plants
+- Bedroom getting-ready setup
+- Outdoor patio/balcony
+
+**Wardrobe Options:**
+- Cozy sweater/cardigan
+- Simple t-shirt
+- Casual button-down
+- Loungewear (if brand appropriate)
+- Athleisure
+
+**Energy Markers:**
+- Conversational rhythm
+- Natural pauses ("honestly?", "okay so...")
+- Expressive facial reactions
+- Genuine enthusiasm without over-selling
+- Relatable body language
+
+**UGC Script Patterns:**
+```
+DISCOVERY: "Okay so I found this [product] and I'm obsessed..."
+REVIEW: "So I've been using [product] for [time] and here's my honest take..."
+COMPARISON: "I used to use [old product] but then I tried [new product]..."
+TRANSFORMATION: "Before [product] I was [problem]. Now? [result]."
+```
+
+**UGC Authenticity Markers:**
+- Slightly imperfect framing
+- Natural lighting (not studio)
+- Casual wardrobe
+- Real reactions, not posed
+- Personal space as backdrop
+- Eye contact with camera
+
+**TTS Voice Selection for Relatable Friend:**
+- Voice qualities: Warm, conversational, natural inflection, mid-range
+- ElevenLabs recommended voices: Young adult, conversational tone
+- Pace: 1.0x-1.1x speed (natural conversational)
+- Avoid: Robotic, overly polished, or broadcasting voices
+
+---
+
+### Energetic Creator
+
+**When to use:** Gen-Z products, entertainment, gaming, trendy DTC, social apps
+
+**Visual Formula:**
+```
+[Young energetic creator] in [22-35], [colorful trendy outfit], in [content
+studio with ring light/neon lights], [bright dynamic lighting], [animated
+expressions], [lots of movement and gestures], [high energy delivery],
+[fast-paced enthusiastic style], YouTube/TikTok creator aesthetic
+```
+
+**Setting Options:**
+- Ring light setup visible
+- LED/neon accent lighting
+- Streaming/gaming setup
+- Colorful backdrop
+- Outdoor action setting
+
+**Wardrobe Options:**
+- Graphic tees
+- Bold colors
+- Branded merch
+- Trendy streetwear
+- Statement accessories
+
+**Energy Markers:**
+- Fast-paced delivery
+- Big expressions
+- Lots of hand movement
+- Pattern interrupts
+- Enthusiasm at 10
+
+**Creator Script Patterns:**
+```
+HOOK: "STOP scrolling. This is important."
+REVEAL: "I literally just discovered [thing] and I'm freaking out."
+CHALLENGE: "I bet you can't guess what [product] does."
+REACTION: "[reaction to trying product]... WAIT what?!"
+```
+
+**TTS Voice Selection for Energetic Creator:**
+- Voice qualities: Bright, energetic, youthful, expressive range
+- ElevenLabs recommended voices: Young, dynamic, high energy
+- Pace: 1.1x-1.3x speed (fast, punchy delivery)
+- Avoid: Flat, monotone, or overly mature voices
+
+---
+
+### Expert Educator
+
+**When to use:** Online courses, professional services, B2B explainers, tutorials
+
+**Visual Formula:**
+```
+[Knowledgeable expert] in [30s-55], [smart casual or academic style],
+in [home study/office with books/whiteboard], [balanced lighting],
+[thoughtful composed expression], [explaining with purposeful gestures],
+[patient instructive tone], educator/thought leader style
+```
+
+**Setting Options:**
+- Study with bookshelves
+- Office with credentials visible
+- Whiteboard/screen behind
+- Standing at presentation
+- Desk with relevant props
+
+**Wardrobe Options:**
+- Button-down shirt
+- Blazer over casual shirt
+- Sweater over collared shirt
+- Glasses (authority signal)
+- Minimal accessories
+
+**Energy Markers:**
+- Patient pace
+- Teaching rhythm
+- Logical structure
+- Illustrative gestures
+- "Here's what matters" moments
+
+**TTS Voice Selection for Expert Educator:**
+- Voice qualities: Clear, articulate, patient, explanatory cadence
+- ElevenLabs recommended voices: Mature professional, teacher tone
+- Pace: 0.9x-1.0x speed (allows comprehension, conveys thoughtfulness)
+- Avoid: Rushed, overly casual, or monotone voices
+
+---
+
+### Lifestyle Aspirational
+
+**When to use:** Luxury brands, high-ticket services, aspirational DTC, travel, real estate
+
+**Visual Formula:**
+```
+[Elegant successful person] in [30s-50s], [elevated casual attire],
+in [beautiful interior/scenic location], [golden hour OR designer lighting],
+[relaxed confident demeanor], [speaking with quiet confidence], [minimal
+but graceful movement], aspirational lifestyle aesthetic
+```
+
+**Setting Options:**
+- Designer living room
+- Travel location (balcony view)
+- Luxury car interior
+- High-end restaurant/hotel
+- Yacht/beach/resort
+
+**Wardrobe Options:**
+- Designer casual
+- Linen/natural fabrics
+- Neutral luxury palette
+- Subtle jewelry/watch
+- Effortlessly elegant
+
+**Energy Markers:**
+- Relaxed confidence
+- No rushing
+- "I have time" energy
+- Subtle smile
+- Quiet success vibes
+
+**TTS Voice Selection for Lifestyle Aspirational:**
+- Voice qualities: Smooth, unhurried, refined, warm confidence
+- ElevenLabs recommended voices: Sophisticated, calm delivery
+- Pace: 0.85x-0.95x speed (unhurried = luxury)
+- Avoid: Hyperactive, sharp, or salesy voices
+
+---
+
+## TTS Integration Guidance
+
+### Why External TTS Matters
+
+The Kling Lip-Sync model includes basic TTS via `text` + `voice_id`, but it has limited voice options and no voice consistency across sessions. For brand-level talking head content, external TTS gives you:
+
+- **Voice cloning** — match a specific voice consistently
+- **Voice selection** — hundreds of voices to match any archetype
+- **Multi-language** — generate the same script in any language
+- **Fine control** — adjust pace, emphasis, pauses, emotion
+- **Brand consistency** — same voice across every video, forever
+
+### ElevenLabs Integration (Recommended)
+
+ElevenLabs is the recommended TTS provider for talking head content due to voice quality, cloning capability, and extensive voice library.
+
+**Voice Selection Process:**
+```
+1. IDENTIFY archetype (Corporate, Friend, Creator, Educator, Aspirational)
+2. BROWSE ElevenLabs voice library filtering by:
+   - Gender matching presenter visual
+   - Age range matching presenter demographic
+   - Accent matching target market
+   - Energy level matching archetype
+3. TEST 3-5 candidate voices with a sample script
+4. SELECT winner based on:
+   - Natural fit with presenter visual
+   - Audience trust factor
+   - Clarity at speaking pace
+   - Emotional range
+5. SAVE voice_id to creative-kit.md alongside presenter spec
+```
+
+**Voice Selection Criteria Per Archetype:**
+
+| Archetype | Gender Match | Age Feel | Tone | Pace | Key Quality |
+|-----------|-------------|----------|------|------|-------------|
+| Corporate Authority | Match visual | 35-55 | Confident, warm | Moderate-slow | Trustworthy |
+| Relatable Friend | Match visual | 25-40 | Conversational | Natural | Authentic |
+| Energetic Creator | Match visual | 22-35 | Bright, dynamic | Fast | Exciting |
+| Expert Educator | Match visual | 30-55 | Clear, patient | Moderate | Credible |
+| Lifestyle Aspirational | Match visual | 30-50 | Smooth, refined | Slow | Aspirational |
+
+**ElevenLabs API Pattern:**
+```
+1. Generate audio from script text using selected voice
+2. Download audio file (MP3 or WAV)
+3. Pass audio_file to Kling Lip-Sync
+4. Result: presenter video with perfectly matched voiceover
+```
+
+**ElevenLabs Script Tips:**
+- Use SSML-style pauses: add "..." for natural breath breaks
+- Capitalize for emphasis: "This is REALLY important"
+- Short sentences sync better than long complex ones
+- Add punctuation for rhythm: periods = full pause, commas = brief pause
+- Test pronunciation of brand names and technical terms
+
+### Alternative TTS Providers
+
+| Provider | Strengths | Best For |
+|----------|-----------|----------|
+| ElevenLabs | Voice cloning, quality, library size | Brand voice, premium content |
+| OpenAI TTS | Fast, affordable, good quality | Quick prototypes, batch content |
+| Google Cloud TTS | Multi-language, SSML support | International campaigns |
+| Amazon Polly | Cost-effective at scale | High-volume personalization |
+| Azure Speech | Neural voices, SSML | Enterprise integrations |
+
+### TTS to Lip-Sync Workflow
+
+```
+SCRIPT → TTS (ElevenLabs) → AUDIO FILE → LIP-SYNC (Kling) → FINAL VIDEO
+
+Step 1: Write script (use duration calculator below)
+Step 2: Generate audio via TTS with selected voice
+Step 3: Download audio as MP3 or WAV
+Step 4: Pass audio_file + presenter video to Kling Lip-Sync
+Step 5: Review sync quality
+Step 6: Deliver or iterate
+```
+
+---
+
+## Script-to-Video Pipeline
+
+The complete pipeline from key message to finished talking head video.
+
+### Pipeline Overview
+
+```
+KEY MESSAGE
+    │
+    ▼
+SCRIPT WRITING (exact duration targeting)
+    │
+    ▼
+TTS AUDIO GENERATION (brand voice)
+    │
+    ▼
+PRESENTER VIDEO GENERATION (multi-model)
+    │
+    ▼
+LIP-SYNC APPLICATION
+    │
+    ▼
+FTC COMPLIANCE CHECK
+    │
+    ▼
+PLATFORM OPTIMIZATION
+    │
+    ▼
+FILE OUTPUT to ./campaigns/{campaign}/video/talking-head/
+```
+
+### Step 1: Key Message Extraction
+
+Start with the core message:
+```
+KEY MESSAGE BRIEF
+─────────────────
+Product/Service: [name]
+Core message: [one sentence]
+Target audience: [who]
+Desired action: [what they should do]
+Tone: [archetype reference]
+Platform: [where it will go]
+Duration target: [seconds]
+```
+
+### Step 2: Script Writing to Exact Duration
+
+Use the duration calculator to write a script that fits the target length precisely.
+
+**Duration Calculation:**
+
+| Word Count | Duration | Use Case |
+|------------|----------|----------|
+| 15 words | ~5 seconds | Social hook |
+| 30 words | ~10 seconds | Instagram Reel |
+| 45 words | ~15 seconds | TikTok optimal |
+| 60 words | ~20 seconds | Short testimonial |
+| 90 words | ~30 seconds | Product explainer |
+| 150 words | ~60 seconds | Full testimonial |
+
+**Rule:** ~150 words per minute at natural conversational pace
+
+**Script Writing Template:**
+```
+TARGET: [X] seconds = [Y] words
+
+SCRIPT DRAFT:
+───────────────────────────
+[Hook - first 1-3 seconds, ~5-10 words]
+[Value/Story - middle section]
+[CTA - final 3-5 seconds, ~10-15 words]
+───────────────────────────
+WORD COUNT: [actual count]
+ESTIMATED DURATION: [seconds]
+ADJUSTMENT NEEDED: [add/cut X words]
+```
+
+### Step 3: TTS Audio Generation
+
+Using the approved voice (saved in creative-kit.md):
+```
+1. Load approved voice_id from creative-kit.md
+2. Feed final script to TTS provider
+3. Set pace to archetype-appropriate speed
+4. Generate audio file
+5. Verify duration matches target (+/- 1 second)
+6. If duration is off: adjust script word count and regenerate
+```
+
+### Step 4: Presenter Video Generation
+
+Generate the presenter video using the approved presenter spec (from creative-kit.md):
+```
+1. Load presenter spec from creative-kit.md
+2. Construct presenter prompt from saved specification
+3. Generate via multi-model parallel (hero) or default model (standard)
+4. Duration should match or slightly exceed audio duration
+5. User selects preferred video output
+```
+
+### Step 5: Lip-Sync Application
+
+Apply the TTS audio to the selected presenter video:
+
+**Audio-driven lip-sync (from MODEL_REGISTRY.md):**
+```json
+{
+  "model": "{{lip_sync_model_from_stack}}",
+  "input": {
+    "video_url": "{{selected_presenter_video_url}}",
+    "audio_file": "{{tts_audio_url}}"
+  }
+}
+```
+
+### Step 6: Quality Review and Delivery
+
+See Quality Checklist section below.
+
+---
+
+## Script Structures
+
+### HOOK-VALUE-CTA (15-30 seconds)
+```
+Hook (0-3 sec): [Attention-grabber - question, statement, or pattern interrupt]
+Value (3-20 sec): [Main message, benefit, or story]
+CTA (20-30 sec): [Clear next step]
+```
+
+### PROBLEM-AGITATE-SOLVE (30-60 seconds)
+```
+Problem (0-10 sec): [Name the pain point]
+Agitate (10-30 sec): [Make them feel it]
+Solve (30-60 sec): [Present the solution + CTA]
+```
+
+### BEFORE-AFTER (15-30 seconds)
+```
+Before (0-10 sec): [Life before product/solution]
+After (10-25 sec): [Transformation/result]
+CTA (25-30 sec): [How to get same result]
+```
+
+### Tone Templates
+
+**Professional/Corporate:**
+```
+"[Name] here with [Company]. Today I want to share how [product/insight]
+can help you [achieve outcome]. Here's what you need to know..."
+```
+
+**Casual/UGC:**
+```
+"Okay so I've been using [product] for [time] and honestly? I'm obsessed.
+Here's why [specific benefit]. If you [problem], you need this."
+```
+
+**Expert/Educational:**
+```
+"One thing I see people get wrong about [topic] is [misconception].
+Here's what actually works: [insight]. Let me show you..."
+```
+
+**Energetic/Sales:**
+```
+"Stop what you're doing. [Product] just changed everything. I'm serious -
+[result] in [timeframe]. You HAVE to try this."
+```
+
+**Aspirational:**
+```
+"[Casual opening]. I wanted to share something that's completely transformed
+[area of life]. [Product] gave me [result]. Here's how it works..."
+```
+
+---
+
+## Lip-Sync Workflow
+
+For adding speech to existing videos using Kling Lip-Sync.
+
+### When to Use Lip-Sync
+
+- You have a great presenter video but need different script
+- Client wants to change messaging after video generation
+- Creating personalized versions of same base video
+- Adding voiceover to product demo videos
+- Dubbing content for different languages
+- Using external TTS for brand voice consistency
+
+### Audio-Driven Lip-Sync (Recommended)
+
+Use when you have a pre-recorded or TTS-generated audio file.
+
+**Payload (from MODEL_REGISTRY.md):**
+```json
+{
+  "model": "{{lip_sync_model_from_stack}}",
+  "input": {
+    "video_url": "{{source_video_url}}",
+    "audio_file": "{{audio_url}}"
+  }
+}
+```
+
+### Text-Driven Lip-Sync (Quick Prototyping)
+
+Use for quick tests when voice quality is not critical.
+
+**Payload (from MODEL_REGISTRY.md):**
+```json
+{
+  "model": "{{lip_sync_model_from_stack}}",
+  "input": {
+    "video_url": "{{source_video_url}}",
+    "text": "{{spoken_text}}",
+    "voice_id": "en_AOT",
+    "voice_speed": 1
+  }
+}
+```
+
+### When to Use Model TTS vs. External Audio
+
+| Scenario | Approach | Why |
+|----------|----------|-----|
+| Quick prototype | Use `text` + `voice_id` | Fastest path. One API call does everything. |
+| Brand-specific voice | External TTS (ElevenLabs) then `audio_file` | Model TTS voices are limited. External gives voice cloning. |
+| Voiceover already recorded | Use `audio_file` | You have the audio. No synthesis needed. |
+| Multiple languages | External TTS then `audio_file` | Model TTS language/accent options are limited. |
+| Testimonial/UGC content | Use `audio_file` with real voice recording | Authenticity matters. Real voices convert better. |
+| Script iteration/testing | Use `text` + `voice_id` | Fast turnaround for script comparison. |
+| Final production delivery | External TTS then `audio_file` | Maximum quality and voice control. |
+
+### Lip-Sync Quality Tips
+
+- Source video should have face visible 70%+ of time
+- Forward-facing shots work better than profiles
+- Avoid videos with heavy face movement/turning
+- Audio should be clear without background noise
+- Script pacing should match natural speech
+- Video must be 2-10 seconds (Kling Lip-Sync requirement)
+- Video resolution must be 720p-1080p
+- Audio file must be under 5MB
+- Keep sentences short (easier sync)
+
+---
+
+## Audio and Voice Considerations
+
+### When Using Veo 3.1 Native Audio
+
+**Strengths:**
+- Generates synchronized audio with video
+- Natural ambient sounds
+- Speech that matches lip movement
+- Good for establishing scenes
+
+**Limitations:**
+- Less control over specific script
+- Audio quality varies
+- May need post-processing
+
+### Voice-Over Tips
+
+**If recording your own VO for lip-sync:**
+```
+[ ] Record in quiet environment
+[ ] Use consistent distance from mic
+[ ] Match energy to presenter style
+[ ] Natural pauses between sentences
+[ ] Clear enunciation
+[ ] Export as MP3 or WAV
+```
+
+**If using TTS (text input via Kling built-in):**
+```
+[ ] Use punctuation for natural pauses
+[ ] Write phonetically for tricky words
+[ ] Keep sentences conversational length
+[ ] Test different phrasings
+[ ] Consider adding "..." for pauses
+```
+
+**If using external TTS (ElevenLabs recommended):**
+```
+[ ] Select voice matching presenter archetype
+[ ] Test voice with sample script before full generation
+[ ] Adjust pace to match archetype (see TTS Voice Selection tables)
+[ ] Export at highest quality (WAV preferred, MP3 acceptable)
+[ ] Verify audio duration matches video duration
+[ ] Save voice_id to creative-kit.md for future use
+```
+
+---
+
+## Presenter Consistency System
+
+Consistency is the key to building brand recognition with AI presenters. One approved presenter used across all videos creates the illusion of a real brand spokesperson.
+
+### Why Consistency Matters
+
+- **Brand recognition:** Viewers remember a face, not a product
+- **Trust building:** Same person appearing repeatedly builds familiarity
+- **Production efficiency:** No need to re-explore styles for every video
+- **Cross-campaign coherence:** All content looks like it comes from one source
+
+### Saving an Approved Presenter to Creative Kit
+
+After the style exploration process identifies a winner, save the complete specification to `./brand/creative-kit.md`:
+
+```markdown
+## Approved Presenter: [Name/Identifier]
+
+**Approved:** [date]
+**Archetype:** [Corporate Authority / Relatable Friend / Energetic Creator / Expert Educator / Lifestyle Aspirational]
+
+### Visual Specification
+- **Demographics:** [age range, gender, ethnicity, overall look]
+- **Setting:** [primary setting, background elements, lighting style]
+- **Wardrobe:** [style, colors, accessories]
+- **Energy:** [level 1-10, specific markers]
+- **Expression:** [default expression, range]
+- **Gestures:** [minimal/moderate/expressive, specific notes]
+
+### Voice Specification
+- **TTS Provider:** [ElevenLabs / OpenAI / etc.]
+- **Voice ID:** [voice_id string]
+- **Voice Name:** [human-readable name from provider]
+- **Pace:** [speed multiplier, e.g., 0.95x]
+- **Tone:** [warm/authoritative/energetic/etc.]
+
+### Generation Prompt (Reusable)
+```
+[The exact prompt that produced the approved presenter video.
+Copy this verbatim for all future videos with this presenter.
+Only change: duration and aspect ratio per platform.]
+```
+
+### Reference Videos
+- **Approval video URL:** [URL of the video that was approved]
+- **Lip-sync sample URL:** [URL of lip-synced version if available]
+
+### Usage Notes
+- [Any specific notes about what works/doesn't work]
+- [Platform-specific adjustments if needed]
+- [Known limitations or artifacts to watch for]
+```
+
+### Reusing an Approved Presenter
+
+When creating new talking head content:
+
+```
+1. CHECK creative-kit.md for approved presenter
+2. IF approved presenter exists:
+   a. Load generation prompt verbatim
+   b. Load voice_id and TTS settings
+   c. Only modify: duration, aspect_ratio, platform-specific adjustments
+   d. Skip style exploration entirely
+3. IF no approved presenter:
+   a. Run full style exploration process
+   b. Save winner to creative-kit.md
+   c. Then proceed with content generation
+```
+
+### Updating an Approved Presenter
+
+Presenters may need updating when:
+- Brand repositions or pivots
+- New campaign requires different energy
+- User explicitly requests a new look
+- Technical quality improves with new models
+
+**Update process:**
+```
+1. Run new style exploration alongside current presenter
+2. Compare new options vs. current presenter
+3. If new presenter wins: update creative-kit.md
+4. If current presenter still wins: keep as-is
+5. Archive old presenter spec (don't delete — may need for legacy content)
+```
+
+---
+
+## FTC Compliance Module
+
+AI-generated presenter videos are subject to evolving regulations. This section covers the current requirements and best practices for disclosure.
+
+### Why FTC Compliance Matters
+
+- **Legal requirement:** The FTC requires disclosure when AI-generated likenesses are used in advertising
+- **Platform enforcement:** Major platforms are implementing AI content labeling requirements
+- **Consumer trust:** Transparent disclosure builds trust rather than eroding it
+- **Future-proofing:** Regulations are tightening, not loosening — build compliance into the workflow now
+
+### Disclosure Requirements
+
+**FTC Guidelines (Current as of 2026):**
+- AI-generated presenters in advertising must be disclosed
+- Disclosure must be "clear and conspicuous"
+- Cannot be buried in fine print or hidden behind clicks
+- Must appear before or during the content, not only after
+
+**Required Disclosure Language:**
+```
+Options (use the one most natural for your format):
+
+1. "This video features an AI-generated presenter"
+2. "Presenter created with AI"
+3. "AI-generated spokesperson"
+4. "[Brand] AI presenter"
+```
+
+### Platform-Specific AI Content Policies
+
+**Meta (Facebook/Instagram):**
+- Requires AI-generated content label on ads
+- Use Meta's built-in "AI-generated" content label
+- Apply during ad setup in Ads Manager
+- Organic content: add text disclosure in caption
+
+**TikTok:**
+- Requires AI-generated content disclosure
+- Use TikTok's AI label feature when uploading
+- Add #AIgenerated or similar hashtag
+- Text overlay disclosure in first 3 seconds recommended
+
+**YouTube:**
+- Requires disclosure for "altered or synthetic" content
+- Use YouTube's AI content declaration in upload settings
+- For ads: disclose in ad setup and in video itself
+- Shorts: text overlay in first frame
+
+**LinkedIn:**
+- Requires transparency for AI-generated content
+- Disclose in post text
+- For video: text overlay or caption disclosure
+- Professional context demands higher transparency standard
+
+**General Web (Landing pages, email, etc.):**
+- FTC guidelines apply regardless of platform
+- Add disclosure near the video player
+- Include in video description or caption
+- Consider text overlay in video itself
+
+### Implementing Disclosure in Video
+
+**Method 1: Text Overlay (Recommended for all platforms)**
+```
+Add text overlay in post-production:
+- Position: Bottom-left or top-right corner
+- Timing: First 3-5 seconds of video
+- Style: Small, readable, non-intrusive
+- Text: "AI-generated presenter" or platform-appropriate variant
+- Font: Match brand typography
+- Opacity: 70-90% (visible but not distracting)
+```
+
+**Method 2: Verbal Disclosure (For longer content)**
+```
+Include in script:
+- Position: Within first 10 seconds
+- Style: Natural, not apologetic
+- Example: "I'm [Brand]'s AI presenter, and today I want to share..."
+- Tone: Confident — AI is a feature, not a flaw
+```
+
+**Method 3: Caption/Description (Supplementary)**
+```
+Add to video description/caption:
+- "This video features an AI-generated presenter"
+- Include relevant hashtags: #AIgenerated #AIpresenter
+- Link to brand's AI disclosure policy if applicable
+```
+
+### Compliance Checklist
+
+Before publishing any talking head video:
+```
+[ ] AI-generated presenter disclosure included in video (text overlay or verbal)
+[ ] Platform-specific AI content label applied (Meta, TikTok, YouTube, etc.)
+[ ] Video description/caption includes disclosure
+[ ] Disclosure appears within first 5 seconds
+[ ] Disclosure is clear and conspicuous (readable, not hidden)
+[ ] No claims of real person endorsement
+[ ] No impersonation of real individuals
+[ ] Testimonial claims are truthful (even with AI presenter)
+[ ] Check platform's current AI content policy (policies evolve)
+```
+
+### What NOT to Do
+
+- Do NOT impersonate real people (celebrities, influencers, competitors)
+- Do NOT claim AI presenter is a real customer giving a real testimonial
+- Do NOT hide or obscure the AI disclosure
+- Do NOT use AI presenter to make claims the brand cannot substantiate
+- Do NOT skip disclosure because "no one will know" — platforms are building detection
+- Do NOT assume today's rules will stay the same — check quarterly
+
+### Compliance-Aware Script Templates
+
+**UGC-Style with Disclosure:**
+```
+"Hi, I'm [Brand]'s AI presenter. Now that that's out of the way —
+let me tell you about something amazing. I've been programmed with
+real customer feedback, and here's what people are saying about [product]..."
+```
+
+**Corporate with Disclosure:**
+```
+"[Pause for text overlay: AI-generated presenter]
+Welcome. Today I want to walk you through how [product] can help you
+[achieve outcome]..."
+```
+
+**Creator-Style with Disclosure:**
+```
+"Before we dive in — full transparency, I'm AI-generated. But the results
+I'm about to show you? Those are 100% real. Let's go..."
+```
+
+---
+
+## Platform-Specific Optimization
+
+### TikTok/Reels (9:16)
+
+**Specs:**
+- Aspect Ratio: 9:16 (vertical)
+- Duration: 15-30 seconds optimal
+- Safe Zone: Keep face/text center 60%
+
+**Style Adjustments:**
+```
+→ Higher energy delivery
+→ Faster pacing
+→ Hook in first 1-2 seconds
+→ Pattern interrupts
+→ Jump cuts acceptable
+→ Casual/authentic feel
+```
+
+**Prompt Modifier:**
+```
+...[base prompt], filmed vertically like TikTok/Reels content,
+energetic creator style, direct eye contact with camera
+```
+
+**AI Disclosure:** Use platform's built-in AI label + text overlay in first 3 seconds
+
+### YouTube (16:9)
+
+**Specs:**
+- Aspect Ratio: 16:9 (landscape)
+- Duration: 30-120 seconds
+- Safe Zone: Standard letterbox
+
+**Style Adjustments:**
+```
+→ More measured pacing
+→ Can be longer form
+→ More professional setups accepted
+→ Room for B-roll integration
+→ Intro/outro structure
+```
+
+**Prompt Modifier:**
+```
+...[base prompt], widescreen YouTube style, professional yet engaging,
+room for graphics/lower thirds
+```
+
+**AI Disclosure:** Use YouTube's AI content declaration in upload settings + text overlay
+
+### LinkedIn (1:1 or 16:9)
+
+**Specs:**
+- Aspect Ratio: 1:1 (square) or 16:9
+- Duration: 30-60 seconds optimal
+- Tone: Professional but personal
+
+**Style Adjustments:**
+```
+→ Professional appearance
+→ Business-appropriate setting
+→ Thought leadership tone
+→ Value-first messaging
+→ Credibility signals
+```
+
+**Prompt Modifier:**
+```
+...[base prompt], professional LinkedIn style, credible expert appearance,
+business casual in modern office environment
+```
+
+**AI Disclosure:** Disclose in post text + text overlay in video
+
+### Instagram Stories (9:16)
+
+**Specs:**
+- Aspect Ratio: 9:16
+- Duration: 15 seconds max per segment
+- Ephemeral feel
+
+**Style Adjustments:**
+```
+→ Casual, in-the-moment feel
+→ Can be "rougher" quality
+→ Direct audience address
+→ Personal/behind-scenes vibe
+→ Clear single message per story
+```
+
+**AI Disclosure:** Use Meta's AI-generated content label + caption disclosure
+
+### Ads (Various)
+
+**Facebook/Instagram Ads:**
+- 1:1, 4:5, or 9:16
+- 15-30 second optimal
+- Hook in 0-3 seconds
+- Clear CTA
+- AI disclosure via Meta Ads Manager labeling
+
+**YouTube Ads:**
+- 16:9
+- 15-30 second (skippable) or 6 second (bumper)
+- Brand visible throughout
+- AI disclosure in ad setup + text overlay
+
+---
+
+## Use Cases Deep Dive
+
+### 1. Lip-Sync Overlay
+
+**Best for:** Adding voiceover to existing video, dubbing, personalization
+
+**Input Requirements:**
+- Video with visible face (front-facing works best)
+- Audio file (MP3, WAV) OR text script
+
+**Workflow:**
+```json
+{
+  "model": "{{lip_sync_model_from_stack}}",
+  "input": {
+    "video_url": "{{source_video_url}}",
+    "audio_file": "{{audio_url}}"
+  }
+}
+```
+
+**Or with text (uses built-in TTS):**
+```json
+{
+  "model": "{{lip_sync_model_from_stack}}",
+  "input": {
+    "video_url": "{{source_video_url}}",
+    "text": "{{spoken_text}}",
+    "voice_id": "en_AOT",
+    "voice_speed": 1
+  }
+}
+```
+
+---
+
+### 2. AI Presenter Generation
+
+**Best for:** Creating presenter content from scratch, brand spokesperson
+
+**Multi-Model Workflow (from MODEL_REGISTRY.md):**
+
+```json
+// Kling 2.5
+{
+  "model": "{{video_model_from_stack}}",
+  "input": {
+    "prompt": "[presenter prompt]",
+    "aspect_ratio": "16:9",
+    "duration": 5,
+    "guidance_scale": 0.5
+  }
+}
+```
+
+```json
+// Veo 3.1 (with native audio)
+{
+  "model": "google/veo-3.1",
+  "input": {
+    "prompt": "[presenter prompt]",
+    "aspect_ratio": "16:9",
+    "duration": 8,
+    "generate_audio": true,
+    "resolution": "1080p"
+  }
+}
+```
+
+```json
+// Sora 2
+{
+  "model": "{{comparison_video_model_from_stack}}",
+  "input": {
+    "prompt": "[presenter prompt]",
+    "seconds": 8,
+    "aspect_ratio": "landscape"
+  }
+}
+```
+
+**Then add lip-sync if specific script needed (from MODEL_REGISTRY.md):**
+```json
+{
+  "model": "{{lip_sync_model_from_stack}}",
+  "input": {
+    "video_url": "[generated video URL]",
+    "audio_file": "[TTS audio URL]"
+  }
+}
+```
+
+---
+
+### 3. UGC-Style Content
+
+**Best for:** Authentic testimonials, product reviews, social proof
+
+**The UGC Formula:**
+```
+[Relatable person] + [Casual setting] + [Natural lighting] +
+[Authentic delivery] + [Genuine reaction] = Believable UGC
+```
+
+**Prompt Template:**
+```
+Friendly [demographic] sitting in [casual setting], natural window light,
+holding/showing [product], genuine excited expression, talking directly to
+camera like filming a selfie video, authentic UGC testimonial style, casual
+comfortable body language, 5 seconds
+```
+
+---
+
+### 4. Personal Brand Series
+
+**Best for:** Thought leaders, course creators, coaches, consultants
+
+**Consistency Formula:**
+```
+ESTABLISH ONCE, USE FOREVER:
+- Same presenter appearance (saved in creative-kit.md)
+- Same setting/background
+- Same wardrobe style
+- Same energy level
+- Same lighting setup
+- Same TTS voice (voice_id saved in creative-kit.md)
+
+Only change: Script and specific content
+```
+
+**Series Prompt Template:**
+```
+[Consistent presenter description from creative-kit.md], [same setting],
+[same lighting], [same wardrobe style], [same energy], discussing [new topic],
+[consistent delivery style], 5 seconds
+```
+
+---
+
+### 5. Multi-Language Dubbing
+
+**Best for:** International campaigns, localized content, global brands
+
+**Workflow:**
+```
+1. Generate presenter video once (approve visual)
+2. Write scripts in each target language
+3. Generate TTS audio per language using ElevenLabs multi-language voices
+4. Run Kling Lip-Sync for each language version
+5. Result: Same presenter, multiple languages
+```
+
+**Cost advantage:** One presenter video generation (~$0.40-1.20) + one lip-sync per language (~$0.30-0.60 each) = dramatically cheaper than filming multiple takes.
+
+---
+
+## Execution Workflow
+
+### Step 1: Clarify Requirements
+
+Before generating:
+```
+[ ] What's the use case? (UGC, corporate, educational, etc.)
+[ ] What platform? (TikTok, YouTube, LinkedIn, ads)
+[ ] What aspect ratio? (9:16, 16:9, 1:1)
+[ ] What duration? (and word count)
+[ ] What presenter style? (see archetypes)
+[ ] What's the script/message?
+[ ] Need lip-sync to specific audio?
+[ ] Is there an approved presenter in creative-kit.md?
+[ ] Budget constraints?
+```
+
+### Step 2: Style Selection
+
+If not predefined and no approved presenter in creative-kit.md:
+```
+[ ] Generate style exploration with 4-5 different presenter styles
+[ ] Present options to user
+[ ] Extract principles from winner
+[ ] Save to creative-kit.md for future reuse
+[ ] Document voice selection alongside visual spec
+```
+
+If approved presenter exists in creative-kit.md:
+```
+[ ] Load presenter spec from creative-kit.md
+[ ] Load voice_id from creative-kit.md
+[ ] Skip to Step 3
+```
+
+### Step 3: Script Writing
+
+```
+[ ] Extract key message
+[ ] Calculate word count for target duration
+[ ] Write script using appropriate structure (Hook-Value-CTA, PAS, Before-After)
+[ ] Apply tone template matching archetype
+[ ] Review word count against duration target
+[ ] Finalize script
+```
+
+### Step 4: Estimate Cost and Confirm
+
+```
+ESTIMATED GENERATION COST
+─────────────────────────
+Models: [list models]
+Clips: [number of clips]
+Duration per clip: [seconds]
+TTS generation: [included/separate]
+Lip-sync: [yes/no]
+Estimated total: ~$X.XX
+Estimated time: ~Xmin
+
+Proceed? [Y/n]
+```
+
+### Step 5: Construct Prompt
+
+Use this formula:
+```
+[PRESENTER DESCRIPTION] + [SETTING] + [LIGHTING] +
+[EXPRESSION/ENERGY] + [ACTION] + [STYLE MODIFIER] + [DURATION]
+```
+
+### Step 6: Multi-Model Generation
+
+```
+Run same prompt through:
+1. Kling 2.5 (~3min) — default, best for faces
+2. Veo 3.1 (~5min) — with audio, highest fidelity
+3. Sora 2 (~6min) — strong prompt adherence
+
+Present all three to user for selection.
+```
+
+For standard content (not hero), use default model only (Kling 2.5).
+
+### Step 7: TTS Audio Generation
+
+If lip-sync is needed:
+```
+1. Load voice_id from creative-kit.md
+2. Generate audio from final script via TTS
+3. Verify audio duration matches video duration
+4. Download audio file
+```
+
+### Step 8: Add Lip-Sync (If Needed)
+
+If specific script delivery required:
+```
+1. User approves video from Step 6
+2. Run through Kling Lip-Sync with TTS audio
+3. Input: selected video + audio file
+4. Output: synced talking head
+```
+
+**Payload (from MODEL_REGISTRY.md):**
+```json
+{
+  "model": "{{lip_sync_model_from_stack}}",
+  "input": {
+    "video_url": "[approved video URL]",
+    "audio_file": "[TTS audio URL]"
+  }
+}
+```
+
+### Step 9: FTC Compliance Check
+
+```
+[ ] AI disclosure text overlay added (or planned for post-production)
+[ ] Platform-specific AI label will be applied on upload
+[ ] Script does not make unsubstantiated claims
+[ ] No impersonation of real individuals
+[ ] Description/caption disclosure prepared
+```
+
+### Step 10: Deliver and Iterate
+
+```markdown
+## Talking Head Video Delivered
+
+**Style:** [Archetype used]
+**Platform:** [Target platform]
+**Duration:** [X seconds]
+**Presenter:** [Name from creative-kit.md or "New — pending approval"]
+
+### Selected Video
+[video URL]
+**Model:** [which model — from MODEL_REGISTRY.md]
+**Estimated Cost:** ~$[amount]
+
+### Lip-Sync Version (if applicable)
+[video URL]
+**Script:** "[excerpt...]"
+**Voice:** [voice name/ID]
+
+### FTC Compliance
+- AI disclosure: [included/pending post-production]
+- Platform label: [to be applied on upload]
+
+**Saved to:** ./campaigns/{campaign}/video/talking-head/[filename].mp4
+
+**Options:**
+- [ ] Approve and use
+- [ ] Adjust script and resync
+- [ ] Try different presenter style
+- [ ] Save presenter to creative-kit.md
+```
+
+---
+
+## File Output
+
+All generated talking head assets are saved to the campaign directory.
+
+### Output Directory Structure
+
+```
+./campaigns/{campaign}/video/talking-head/
+├── presenter-exploration/
+│   ├── corporate-authority-kling-v1.mp4
+│   ├── relatable-friend-kling-v1.mp4
+│   ├── energetic-creator-kling-v1.mp4
+│   ├── expert-educator-kling-v1.mp4
+│   └── lifestyle-aspirational-kling-v1.mp4
+├── generation/
+│   ├── presenter-kling-v1.mp4
+│   ├── presenter-veo-v1.mp4
+│   ├── presenter-sora-v1.mp4
+│   └── presenter-kling-v2.mp4       (iteration)
+├── audio/
+│   ├── script-v1-elevenlabs.mp3
+│   ├── script-v2-elevenlabs.mp3
+│   └── script-v1-spanish.mp3        (localization)
+├── lip-sync/
+│   ├── presenter-synced-v1.mp4
+│   ├── presenter-synced-v2.mp4
+│   └── presenter-synced-spanish.mp4  (localization)
+└── approved/
+    ├── final-talking-head.mp4        (selected winner)
+    └── final-talking-head-spanish.mp4
+```
+
+### File Naming Convention
+
+```
+{purpose}-{model}-v{version}.mp4
+
+Examples:
+  presenter-kling-v1.mp4
+  ugc-testimonial-kling-v1.mp4
+  corporate-intro-veo-v1.mp4
+  product-explainer-sora-v1.mp4
+  presenter-synced-v1.mp4            (lip-synced version)
+  presenter-synced-spanish-v1.mp4    (localized lip-sync)
+```
+
+### Saving Deliverables
+
+After the user approves a video:
+1. Download the video from the generation URL
+2. Save to `./campaigns/{campaign}/video/talking-head/generation/` with the naming convention
+3. If lip-synced: save synced version to `./campaigns/{campaign}/video/talking-head/lip-sync/`
+4. Copy the approved final version to `./campaigns/{campaign}/video/talking-head/approved/`
+5. Log the generation metadata (model, prompt, cost, timestamp, voice_id)
+6. If new presenter approved: save spec to `./brand/creative-kit.md`
+
+---
+
+## Quality Checklist
+
+### Technical Quality
+- [ ] Face clearly visible throughout
+- [ ] No uncanny valley artifacts
+- [ ] Consistent appearance (no morphing)
+- [ ] Smooth natural movement
+- [ ] Appropriate resolution for platform
+
+### Presenter Quality
+- [ ] Matches intended archetype
+- [ ] Expression appropriate for message
+- [ ] Energy level fits content type
+- [ ] Wardrobe matches brand/context
+- [ ] Setting supports message
+- [ ] Consistent with creative-kit.md specification (if approved presenter exists)
+
+### Lip-Sync Quality (if applicable)
+- [ ] Mouth movement matches audio
+- [ ] Natural speech rhythm
+- [ ] No obvious desync
+- [ ] Head movement doesn't break sync
+- [ ] Audio quality clear
+
+### Content Quality
+- [ ] Script delivered clearly
+- [ ] Pacing appropriate for platform
+- [ ] Hook captures attention
+- [ ] Message comes through
+- [ ] CTA clear (if applicable)
+
+### Voice Quality (if TTS used)
+- [ ] Voice matches presenter archetype
+- [ ] Pace feels natural for the content
+- [ ] Pronunciation correct (especially brand names)
+- [ ] Emotional tone matches script intent
+- [ ] No robotic artifacts or unnatural pauses
+
+### FTC Compliance Quality
+- [ ] AI disclosure present and visible
+- [ ] Disclosure appears within first 5 seconds
+- [ ] Platform-specific AI label will be applied
+- [ ] No misleading claims about real person endorsement
+- [ ] Description/caption disclosure prepared
+
+### Platform Fit
+- [ ] Correct aspect ratio for destination platform
+- [ ] Duration within platform optimal range
+- [ ] Safe zones respected (mobile UI overlays)
+- [ ] Thumbnail/first frame compelling
+- [ ] Works with platform compression
+
+---
+
+## Common Issues and Solutions
+
+| Issue | Cause | Solution |
+|-------|-------|----------|
+| Uncanny valley feel | Model limitations | Use Kling v2.5 for most realistic faces |
+| Face morphing mid-video | Long duration | Keep videos shorter (5-10 sec), extend with cuts |
+| Lip-sync drift | Audio/video mismatch | Use shorter scripts, clear enunciation |
+| Wrong energy level | Prompt too vague | Be explicit about energy: "calm" vs "enthusiastic" |
+| Generic stock presenter | No specific direction | Add detailed demographic and style descriptors |
+| Setting doesn't match | Prompt conflict | Prioritize setting description, remove conflicts |
+| Awkward hand movement | Unspecified gestures | Add gesture direction or specify "minimal movement" |
+| Bad lighting | Missing lighting prompt | Always include lighting: "warm natural light" |
+| Doesn't look like brand | No style consistency | Save presenter spec to creative-kit.md and reuse |
+| Audio quality poor | Model TTS limitations | Use external TTS (ElevenLabs) instead of text input |
+| Voice doesn't match visual | Wrong TTS voice selected | Re-select voice per archetype guidelines |
+| Lip-sync rejected | Video too long/short | Trim video to 2-10 seconds per Kling Lip-Sync requirements |
+| Audio file too large | Uncompressed audio | Compress to MP3 under 5MB |
+| Side profile sync fails | Non-frontal face | Use front-facing presenter videos for lip-sync |
+| Script too long for duration | Word count miscalculation | Recalculate: ~150 words per minute |
+| API parameter error | Wrong param name for model | Check cross-model cheat sheet in MODEL_REGISTRY.md |
+| Aspect ratio rejected | Model doesn't support it | Check MODEL_REGISTRY.md — Veo/Sora have no square support |
+| Generation timeout | Model overloaded | Retry; allow extra buffer for Sora 2 |
+| Presenter inconsistent across videos | Not using saved spec | Load presenter prompt from creative-kit.md |
+
+---
+
+## Iteration Strategies
+
+### When Presenter Is Close But Not Right
+
+**Problem:** Almost there but something's off
+**Strategy:** Targeted prompt adjustments
+
+```
+Wrong energy → adjust energy descriptors: "calm measured" vs "enthusiastic dynamic"
+Wrong age → adjust age range: "early 30s" vs "late 40s"
+Wrong setting → change environment descriptors
+Wrong wardrobe → update clothing description
+Expression off → add explicit expression: "warm genuine smile" vs "confident neutral"
+```
+
+### When Presenter Is Completely Wrong
+
+**Problem:** Output doesn't match intent
+**Strategy:** Different approach entirely
+
+Don't iterate on broken foundation:
+1. Try different archetype entirely
+2. Try different model
+3. Simplify prompt dramatically
+4. Check if Kling v2.5 gives better face quality
+
+### When Lip-Sync Quality Is Poor
+
+**Problem:** Audio doesn't match mouth movement
+**Strategy:** Technical fixes
+
+```
+- Use shorter scripts (fewer words = easier sync)
+- Ensure audio has clear enunciation
+- Use front-facing video (not profiles)
+- Try generating a new presenter video with more frontal positioning
+- Use recorded audio instead of model TTS for better quality
+- Verify video is 2-10 seconds
+- Check audio file is under 5MB
+```
+
+### When Voice Doesn't Match
+
+**Problem:** TTS voice feels wrong for the presenter
+**Strategy:** Voice re-selection
+
+```
+1. Review archetype voice guidelines (see TTS Voice Selection tables)
+2. Test 3-5 alternative voices from ElevenLabs library
+3. Prioritize: natural match to visual > quality > cost
+4. Re-generate lip-sync with new voice
+5. Update voice_id in creative-kit.md when satisfied
+```
+
+### When Budget Is a Concern
+
+**Problem:** Need to minimize generation costs
+**Strategy:** Targeted model selection
+
+```
+- Use Kling 2.5 only (cheapest at ~$0.40/clip)
+- Use model TTS (text input) instead of external TTS for prototyping
+- Skip parallel comparison for non-hero content
+- Reuse approved presenter video for multiple scripts via lip-sync
+- Plan scripts carefully before generating (avoid wasted generations)
+```
+
+---
+
+## Output Format
+
+### Style Exploration Output
+```markdown
+## Presenter Style Exploration
+
+**Brand/Project:** [Name]
+**Use Case:** [What videos will be used for]
+**Platform:** [Target platform]
+
+### Style 1: Corporate Authority
+[video URL]
+- Demographic: [specifics]
+- Setting: [description]
+- Energy: [level]
+- Saved to: ./campaigns/{campaign}/video/talking-head/presenter-exploration/
+
+### Style 2: Relatable Friend
+[video URL]
+- Demographic: [specifics]
+- Setting: [description]
+- Energy: [level]
+
+### Style 3: Energetic Creator
+[video URL]
+- Demographic: [specifics]
+- Setting: [description]
+- Energy: [level]
+
+### Style 4: Expert Educator
+[video URL]
+- Demographic: [specifics]
+- Setting: [description]
+- Energy: [level]
+
+### Style 5: Lifestyle Aspirational
+[video URL]
+- Demographic: [specifics]
+- Setting: [description]
+- Energy: [level]
+
+**Recommendation:** Style [X] best fits because [reasons]
+**Feedback needed:** Which direction resonates?
+**Next step:** Approve a style to save to creative-kit.md
+```
+
+### Generated Video Output
+```markdown
+## Talking Head Video Generated
+
+**Style:** [Archetype]
+**Platform:** [Target]
+**Duration:** [X seconds]
+**Presenter:** [From creative-kit.md or "New"]
+
+### Model Outputs:
+
+**Kling 2.5:** [URL] (~$0.40)
+**Veo 3.1:** [URL] (includes audio) (~$1.00)
+**Sora 2:** [URL] (~$0.80)
+
+**Total generation cost:** ~$2.20
+
+**Prompt Used:**
+> [full prompt for reference]
+
+**Next Steps:**
+- [ ] Select preferred video
+- [ ] Generate TTS audio for lip-sync
+- [ ] Add lip-sync to specific script
+- [ ] Request variation
+- [ ] Approve for use
+- [ ] Save presenter to creative-kit.md
+```
+
+### Lip-Sync Output
+```markdown
+## Lip-Sync Video Delivered
+
+**Source Video:** [URL]
+**Script:** "[excerpt...]"
+**Duration:** [X seconds]
+**Voice:** [voice name, voice_id]
+**TTS Provider:** [ElevenLabs / Model built-in / Recorded]
+
+**Final Video:** [URL]
+**Estimated Cost:** ~$[lip-sync cost]
+**Saved to:** ./campaigns/{campaign}/video/talking-head/lip-sync/[filename].mp4
+
+**Quality Check:**
+- Sync accuracy: [pass/fail]
+- Natural rhythm: [pass/fail]
+- Audio clarity: [pass/fail]
+- Expression match: [pass/fail]
+
+**FTC Compliance:**
+- AI disclosure: [included/pending]
+- Platform label: [to apply on upload]
+
+**Options:**
+- [ ] Approve and use
+- [ ] Adjust script and resync
+- [ ] Try different source video
+- [ ] Try different voice
+```
+
+---
+
+## Pipeline Integration
+
+```
+TALKING HEAD PIPELINE
+
+┌─────────────────────────────────────────┐
+│  Request arrives (direct or routed)     │
+│  → Clarify: platform, duration, style   │
+│  → Determine: generation vs lip-sync    │
+│  → Check creative-kit.md for presenter  │
+└─────────────────────────────────────────┘
+                    │
+        ┌───────────┴───────────┐
+        ▼                       ▼
+┌──────────────────┐   ┌──────────────────┐
+│  Style Undefined │   │  Style Defined   │
+│  (no approved    │   │  (approved       │
+│   presenter in   │   │   presenter in   │
+│   creative-kit)  │   │   creative-kit)  │
+│  → Run style     │   │  → Load spec     │
+│    exploration   │   │  → Skip to       │
+│  → Save winner   │   │    generation    │
+└──────────────────┘   └──────────────────┘
+                    │
+                    ▼
+┌─────────────────────────────────────────┐
+│  Script Writing                         │
+│  → Key message extraction               │
+│  → Duration-targeted word count         │
+│  → Tone matching archetype              │
+└─────────────────────────────────────────┘
+                    │
+                    ▼
+┌─────────────────────────────────────────┐
+│  Cost Estimation                        │
+│  → Calculate based on models + clips    │
+│  → Include TTS + lip-sync costs         │
+│  → Present to user for confirmation     │
+└─────────────────────────────────────────┘
+                    │
+                    ▼
+┌─────────────────────────────────────────┐
+│  talking-head mode (THIS MODE)          │
+│  → Multi-model generation               │
+│  → Present options                      │
+│  → TTS audio generation                 │
+│  → Add lip-sync                         │
+│  → FTC compliance check                 │
+│  → Quality check                        │
+│  → Save to campaigns/{campaign}/        │
+│    video/talking-head/                  │
+└─────────────────────────────────────────┘
+                    │
+                    ▼
+┌─────────────────────────────────────────┐
+│  Delivery                               │
+│  → Platform-optimized output            │
+│  → AI disclosure included               │
+│  → Ready for ads/social/content         │
+│  → Presenter saved to creative-kit.md   │
+└─────────────────────────────────────────┘
+```
+
+---
+
+## Handoff Protocols
+
+### Receiving from creative workflow
+```yaml
+Receive:
+  use_case: "talking head" | "UGC" | "presenter" | "lip-sync"
+  campaign: "[campaign name]"
+  platform: "[target platform]"
+  aspect_ratio: "[ratio]"
+  duration: "[seconds]"
+  style: "[archetype or custom]"
+  script: "[text]"
+  audio_url: "[if lip-sync with audio]"
+  video_url: "[if lip-sync to existing]"
+  budget_limit: "[if specified]"
+```
+
+### Returning to Workflow
+```yaml
+Return:
+  status: "complete" | "needs_selection" | "needs_iteration"
+  deliverables:
+    - url: "[video URL]"
+      local_path: "./campaigns/{campaign}/video/talking-head/[filename].mp4"
+      model: "[which model — from MODEL_REGISTRY.md]"
+      has_audio: true | false
+      has_lip_sync: true | false
+      duration: "[seconds]"
+      voice_id: "[if TTS used]"
+      estimated_cost: "[cost]"
+  presenter_saved: true | false
+  ftc_disclosure: "included" | "pending_post_production"
+  total_cost: "[sum of all generations]"
+  feedback_needed: "[any questions]"
+```
+
+### Receiving Video from product-video mode
+```yaml
+Receive for lip-sync:
+  video_url: "[product video URL]"
+  local_path: "./campaigns/{campaign}/video/[filename].mp4"
+  aspect_ratio: "[ratio]"
+  script: "[voiceover text]"
+  audio_url: "[optional, if pre-recorded]"
+```
+
+### Routing to product-video mode
+```yaml
+Route to product-video mode:
+  reason: "Need product footage before adding presenter overlay"
+  product: "[product name]"
+  video_style: "[motion style from product-video mode]"
+  return_to: "talking-head mode for voiceover/lip-sync"
+```
+
+---
+
+## Tips from Experience
+
+### What Works
+
+1. **Consistency beats variety** — Same presenter across videos builds recognition
+2. **Kling v2.5 for faces** — Most realistic human generation
+3. **Shorter is safer** — 5-10 second clips avoid quality degradation
+4. **Explicit energy levels** — "calm and measured" vs "enthusiastic and dynamic"
+5. **Multi-model approach** — Always generate with 2-3 models for hero content, let user pick
+6. **Lip-sync extends value** — One good video can become many scripts
+7. **External TTS for production** — ElevenLabs voice quality far exceeds model built-in TTS
+8. **Save everything to creative-kit.md** — Presenter spec + voice_id = instant reuse
+9. **FTC disclosure builds trust** — Be transparent about AI, audiences respect honesty
+10. **Duration-target your scripts** — Calculate word count before writing, not after
+
+### What Doesn't Work
+
+1. **Vague presenter description** — "A person talking" = generic results
+2. **Long continuous takes** — Quality degrades after 10-15 seconds
+3. **Ignoring setting** — Presenter without context looks artificial
+4. **Skipping style exploration** — First idea rarely best for brand
+5. **Mismatched energy** — Corporate script + UGC style = awkward
+6. **Complex movements** — Walking + talking + gesturing = artifacts
+7. **Model TTS for final delivery** — Built-in TTS voices are limited and inconsistent
+8. **Skipping FTC disclosure** — Platforms are building AI detection, regulations are tightening
+9. **Regenerating presenter for every video** — Save and reuse from creative-kit.md
+10. **Hardcoding model IDs** — Always use MODEL_REGISTRY.md
+
+### The 80/20
+
+80% of talking head success comes from:
+1. Clear presenter archetype selection
+2. Matching energy to platform
+3. Short, punchy scripts at exact word count
+4. Using Kling v2.5 for realism
+
+Get these four right, and you'll get good results.
+
+### Multi-Model Strategy
+
+When to invest in parallel comparison:
+- **Hero content** (website presenters, launch videos) — Always run all three
+- **Social content** (feed posts, stories) — Default model is fine
+- **UGC batch** (multiple testimonial variants) — Default model, iterate on scripts
+- **Ad creative** (paid campaigns) — Run all three, A/B test the winners
+- **Personalized outreach** (sales videos) — Default model + lip-sync
+
+### Budget Management
+
+```
+BUDGET TIERS
+─────────────────────────────────
+Lean:     1 model, no lip-sync           ~$0.40
+Standard: 1 model + lip-sync            ~$0.70-1.00
+Premium:  3 models, lip-sync on winner   ~$2.50-3.10
+Hero:     3 models + exploration + sync  ~$4.50-6.00
+Campaign: 5 variants + 3 languages       ~$5.00-10.00+
+```
+
+---
+
+## Quick Reference
+
+| Task | Model | Process |
+|------|-------|---------|
+| Generate presenter video | All 3 models | Multi-model parallel, user picks |
+| Add speech to existing video | Kling Lip-Sync | Direct, ~1min |
+| Presenter + specific script | Generate → TTS → Lip-Sync | Three-step pipeline |
+| Video with built-in audio | Veo 3.1 | Single generation |
+| Most realistic face | Kling v2.5 | Single or multi-model |
+| Fastest generation | Kling v2.5 | Single generation |
+| UGC style | Kling v2.5 | Handles casual movement best |
+| Brand voice consistency | ElevenLabs TTS → Lip-Sync | External TTS, saved voice_id |
+| Multi-language dubbing | Kling Lip-Sync (per language) | One video + multiple audio files |
+| Quick script test | Kling Lip-Sync with `text` | Built-in TTS, fastest path |
+| Reuse approved presenter | Load from creative-kit.md | Skip exploration, direct generation |
+
+---
+
+## Appendix: Model API Quick Reference
+
+**Always verify against MODEL_REGISTRY.md before executing.** This appendix is a convenience reference only.
+
+### Video Generation Call Patterns
+
+**Kling 2.5 T2V (Presenter Generation):**
+```
+Model: {{default_video_model from brand/stack.md}}
+Duration param: duration (5 or 10)
+Ratio param: aspect_ratio ("16:9", "9:16", "1:1")
+Extras: guidance_scale (0-1), negative_prompt
+```
+
+**Veo 3.1 T2V (Presenter with Audio):**
+```
+Model: google/veo-3.1
+Duration param: duration (4, 6, or 8)
+Ratio param: aspect_ratio ("16:9", "9:16")
+Extras: generate_audio (bool), resolution ("720p", "1080p"), negative_prompt, seed
+```
+
+**Sora 2 T2V (Presenter Generation):**
+```
+Model: {{comparison_video_model from brand/stack.md}}
+Duration param: seconds (4-12)
+Ratio param: aspect_ratio ("landscape", "portrait")
+Extras: openai_api_key (optional)
+```
+
+### Lip-Sync Call Patterns
+
+**Kling Lip-Sync (Audio-Driven):**
+```
+Model: {{lip_sync_model from brand/stack.md}}
+Video param: video_url (2-10s, 720p-1080p, <100MB)
+Audio param: audio_file (.mp3, .wav, .m4a, .aac, <5MB)
+```
+
+**Kling Lip-Sync (Text-Driven):**
+```
+Model: {{lip_sync_model from brand/stack.md}}
+Video param: video_url (2-10s, 720p-1080p, <100MB)
+Text param: text (free text)
+Voice param: voice_id (e.g., "en_AOT")
+Speed param: voice_speed (0.5-2.0, default 1)
+```
+
+**Important:** `audio_file` and `text` are mutually exclusive. `video_url` and `video_id` are mutually exclusive.
+
+---
+
+*This mode is part of the Vibe Marketing Skills v2 creative engine. Model configurations are maintained in `brand/stack.md`. When model schemas change upstream, MODEL_REGISTRY.md is updated first, then this mode file is updated to match.*