---
name: content-generator
description: Multi-provider automated video content generation platform. Generates TikTok 9:16 vertical videos (1 minute) from text prompts using NVIDIA NIM + BytePlus Seedance + FFmpeg. Implements Larry Playbook viral formula. Use when creating TikTok content, product videos, or any AI-generated video.
persona:
  name: "MrBeast (Jimmy Donaldson)"
  title: "Master of Viral Content"
  expertise: ["retention optimization", "thumbnail psychology", "pacing mastery", "audience psychology"]
  philosophy: "Every second matters. If they're not entertained, they leave. Make every frame count."
  credentials:
    - "300+ million YouTube subscribers across channels"
    - "Pioneered high-production challenge videos with massive budgets"
    - "Average 100M+ views per video through retention optimization"
    - "Built Feastables to $100M+ revenue through content-driven marketing"
  principles:
    - "Hook in 3 seconds - grab attention immediately or lose them forever"
    - "Pacing is everything - cut dead air, maintain momentum relentlessly"
    - "Thumbnails sell clicks - invest in visual psychology, test everything"
    - "Retention over length - 8 minutes at 80% beats 20 minutes at 40%"
    - "Scale creates spectacle - bigger stakes, bigger emotions, bigger views"
    - "Data drives decisions - A/B test titles, thumbnails, hooks constantly"
    - "Reinvest everything - compound growth by putting revenue back into content"
---

# Content Generator Skill

End-to-end AI video pipeline: LLM hook → NVIDIA image → BytePlus Seedance video → FFmpeg loop/compress.

## Quick Start

### Python API
```python
from scripts.generator import ContentGenerator

gen = ContentGenerator()
result = await gen.generate(
    concept="landlord_kitchen",   # Larry Playbook preset
    platform="tiktok",
    ratio="9:16",
    target_duration=60,           # 1 minute
)
print(result["video"])   # /path/to/final.mp4
print(result["hook"])    # "My landlord said I can't change anything..."
print(result["caption"]) # Story-style caption with hashtags
```

### CLI Script (TikTok viral)
```bash
cd skills/1ai-skills/content/content-generator

# TikTok viral (Larry Playbook formula, 1 min, 9:16)
python3 scripts/generate_tiktok_viral.py --niche motivation --ratio 9:16

# Full generator
python3 scripts/generator.py --concept landlord_kitchen --duration 60
```

---

## Provider Status (Verified 2026-02-27)

| Provider | Type | Status | Endpoint |
|----------|------|--------|----------|
| **NVIDIA NIM** | Image | ✅ WORKING | `ai.api.nvidia.com/v1/genai/` |
| **BytePlus Seedance** | Video | ⚠️ BLOCKED | `ark.ap-southeast.bytepluses.com/api/v3` |
| **NVIDIA LLM** | LLM | ✅ WORKING | `integrate.api.nvidia.com/v1/chat/completions` |
| Groq | LLM | ✅ WORKING | `api.groq.com/openai/v1` |
| Ollama Cloud | LLM | ⚠️ PENDING | `api.ollama.com` (key set, not verified) |
| XAI | Video | ❌ DISABLED | Credits exhausted |
| **Vast.ai Serverless** | Video | ⚠️ NEEDS $5 | ComfyUI on Vast.ai GPU fleet |
| **Runware** | Video | ⚠️ NEEDS KEY | Unified API — 300K+ models, $0.14/clip |

### Critical API Notes

**NVIDIA Image:**
- Endpoint: `POST https://ai.api.nvidia.com/v1/genai/{provider}/{model}`
- Working model: `black-forest-labs/flux.1-dev`
- Payload: `{"prompt": "..."}` ONLY — no `num_images`, no null fields
- Response: `artifacts[0].base64` (JPEG)

**BytePlus Seedance (⚠️ BLOCKED — Safe Experience Mode):**
- Base URL: `https://ark.ap-southeast.bytepluses.com/api/v3`
- Create task: `POST /contents/generations/tasks`
- Poll task: `GET /contents/generations/tasks/{task_id}`
- Payload: `{"model": "...", "content": [{"type": "text", "text": "..."}], "ratio": "9:16"}`
- ⚠️ Do NOT include `resolution` param — causes HTTP 400 on lite model
- Generation time: ~20s (lite), ~60s (pro)

**Runware (⚠️ SETUP REQUIRED):**
- Sign up: https://runware.ai
- API key: dashboard → API Keys
- SDK: `pip install runware` (WebSocket-based)
- Models: `runware:100@1` (Seedance Lite $0.14/clip), `runware:100@2` (Seedance Pro)
- Supports T2V, I2V, V2V via unified interface
- Default timeout: 480s (8 min for video)

**Vast.ai Serverless (⚠️ NEEDS $5 TOP-UP):**
- Deploy ComfyUI template → Serverless → Create Endpoint at cloud.vast.ai
- Get serverless API key from `vastai show endpoints`
- Set env vars: `VASTAI_SERVERLESS_KEY` + `VASTAI_ENDPOINT_NAME` (e.g. `ozeap3mc`)
- Uses `vastai-sdk` Python package — auto-installs if not present
- Supports T2V and I2V via ComfyUI workflow
- Worker auto-provisions GPU on-demand, pay-per-use

---

## Larry Playbook Viral Formula

Based on Oliver Henry's proven results: **234K views single post, 500K+ total, $588 MRR**.

### The Hook Formula
```
[Third-party person's problem] + [Doubt/conflict]
→ "Showed them what AI thinks..."
→ They reacted / changed their mind
```

### Why It Works
- Creates curiosity (what happened?)
- Third-party = relatable (not self-promo)
- AI result = concrete solution
- Triggers action (show YOUR landlord/mum/friend!)

### Hook Templates by Performance

| Hook Type | Views | Template |
|-----------|-------|----------|
| Landlord + AI | 234K avg | "My landlord said {constraint}, so I showed {them} what AI thinks {space} could look like" |
| Parent + AI | 80K avg | "My mum was skeptical about AI until I showed her {result} for our {room}" |
| Roommate + AI | 60K avg | "My flatmate thinks {X} is impossible, so I proved them wrong with AI {result}" |

### ❌ What Kills Virality
- "I built an app that does X" → Self-promotion
- "Check out my new feature Y" → Feature-focused
- "Download now for Z" → Direct CTA
- Not exactly 6 slides → Wrong count
- Different rooms across slides → Must be same room

### Caption Formula (Story Style)
```
[Hook context — 1 line]
My [relationship] [reaction/emotion] when I showed them [AI result]
[Subtle CTA — never pushy]
[max 5 hashtags]
```
Max 200 characters. Natural tone. NOT marketing language.

---

## TikTok 9:16 Pipeline (1-Minute Video)

```
LLM hook+prompt → BytePlus Seedance 5s clip → FFmpeg loop×12 = 60s → Compress CRF28 → ~8.6MB MP4
```

**Step by step:**
1. `generate_content(concept)` — LLM generates hook + video prompt + caption
2. `generate_video(prompt, ratio="9:16")` — BytePlus creates 5s portrait video
3. `loop_to_minute(clip, output, secs=60, loops=12)` — FFmpeg -stream_loop 12
4. `compress_video(looped, final, crf=28)` — 44MB → 8.6MB

**Format specs:**
- Ratio: 9:16
- Resolution: 704×1248 (Seedance lite output)
- FPS: 24
- Duration: 60 seconds
- Filesize: ~8.6MB (Telegram safe: < 16MB)
- Codec: H.264

**FFmpeg loop command:**
```bash
ffmpeg -stream_loop 12 -i input.mp4 -c copy -t 60 -y output_60s.mp4
```

**FFmpeg compress command:**
```bash
ffmpeg -i input.mp4 -c:v libx264 -crf 28 -preset fast -c:a copy -y output_compressed.mp4
```

---

## Preset Concepts (Larry Playbook)

| Concept | Hook | Best For |
|---------|------|---------|
| `landlord_kitchen` | "Landlord won't let me renovate..." | Interior design niche |
| `parent_bedroom` | "Mum was skeptical about AI..." | Home transformation |
| `motivation` | "Your only competition is yesterday's you" | Lifestyle / motivational |
| `money` | "Money follows action not wishes" | Finance / entrepreneur |
| `product` | Premium product showcase | E-commerce |

---

## Platform Specs

| Platform | Ratio | Resolution | FPS | Duration | Seedance Model |
|----------|-------|------------|-----|----------|----------------|
| **TikTok** | 9:16 | 1080×1920 | 24 | 60s | lite-t2v |
| YouTube Shorts | 9:16 | 1080×1920 | 30 | 60s | lite-t2v |
| Instagram Reels | 9:16 | 1080×1920 | 30 | 60s | lite-t2v |
| Facebook | 16:9 | 1920×1080 | 30 | any | lite-t2v |

---

## Strategies

| Strategy | LLM | Image | Video | Use When |
|----------|-----|-------|-------|----------|
| `fast` | NVIDIA | skip | BytePlus lite | Quick drafts |
| `quality` | Groq | NVIDIA Flux | BytePlus pro | Final content |
| `cheap` | Ollama | skip | BytePlus lite | Batch testing |
| `balanced` | Groq | NVIDIA Flux | BytePlus lite | Default |

---

## Environment Variables

```bash
export NVIDIA_API_KEY="nvapi-..."        # Required for image + LLM
export BYTEPLUS_API_KEY="..."            # Video (currently blocked)
export GROQ_API_KEY="gsk_..."            # Optional (faster LLM)
export OLLAMA_API_KEY="..."              # Optional (cheap LLM)
export VASTAI_SERVERLESS_KEY="..."       # Optional — Vast.ai serverless API key
export VASTAI_ENDPOINT_NAME="ozeap3mc"   # Optional — Vast.ai serverless endpoint name
export RUNWARE_API_KEY="rw_..."          # Optional — Runware unified video API
```

---

## File Structure

```
content/content-generator/
├── SKILL.md                           ← This file
├── config.yaml                        ← Provider config (verified endpoints)
├── STATUS.md                          ← Provider test results
├── scripts/
│   ├── generator.py                   ← Main ContentGenerator class ← START HERE
│   ├── generate_tiktok_viral.py       ← TikTok 9:16 1-min CLI script
│   ├── cli.py                         ← General CLI
│   ├── providers/
│   │   ├── base.py                    ← AIProvider abstract base class
│   │   ├── nvidia.py                  ← NVIDIA image gen
│   │   ├── byteplus.py                ← BytePlus Seedance (T2V/I2V)
│   │   ├── vastai.py                  ← Vast.ai serverless ComfyUI (T2V/I2V) ← NEW
│   │   ├── groq.py                    ← Groq LLM
│   │   ├── ollama.py                  ← Ollama cloud LLM
│   │   ├── xai.py                     ← XAI (disabled)
│   │   ├── fallback.py                ← FallbackChain + FallbackManager
│   └── ffmpeg_editor.py              ← FFmpeg wrapper
└── references/
    ├── providers/
    │   ├── nvidia.md                  ← NVIDIA API docs
    │   └── byteplus.md                ← BytePlus API docs
    └── workflow.md                    ← Workflow guide
```

---

## Dependencies

- Python 3.11+
- FFmpeg with libx264 (for loop + compress)
  - Path: `/home/linuxbrew/.linuxbrew/bin/ffmpeg`
  - Note: NOT built with libfreetype — no text overlay
- urllib (built-in)
- asyncio (built-in)
- ssl (built-in)

---

## Cost Estimates

| Operation | Provider | Cost |
|-----------|----------|------|
| Image generation | NVIDIA Flux | ~$0.004 |
| Video 5s clip | BytePlus Seedance lite | ~$0.026 |
| LLM storyboard | NVIDIA/Groq | ~$0.001 |
| **Full TikTok 1-min** | All combined | **~$0.031** |

---

## Advanced Content Operations (From Reference Libraries)

### 1. Content Quality Gate

**Strategy:** Score content with expert panel before publishing

```python
def content_quality_gate(content, content_type):
    """
    Multi-dimensional content quality scoring
    """
    expert_panel = {
        "clarity_expert": score_clarity(content),
        "engagement_expert": score_engagement_potential(content),
        "seo_expert": score_seo_optimization(content),
        "brand_expert": score_brand_alignment(content),
        "audience_expert": score_audience_fit(content)
    }
    
    scores = {}
    for expert, score in expert_panel.items():
        scores[expert] = score
    
    composite_score = sum(scores.values()) / len(scores)
    
    return {
        "composite_score": composite_score,
        "dimension_scores": scores,
        "status": "APPROVED" if composite_score >= 90 else "NEEDS_REVISION",
        "feedback": generate_improvement_suggestions(scores)
    }
```

**Quality Thresholds:**
```
90-100: EXCELLENT - Publish immediately
80-89:  GOOD - Minor improvements suggested
70-79:  ACCEPTABLE - Revision recommended
<70:    POOR - Significant rework required
```

### 2. Autoresearch for Content

**Strategy:** Karpathy-inspired optimization loops for conversion content

```python
def autoresearch_content(variants, target_metric):
    """
    Generate and test multiple content variants
    """
    # Generate 50+ variants
    all_variants = []
    for template in content_templates:
        for hook in hook_library:
            for cta in cta_library:
                variant = generate_variant(template, hook, cta)
                all_variants.append(variant)
    
    # Expert panel scoring
    scored_variants = []
    for variant in all_variants:
        score = content_quality_gate(variant, "short_form")
        scored_variants.append({
            "variant": variant,
            "score": score["composite_score"]
        })
    
    # Select top performers for A/B test
    top_variants = sorted(scored_variants, 
                         key=lambda x: x["score"], 
                         reverse=True)[:10]
    
    return top_variants
```

### 3. Multi-Platform Content Adaptation

```python
def adapt_content_for_platforms(base_content):
    """
    Repurpose one content piece for multiple platforms
    """
    adaptations = {
        "tiktok": {
            "format": "9:16_video",
            "duration": 60,
            "hook_style": "fast_paced_visual",
            "caption_length": 200,
            "hashtag_count": 5
        },
        "instagram_reels": {
            "format": "9:16_video",
            "duration": 60,
            "hook_style": "trending_audio",
            "caption_length": 150,
            "hashtag_count": 10
        },
        "youtube_shorts": {
            "format": "9:16_video",
            "duration": 60,
            "hook_style": "story_driven",
            "title_length": 60,
            "description_length": 300
        },
        "twitter": {
            "format": "carousel",
            "slides": 6,
            "hook_style": "thread_hook",
            "text_length": 280,
            "alt_text": True
        },
        "linkedin": {
            "format": "long_form_post",
            "style": "professional_storytelling",
            "length": 1300,
            "cta": "comment_engagement"
        }
    }
    
    return adaptations
```

### 4. Content Performance Prediction

```python
def predict_content_performance(content, platform):
    """
    Predict viral potential before publishing
    """
    features = extract_content_features(content)
    
    prediction_model = {
        "hook_strength": score_hook(features["hook"]),
        "visual_quality": score_visuals(features["media"]),
        "engagement_triggers": count_triggers(features["cta"]),
        "timing_factor": score_posting_time(content.schedule),
        "audience_match": calculate_audience_fit(content, platform)
    }
    
    viral_probability = calculate_viral_score(prediction_model)
    
    return {
        "viral_probability": viral_probability,
        "estimated_reach": predict_reach(viral_probability, platform),
        "estimated_engagement": predict_engagement(viral_probability),
        "improvement_suggestions": suggest_optimizations(prediction_model)
    }
```

### 5. Content Calendar Intelligence

```python
def intelligent_content_calendar(goals, resources):
    """
    AI-powered content calendar with strategic timing
    """
    calendar = []
    
    # Identify optimal posting windows
    best_times = analyze_audience_activity(platform="all")
    
    # Map content to business goals
    for goal in goals:
        content_needed = calculate_content_requirements(goal)
        
        for week in range(52):
            slots = find_optimal_slots(
                week, 
                goal.priority, 
                best_times
            )
            
            calendar.append({
                "week": week,
                "goal": goal.name,
                "content_type": goal.content_type,
                "platform": select_best_platform(goal),
                "timing": slots,
                "expected_impact": predict_impact(goal, slots)
            })
    
    return optimize_calendar(calendar, resources)
```

## Content Analytics Integration

### Attribution and ROI Tracking

```python
def track_content_roi(content_pieces, conversions):
    """
    Attribute conversions to content pieces
    """
    attribution = {}
    
    for content in content_pieces:
        touchpoints = get_user_touchpoints(conversions)
        
        # Multi-touch attribution
        content_conversions = calculate_attribution(
            touchpoints, 
            model="time_decay"
        )
        
        attribution[content.id] = {
            "content_id": content.id,
            "conversions": content_conversions,
            "revenue": content_conversions * avg_conversion_value,
            "cost": content.production_cost,
            "roi": (content_conversions * avg_conversion_value - content.production_cost) / content.production_cost
        }
    
    return attribution
```

## Output Format

```yaml
content_operations_report:
  content_id: "CNT-2025-089"
  platform: "tiktok"
  
  quality_scores:
    composite: 92
    clarity: 95
    engagement: 90
    seo: 88
    brand: 94
    status: "EXCELLENT"
    
  viral_prediction:
    probability: 0.78
    estimated_reach: "500K-1M"
    estimated_engagement: "45K-90K"
    confidence: "HIGH"
    
  platform_adaptations:
    tiktok:
      format: "9:16_video"
      duration: 60
      hook: "Landlord constraint + AI solution"
      
    instagram:
      format: "9:16_video"
      trending_audio: "Recommended"
      
    youtube_shorts:
      format: "9:16_video"
      title: "I showed my landlord AI could do this"
      
  posting_strategy:
    optimal_time: "Tuesday 7PM EST"
    hashtags: ["#AITransformation", "#InteriorDesign", "#Tech", "#BeforeAndAfter", "#HomeDecor"]
    caption: "My landlord didn't believe in AI..."
    
  performance_forecast:
    week_1: "100K views"
    week_2: "300K views"
    month_1: "750K views"
    
  related_content:
    - "Follow-up: Kitchen transformation"
    - "Series: Room makeovers"
    - "Behind the scenes: AI tools"
```

## Integration Points

**Cross-Skill Dependencies**
- `marketing/growth-engine` - For experiment tracking
- `marketing/seo-optimizer` - For discoverability
- `marketing/social-media-upload` - For distribution
- `marketing/analytics-dashboard` - For performance tracking

**Tool Integrations**
- Content Management: Notion, Airtable
- Scheduling: Buffer, Hootsuite, Later
- Analytics: Native platform APIs
- Creative: Canva, Adobe Creative Suite

---

## See Also

- `larry-playbook/SKILL.md` — Full viral formula + confidence system
- `tiktok-automation/SKILL.md` — Browser-based TikTok posting
- `social-media-upload/SKILL.md` — Multi-platform upload
- `humanizer/SKILL.md` — Make captions sound natural
- `marketing/content-ops` — Content operations workflows