@jairodri/huckleberry-cli 1.0.2 → 1.1.0

package/skills/hook-generator/SKILL.md

@@ -0,0 +1,438 @@
+---
+name: hook-generator
+description: Generate platform-optimized hooks using Unified Hook Generator v3.0 (Volume generation + Jev ranking + Ralph). 3-step system - volume generation produces 50 hooks (85-90/100), jev_jev_rerank ranks all 50 against ICP + funnel stage + voice (sub-second), Ralph enhances top hooks to 90-95/100. Uses examples-first training with real engagement data across all 6 platforms (Twitter, LinkedIn, Instagram, YouTube, TikTok, Medium).
+allowed-tools: Read, Grep, Glob, Bash
+license: Internal use only
+---
+
+# Unified Hook Generator v3.0
+**Status**: PRODUCTION-READY (upgraded 2026-01-23)
+**Quality**: 90-95/100 (v3.0 with Ralph enhancement)
+
+## Awareness Gate (MANDATORY)
+
+Before generating hooks, determine the target audience's awareness level. See `references/awareness-level-guide.md`. If unclear, ask the user. The awareness level determines which hook patterns will resonate — problem-aware audiences respond to validation hooks, solution-aware to comparison hooks, product-aware to proof hooks.
+
+## Quick Start
+```bash
+
+# Generate Twitter hooks (auto-detects niche from brand name)
+~/clawd/skills/hook-generator-openclaw/scripts/generate.sh \
+ --content "Building AI agents with OpenClaw" \
+ --platform twitter \
+ --brand aiwithjai \
+ --count 3
+
+# Generate hooks for Enchantents (auto-detects glamping niche)
+--content "Glamping tent ROI" \
+ --brand enchantents \
+
+# Explicit niche override
+node ~/clawd/skills/hook-generator-openclaw/scripts/generate-hooks-unified.js \
+ --content "Your content" \
+ --platform tiktok \
+ --brand default \
+ --niche enchantents-glamping
+```
+
+**Parameters:**
+- `--content` (required): The content/idea to generate hooks for
+- `--platform` (required): twitter, linkedin, instagram, youtube, tiktok, medium
+- `--brand` (optional): Brand name for context + auto-niche detection, defaults to "default"
+- `--niche` (optional): Explicit niche filter. Overrides brand-based detection
+- `--count` (optional): Number of hooks to generate, default 3
+
+**Brand → Niche Auto-Detection:**
+- enchantents, nomadixgear, nomadix: `enchantents-glamping`
+- aiwithjai, linkedgraphix, clawver: `ai-agents`
+- (other/default): all examples, no filtering
+
+## Overview
+Single authoritative system for generating platform-optimized hooks across all 6 platforms using **examples-first training** instead of abstract formulas.
+
+**Key Innovation**: Show AI 5-7 PROVEN hooks with real engagement data before generation → Consistent 85/100+ quality vs 76-81/100 baseline.
+
+## When to Invoke
+Use this skill when:
+- **Creating hooks for social content** (platform-content-generator Phase 2)
+- **Generating newsletter subject lines** (newsletter-strategist)
+- **Writing blog post titles** (blog-workflow)
+- **A/B testing hooks** for engagement optimization
+- **Repurposing content** across platforms with consistent quality
+
+## 4-Step Model Pipeline (v3.2 — two-pass generation, 2026-10-17)
+
+> **Dead model warning (verified 2026-10-17):** this skill previously pinned `z-ai/glm-4.7-flash` (Step 1) and `z-ai/glm-4.7` (Step 3). Both now return **HTTP 200 with EMPTY content**. OpenRouter credits are NOT the problem — `anthropic/claude-haiku-4.5`, `google/gemini-2.5-flash-lite` and `openai/gpt-4o-mini` all respond normally on the same key. Repoint any pin before running, or the step silently produces nothing.
+
+### Step 1a: Volume Generation (examples-first)
+**Model**: `anthropic/claude-haiku-4.5` via OpenRouter (verified working 2026-10-17)
+**Why**: Fast, cheap, optimized for volume
+**Output**: 40-50 hooks — a relevance floor and ICP coverage
+
+### Step 1b: Diversity Generation (REQUIRED — added 2026-10-17)
+**Model**: same as 1a, **different prompt**
+**Why**: the examples-first pass alone cannot produce a non-obvious hook. Measured: its genericness *floor* was 0.71 across 10 candidates — structurally zero hooks an average AI wouldn't write, so no selection step can find a tail that was never generated.
+**Output**: 40 more hooks from a prompt that widens the distribution:
+- Audience described as **ONE named individual** — age, business, the exact tool they bought, the phrase they used to dismiss it, a recurring scene from their week. Never a segment.
+- Banned structures: `I watched`, `I saw`, `Every <X> you do manually`, `Here is why`, `Most people`, `Imagine`, `X just Y` news framing, numbered lists.
+- Forced angle variety: second-person accusation, specific scene, question, confession, fragment, comparison to physical work the reader already understands.
+- Explicit instruction: hooks "an average AI would almost never produce".
+
+**Measured effect of adding 1b**: genericness floor 0.71 → 0.14, distribution range widened 4.1×, top-of-range held. On a named-person ICP query the examples-first batch scored R 0.36-0.67 (every one below the 0.70 ship floor) while the diversity batch took the entire top 6.
+
+### Step 2: Hook Selection
+**REPLACED 2026-09-26, EXTENDED 2026-10-17:** selection runs through the `jev-hook-gate` skill — TWO `jev_jev_rerank` calls over the pooled candidates (relevance R + genericness G), selected on `R − λ·G` with a 0.70 relevance floor. The old qwen/Sonnet/Gemini selection subagents are retired — do NOT use them.
+**Why**: TypeSafe judgment model, typed scores, no wasted reasoning tokens, no rate limits.
+**Output**: ranked list with both scores; best `R − λ·G` ships, next 2 kept as A/B alternates.
+See: `jev-hook-gate` SKILL.md (gate rule: global AGENTS.md §1b-s).
+
+### Step 3: Ralph Enhancement
+**Model**: a stronger model than the volume pass — **verify the pin resolves before relying on it** (see dead model warning above)
+**Why**: Consistent scoring, stable multi-step reasoning
+**Output**: 3 hooks at 90-95/100 (+7-10 points)
+
+**Fallback Pattern**: If Step 1 fails, retry with a different working model — never the dead GLM pins
+**Cost**: ~$0.25-0.35 per batch (80-90 hooks → 3 final hooks)
+
+## Architecture
+
+### Four-Step System (v3.2 — two-pass generation)
+Input: Content + Platform + Brand + NAMED-PERSON ICP
+ ↓
+Step 1a: Volume Generation — examples-first        → 40-50 hooks
+ ├─ Show 5-7 proven hooks from the platform database
+ └─ Purpose: relevance floor + ICP coverage
+Step 1b: Diversity Generation — different prompt    → 40 hooks   [REQUIRED]
+ ├─ Audience = one named individual, never a segment
+ ├─ Template openers banned; angle variety forced
+ └─ Purpose: make a tail exist so Step 2 can find it
+Step 2: Jev Gate (jev_jev_rerank MCP — 2 calls, since 2026-09-26 + 2026-10-17)
+ ├─ 3a relevance R — against the NAMED PERSON, not a segment
+ ├─ 3b genericness G — inverted query, "what would every AI say"
+ └─ Score = R − λ·G  (λ = 0.50), floor R ≥ 0.70; top 3 kept
+Step 3: Ralph Enhancement
+ ├─ Score top 3 (Ralph methodology)
+ ├─ Generate improvement prompts
+ ├─ Iterative enhancement (max 3 iterations)
+ └─ Early stop if no progress
+Output: 3 hooks at 90-95/100 (vs 85-88/100 v2.0)
+
+**Quality Improvement**: +7-10 points per hook
+**Cost**: ~$0.50-0.70 per batch (80-90 hooks + 2 gate calls + enhancement)
+**Time**: ~2-3 minutes total
+
+## Platform-Specific Databases
+
+### Database Structure
+Each platform has a JSON database with 20+ proven hooks:
+
+```json
+{
+ "platform": "twitter",
+ "description": "Real high-performing Twitter hooks for AI training",
+ "source": "Bird CLI competitive data + 2026 viral research",
+ "updated": "2026-01-15",
+ "voice_requirements": {
+ "tone": "Conversational, punchy, quotable",
+ "structure": "Short lines with white space",
+ "forbidden": ["Most people think", "Imagine", "Picture this"]
+ },
+ "examples": [
+ "hook": "I watched Kip Bodnar build...",
+ "engagement": {"likes": 7005, "retweets": 666},
+ "why_it_works": "First-person witness, specific timeframe",
+ "pattern": "First-Person Witness + Transformation",
+ "validated": true
+ }
+ ],
+ "anti_patterns": [
+ "hook": "Kip Bodnar built an AI app...",
+ "why_it_fails": "Third-person reporter, information overload",
+ "pattern": "Information Overload",
+ "validated": "2026-01-15 linkedgraphix test (scored 76/100)"
+ ]
+
+### Available Databases
+- `database/twitter-hooks.json` - 39+ Twitter hooks (7k+ likes validated, 19 AI agent + 10 Enchantents)
+- `database/linkedin-hooks.json` - 38+ LinkedIn hooks (81/100 validated, 8 AI agent + 10 Enchantents)
+- `database/instagram-hooks.json` - 30+ Instagram Reels hooks (10 AI agent + 10 Enchantents)
+- `database/youtube-hooks.json` - 22+ YouTube hooks (8 AI agent + 10 Enchantents)
+
+**Niche Tags** (each hook has a `niche` field for filtering):
+- `"ai-agents"` — AI tools, automation, AI agent content (aiwithjai, linkedgraphix, clawver)
+- `"enchantents-glamping"` — Glamping tents, hospitality investment (enchantents, nomadixgear)
+- `(no niche field)` — General hooks, used as fallback for any brand
+
+**How niche filtering works:**
+1. Pass `--brand enchantents` → auto-detects niche `enchantents-glamping`
+2. Loads niche-specific hooks FIRST, then general hooks as fallback
+3. Generator sees the most relevant examples → better output
+4. Override with `--niche ai-agents` if needed
+
+# Automatically loads appropriate database
+node .claude/skills/hook-generator/scripts/generate-hooks-unified.js \
+ --content "path/to/blog.md" \
+ --brand linkedgraphix.app \
+
+**What it does**:
+- Loads 20+ proven hooks for target platform
+- Selects top 5-7 examples to show Gemini
+- Includes engagement data, patterns, "why it works" explanations
+
+### Step 2: Extract Content Elements
+Simple extraction (Gemini does heavy analysis):
+- **Numbers**: `40%`, `3 minutes`, `16 talking points`
+- **Timeframes**: `under 10 minutes`, `30-minute video`
+- **Quotes**: Direct quotes from content
+- **Context**: First 500 characters for understanding
+
+### Step 3: Generate with Gemini CLI
+**Prompt Structure**:
+Study these 7 PROVEN [platform] hooks:
+
+EXAMPLE 1:
+Hook: [actual hook text]
+Engagement: [real data]
+Why it works: [specific reasons]
+Pattern: [pattern name]
+
+[... 6 more examples ...]
+
+TASK: Generate 3 hooks using the SAME PATTERNS as proven examples.
+
+For each hook, provide:
+1. hook_text: The actual hook
+2. pattern: Which example pattern it follows
+3. why_it_works: Specific reason for performance
+4. estimated_score: 0-100 (based on pattern match + specificity)
+5. similar_to_example: Which example (1-7) it's most similar to
+
+**Token Savings**: Hook examples (2-3k tokens) loaded in Gemini's context ($0.075/1M), not Claude's ($3/1M) = **95%+ savings**
+
+### Step 4: Score & Rank
+Ranks hooks by:
+- Pattern match strength (does it follow proven pattern?)
+- Specificity (concrete numbers, names, timeframes?)
+- Curiosity gap (creates desire to learn more?)
+- Platform fit (matches platform voice requirements?)
+
+## Usage Examples
+
+### Command Line
+--content "brands/linkedgraphix.app/content/03-production/blog/article.md" \
+
+# Generate LinkedIn hooks
+--platform linkedin \
+
+### From Claude Code (Skill Invocation)
+Skill: hook-generator
+Platform: twitter
+Content: brands/{brand}/content/03-production/blog/{slug}/03-final.md
+Brand: {brand}
+Count: 3
+
+### Programmatic (Node.js)
+```javascript
+const { generateHooksWithGemini, loadPlatformHooks } = require('./.claude/skills/hook-generator/scripts/generate-hooks-unified.js');
+
+const hookDatabase = loadPlatformHooks('twitter');
+const content = fs.readFileSync('article.md', 'utf8');
+const elements = extractContentElements(content);
+
+const hooks = await generateHooksWithGemini('twitter', hookDatabase, content, elements, 3);
+console.log(hooks);
+
+## Validated Results (LinkedGraphix Test - 2026-01-15)
+
+### Twitter Hooks Generated
+1. **88/100**: "I watched Kip build an AI app live. Under 10 minutes. 16 viral talking points from ONE video..." (Building in Public Journey)
+2. **85/100**: "Kip Bodnar just killed the content repurposing industry..." (Shocking Action + Transformation)
+3. **87/100**: "every YouTube video you watch manually is now a wasted opportunity..." (Contrarian + Hidden Opportunity)
+
+### LinkedIn Hooks Generated
+1. **81/100**: "I watched Kip Bodnar build an AI content repurposing app live..." (First-Person Witness - exact validated pattern)
+2. **84/100**: "Two content creators, same YouTube video..." (A/B Comparison)
+3. **82/100**: "A marketing exec just proved why most content creators are wasting hours..." (Exec Validation)
+
+**Average**: 84.5/100 across both platforms
+**Improvement**: +8.5 points over 76/100 baseline
+**Consistency**: All hooks 81-88/100 (no quality variance)
+
+## Key Principles
+
+### 1. Examples Over Formulas
+**OLD Approach**:
+"Be contrarian. Create curiosity gap. Use specific numbers."
+
+**NEW Approach**:
+"Study these 7 PROVEN hooks with 7k+ likes.
+Generate hooks using the SAME PATTERNS."
+
+**Result**: AI learns from concrete examples, not abstract rules.
+
+### 2. Platform-Specific Libraries
+**Why**: Twitter hooks ≠ LinkedIn hooks ≠ Instagram hooks
+
+**Example**:
+- **Twitter**: "I watched Kip build..." (conversational, punchy)
+- **LinkedIn**: "I watched Kip Bodnar build an AI content repurposing app live in under 10 minutes." (professional depth)
+- **Instagram**: Text overlay: "This feels illegal to know." + Suspenseful audio (visual-first)
+
+### 3. Validated 2026 Data
+**Sources**:
+- LinkedIn: 300 posts tested, algorithm analysis
+- Instagram: 1.9M+ Reels analyzed
+- Twitter: Real engagement (7k+ likes, 666 RTs)
+- LinkedGraphix validation: 81/100 (LinkedIn), 76/100 (Twitter) - identified what works vs what doesn't
+
+### 4. Token Efficiency
+**Cost Comparison**:
+- **Inline generation** (Claude): 40k tokens @ $3/1M = $0.12
+- **Unified generator** (Gemini CLI): 3k tokens @ $0.075/1M = $0.0002
+- **Savings**: 95%+ per generation
+
+### 5. Continuous Improvement
+**Process**:
+1. Track published hooks + engagement (monthly)
+2. Add top performers to databases
+3. Remove underperformers
+4. System improves automatically
+
+### 6. Quantifiable Tension First (SkillOpt-trained, 2026-09-27)
+**Rule**: Immediately present a quantifiable problem or tension that addresses the reader's exact situation. No generic "people" or abstract concepts — each hook is 1-2 sentences.
+
+**Source**: Validation-gate-accepted edit from a live SkillOpt training run (baseline 0.60 → 1.00 on held-out hook tasks). Apply this as a hard requirement when scoring Step 1 volume output: hooks that open with a vague audience or an unquantified tension get regenerated before Ralph enhancement.
+
+## Integration Points
+
+### Current Integrations (UPDATED 2026-01-15)
+**platform-content-generator** (Phase 2):
+- INTEGRATED: Invokes unified hook generator
+- Generates 3 hooks per platform
+- Uses platform-specific databases
+- Selects best hook for final content
+
+**newsletter-strategist**:
+- ⏳ PENDING: Update to use unified generator for subject lines
+
+**social-content-generator**:
+- ⏳ PENDING: Update to use unified generator for social hooks
+
+## Performance Metrics
+
+### Performance Metrics (v3.0)
+**Quality Targets**:
+- v2.0 Baseline: 85-88/100 (Gemini/Haiku generation only)
+- v3.0 Target: 90-95/100 (Haiku + Sonnet + Ralph)
+- Success Rate: 90%+ platforms reach target (vs 50% v2.0)
+- Voice Quality: 100/100 maintained (zero AI clichés)
+
+**Efficiency Targets**:
+- Token Savings: 95%+ (Gemini vs Claude)
+- Cost: $0.50-0.70 per batch (50 hooks + selection + enhancement)
+- Consistency: Same quality across all platforms
+- Speed: 2-3 minutes per generation
+
+## Ralph Enhancement Methodology
+
+### What is Ralph Enhancement?
+Ralph is a quality scoring system that enhances hooks from 85-88/100 to 90-95/100 through:
+1. **10-dimensional scoring** (Hook Effectiveness, Voice, Specificity, Emotion, etc.)
+2. **Targeted improvement prompts** (focus on weak dimensions < 7/10)
+3. **Pattern teaching** (explain WHY patterns work)
+4. **Iterative refinement** (repeat until 90/100 or early stop)
+
+### Enhancement Process
+**For each of top 3 hooks**:
+
+**Step 1: Score** (6 dimensions):
+- Hook Effectiveness (40% weight) - Stops scrolling in 0.3 seconds?
+- Voice Authenticity (20% weight) - Sounds like real person?
+- Specificity (15% weight) - Concrete details vs vague claims?
+- Emotion Strength (10% weight) - Single clear emotion at 7+/10?
+- Pattern Recognition (10% weight) - Uses proven formulas?
+- Platform Optimization (5% weight) - Platform-specific format?
+
+**Step 2: Generate Improvement Prompt** (if score < 90/100):
+- Load brand voice foundation (with 9/10 examples)
+- Identify weak dimensions (score < 7/10)
+- Inject banger hook options (if hook effectiveness weak)
+- Teach patterns (Zeigarnik effect, cognitive dissonance, etc.)
+- Preserve strong dimensions (score ≥ 7/10)
+- Apply AI slop blacklist (remove generic phrases)
+- Provide implementation steps
+
+**Step 3: Improve** (Gemini/Claude):
+- Apply improvement prompt, Generate enhanced hook, Re-score enhanced version
+- Repeat if needed (max 3 iterations)
+- Early stop if no progress
+
+**Step 4: Output**:
+- Enhanced hook at 90-95/100
+- Dimension breakdown
+- Improvement delta (+X points)
+- Number of iterations
+
+### Why Ralph Enhancement Works
+**Compounds Quality**:
+- Haiku baseline: 85-90/100 (zero AI slop)
+- Sonnet selection: Best patterns from 50 options
+- Ralph enhancement: Targeted improvements per dimension
+- **Result: 90-95/100 consistently**
+
+**Teaches Patterns**:
+- Explains WHY hooks work (neuroscience, psychology)
+- Shows 9/10 examples before improvement
+- Injects banger hook alternatives
+- System learns from improvements
+
+**Preserves Strengths**:
+- Doesn't break what's working (dimensions ≥ 7/10)
+- Focuses only on weak areas
+- Maintains voice authenticity
+- Prevents trade-offs
+
+## Troubleshooting
+
+### "Generated hooks are generic"
+- **Check**: Are you loading platform-specific database?
+- **Fix**: Verify database path and platform parameter match
+
+### "Quality varies between platforms"
+- **Check**: Are examples being shown before generation?
+- **Fix**: Ensure 5-7 examples loaded in Gemini's context
+
+### "Hooks don't match brand voice"
+- **Check**: Does database include brand-specific examples?
+- **Fix**: Add brand's top performers to platform database
+
+### "Gemini CLI error"
+- **Check**: `gemini` command available in PATH?
+- **Fix**: Use Task tool with examples as workaround pattern
+
+## Supporting Resources
+- `README.md` - Complete system documentation (110+ lines)
+- `database/2026-PLATFORM-RESEARCH.md` - Research sources and findings
+- `database/{platform}-hooks.json` - 6 platform-specific databases
+- `scripts/generate-hooks-unified.js` - Main generator script
+- `references/hook-generation-system.md` - Original 6-formula system (deprecated)
+- `references/platform-optimization.md` - Platform-specific rules
+
+## Migration from Old System
+**OLD System** (7-formula):
+- Abstract formulas: "Be contrarian", "Create curiosity gap"
+- Inconsistent results (LinkedIn 81, Twitter 76)
+- No examples shown during generation
+
+**NEW System** (Unified v2.0):
+- Examples-first: Show 5-7 proven hooks
+- Consistent 85/100+ across platforms
+- Real engagement data validates patterns
+
+**Migration Steps**:
+1. Update platform-content-generator SKILL.md (completed 2026-01-15)
+2. ⏳ Update newsletter-strategist agent
+3. ⏳ Update social-content-generator agent
+4. ⏳ Archive old 7-formula system
+
+**Status**: Production-ready | Validated 2026-01-15 | 95%+ token savings | 85/100+ quality

package/skills/hook-generator/database/2026-PLATFORM-RESEARCH.md

@@ -0,0 +1,211 @@
+# 2026 Platform Research: What's Actually Working
+**Last Updated**: 2026-01-13
+**Sources**: LinkedIn algorithm analysis (300 posts tested), Instagram Reels trends, LinkedGraphix validation test
+
+---
+
+## LinkedIn (B2B Professional)
+
+### Algorithm Changes (Early 2026)
+**Major Shifts**:
+1. **Link Penalty**: Posts with external links get **60% less reach** vs identical posts without links
+2. **Comments > Likes**: Post with 50 comments outperforms post with 500 likes
+3. **Engagement Bait Detection**: Phrases like "Agree or disagree?" and "What would you add?" are now penalized
+
+**Sources**:
+- [The 2026 LinkedIn Algorithm: What Actually Works Right Now](https://rivereditor.com/blogs/2026-linkedin-algorithm-what-works-300-posts-tested)
+- [LinkedIn's 2026 Algorithm: The Engagement Bait Era Is Finally Over](https://synergist-digital-media.ghost.io/linkedins-2026-algorithm-the-engagement-bait-era-is-finally-over/)
+
+### Top-Performing Formats
+1. **Text-only posts**: Most reach per follower (no image penalty)
+2. **PDF carousels**: Highest engagement (keeps users on LinkedIn)
+3. **Video clips (30-90s)**: Thriving (platform-native engagement)
+4. **Image posts**: 30% LESS reach than text-only (underperforming in 2026)
+
+### Hook Requirements
+**Critical**: First 2 lines determine engagement (before "see more" cutoff)
+
+**Top Hook Types**:
+1. **First-Person Witness**: "I watched [person] do [surprising thing]..."
+ - Example: "I watched Kip Bodnar build an AI app live in under 10 minutes." (81/100 score)
+ - Why: Personal, credible, creates FOMO
+
+2. **Surprising Statistics**: "I analyzed 10,000 [things]. [X%] had [finding]."
+ - Creates credibility + curiosity gap
+
+3. **Problem-Solution Teaser**: Identifies pain point, promises framework
+ - Example: "Most business owners don't know if SEO makes them money. Here's how I track..."
+
+### Validated Performance (Real Data)
+**Post 1** (SEO Lead Tracking):
+- Hook: "How I track SEO leads & make sure SEO is ROI positive"
+- Engagement: 172 likes, 123 comments, 4 shares
+- Why: Tackled widespread pain point with framework promise
+
+**Post 2** (AI Skills):
+- Hook: "AI skills will be crucial for years to come. Knowing how to use ChatGPT isn't enough."
+- Engagement: 1,847 likes, 321 comments, 285 shares
+- Why: Fear-based motivation + save-worthy visual reference
+
+**Source**: [Anatomy of a Viral LinkedIn B2B Post](https://gracker.ai/blog/viral-linkedin-b2b-post-anatomy)
+
+### Posting Cadence
+- **Optimal**: 3-4 high-quality posts/week
+- **Timing**: 7-9 AM audience timezone (fastest initial engagement)
+- **Days**: Tuesday-Thursday > Monday/Friday
+
+### What to Avoid
+- Engagement pods (detected and penalized)
+- Overly promotional content (suppressed)
+- Generic engagement bait phrases
+- External links in posts (60% reach penalty)
+
+## Instagram Reels (Visual/Creator)
+
+### Algorithm Priorities (2026)
+**Ranking Factors**:
+1. **Completion rate** (15-60 second sweet spot)
+2. **Saves/shares > likes** (high-intent actions prioritized)
+3. **First 3 seconds** = critical hook window
+
+### Top Audio Trends (Jan 2026)
+**#1: "Every Breath You Take"** (812K Reels)
+- Pattern: Emotional, nostalgic reflection + milestone moments
+- Works with: Slow transitions, close-ups, sentimental reveals
+
+**#2: "I Wanna Thank Me"** (18K Reels)
+- Pattern: Self-empowerment + celebration of personal wins
+- Best for: Before-after transformations, achievement showcases
+
+### December 2025 Standouts
+**"Twinkling Lights (Reimagined)"** (1.9M Reels) - *Top Performer*
+- Hook: Soft sparkle atmosphere without oversaturation
+- Strategy: Minimal movement, steady framing, let light carry narrative
+
+**"Turn the Lights Off"** (310K Reels)
+- Hook: Builds anticipation through contrast (stillness → motion)
+- Pattern: Viewers stay waiting for shift
+
+### Visual Patterns That Work
+**What's Winning**:
+- Slow transitions > quick cuts (for emotional content)
+- Steady framing > aggressive camera movement (reflective content)
+- Natural light + textures drive engagement
+- Intentional pacing > polish (authenticity beats overproduction)
+
+**What's Underperforming**:
+- Fast edits for emotional audio
+- Overly polished, corporate-feeling content
+
+### Hook Formulas (Text Overlays)
+Based on trending patterns:
+1. "This feels illegal to know."
+2. "I've been keeping a secret..."
+3. "You're not going to believe what I just did/found out."
+4. "MYTH: [common belief]"
+5. "I was wrong..."
+
+**Source**: [Instagram Reels Trends 2026](https://www.socialpilot.co/blog/instagram-reels-trends)
+
+## Twitter/X (Conversational/Viral)
+
+### Validated Patterns (From Existing Database)
+**Top Performers** (Real Engagement Data):
+
+1. **"10 days into 2026..."** (rauchg)
+ - 7,005 likes, 666 retweets
+ - Pattern: Timely Event List + Bold Prediction
+ - Why: Specific credible sources + urgency
+
+2. **"in an AI world, your ability to clearly describe..."** (gregisenberg)
+ - 2,910 likes, 314 retweets
+ - Pattern: Contrarian Prediction
+ - Why: Challenges assumption (not coding, but communication)
+
+3. **"every failed VC-backed consumer app..."** (gregisenberg)
+ - 2,989 likes, 130 retweets
+ - Pattern: Hidden Opportunity Reveal
+ - Why: Specific opportunity ($1-10M) + actionable CTA
+
+### LinkedGraphix Validation (2026-01-13)
+**What DIDN'T Work** (76/100 score):
+- Hook: "Kip Bodnar built an AI app live that extracts 16 viral talking points..."
+- Problem: Third-person reporter, too much info crammed, no emotion
+
+**What WORKED** (LinkedIn, 81/100 score):
+- Hook: "I watched Kip Bodnar build an AI app live. Under 10 minutes."
+- Why: First-person witness, dramatic pacing, simple structure
+
+### Key Insights
+**Voice Requirements**:
+- Conversational, punchy, quotable
+- Short lines with white space
+- Use → arrows and bullets
+- NO: "Most people think", "Imagine", "Picture this"
+
+**Length**: Quality-driven, no arbitrary limits, optimize for feed readability
+
+## YouTube (Long-Form Video)
+**First 8 seconds critical** (viewer decides to stay or leave)
+
+1. **Problem Reveal**: "The biggest mistake [audience] make is..."
+2. **Bold Promise**: "By the end of this video, you'll know exactly how to..."
+3. **Pattern Interrupt**: Show unexpected result first, then explain
+
+### Title Patterns
+- Specificity > vagueness ("16 talking points" not "many tips")
+- Numbers work (lists, timeframes, results)
+- Keyword at beginning for search
+
+## TikTok (Short-Form Video)
+
+### Algorithm Priorities
+**Watch time** = primary ranking factor
+
+**First 2 seconds** = decide to watch or scroll
+
+### Hook Patterns
+1. **Shock/Surprise**: "I can't believe this actually worked..."
+2. **Relatability**: "POV: You're trying to..."
+3. **Question**: "Why does nobody talk about..."
+4. **Challenge**: "I tried [trend] for 30 days. Here's what happened."
+
+## Medium (Long-Form Written)
+
+### Article Performance
+**Title patterns**:
+1. "I [did surprising thing]. Here's what I learned."
+2. "[Number] [things] that [surprising result]"
+3. "The [thing] nobody talks about"
+
+**Structure**: Similar to LinkedIn but longer (1500-3000 words)
+
+**SEO**: Cross-post from blog with canonical URL
+
+## Cross-Platform Insights
+
+### Universal Patterns (Work Everywhere)
+1. **Specificity beats vagueness**: "16 talking points in 10 minutes" > "lots of content quickly"
+2. **First-person > third-person**: "I watched..." > "Kip Bodnar did..."
+3. **Curiosity gaps**: Promise reveal ("Here's the framework:")
+4. **Emotional triggers**: Surprise, opportunity, concern (NOT generic excitement)
+5. **Contrarian angles**: Challenge common beliefs
+
+### What's Dead in 2026
+- "Most people think..." (AI cliché)
+- Generic promises ("Save time", "Make money")
+- Ranges instead of specifics ("10-35 hours")
+- Engagement bait phrases
+- Overly promotional content
+
+## Validation Methodology
+1. Algorithm research (300+ posts tested)
+2. Real engagement data (likes, comments, shares)
+3. LinkedGraphix test (81/100 LinkedIn, 76/100 Twitter)
+4. Platform trend analysis (monthly top performers)
+
+**Next Steps**:
+1. Build platform-specific hook databases
+2. Tag hooks by pattern + engagement data
+3. Create unified generator that loads relevant examples
+4. Continuous learning: add monthly top performers