Tiktok Script Generator

TikTok Script Generator Skill v2.3

PURPOSE: Generate production-ready UGC TikTok ad scripts with official ElevenLabs v3 audio tags (German VO + mandatory Chinese translation) EXECUTOR: Claude Code (for quality creative writing) INPUT: Analysis files from tiktok-product-analysis skill OUTPUT: 3 scripts + Campaign Summary (references analysis, no duplication) STYLE: Fast-paced UGC ads with mandatory emotion cues (1-2 per line), engaging delivery, dynamic performance OPTIMIZATION: Batched Write calls (4 files per product in single message) ⭐ 2x faster

⚠️ CRITICAL FORMAT RULE - READ THIS FIRST:

• INLINE CUES ONLY: [emotion] Text here. (cue + text on SAME LINE)
• NEVER BROKEN LINES: Do NOT write [emotion] on one line followed by text on the next line
• EVERY LINE NEEDS CUES: No orphan text without emotion tags
• See templates/voiceover_format.md for complete format specification

Agent Responsibility

Key Lesson: Gemini produces generic placeholder scripts. Claude produces detailed, product-specific content with natural voiceovers.

Workflow (4 Steps - OPTIMIZED)

[Step 0: Pre-Check] → GATE
        │
        ▼
[Step 1: Read Analysis Files in Parallel] → Extract key insights
        │                                     (5+ Read calls at once)
        ▼
[Step 2: Generate All Scripts + Summary] → Create all 4 files
        │                                    (in single response)
        ▼
[Step 3: Write All Files in Parallel] → 4 Write calls simultaneously
        │                                ⭐ 2x faster
        ▼
[Step 4: Quality Gate] → Verify completeness

Key Optimization (v2.3.0): Steps 2-3 now execute in ONE MESSAGE with parallel Write tool calls, eliminating sequential overhead.

Step 0: Pre-Check Gate (MANDATORY)

DO NOT PROCEED if analysis files are missing.

product_id="{product_id}"
# Use the correct dated batch folder:
date="YYYYMMDD"
base="product_list/$date/$product_id"

echo "=== PRE-CHECK: $product_id ==="

# MANDATORY: Video synthesis must exist
if [ ! -f "$base/ref_video/video_synthesis.md" ]; then
    echo "❌ BLOCKED: video_synthesis.md missing"
    echo "Run tiktok-product-analysis skill first"
    exit 1
fi

lines=$(wc -l < "$base/ref_video/video_synthesis.md" | tr -d ' ')
if [ "$lines" -lt 150 ]; then
    echo "❌ BLOCKED: video_synthesis.md too short ($lines lines, need 150+)"
    exit 1
fi

# CRITICAL: Verify synthesis has properly flagged compliance risks
echo "Checking compliance flagging in analysis..."
if ! python3 scripts/validate_compliance_flags.py "$base/ref_video/video_synthesis.md" >/dev/null 2>&1; then
    echo "⚠️ WARNING: video_synthesis.md has compliance risks not properly flagged"
    echo "   Review and ensure risks are documented in 'Compliance & Trust Signals' section"
    echo "   Run: python3 scripts/validate_compliance_flags.py \"$base/ref_video/video_synthesis.md\""
    # Don't block, but warn - analysis files can have flagged risks
fi

echo "✅ Pre-check passed. Ready for script generation."

If blocked: Run tiktok-product-analysis skill first.

Recommended (Repo Verifier)

bash scripts/verify_gate.sh --date YYYYMMDD --csv scripts/products.csv --phase analysis

Step 1: Read Analysis Files (PARALLEL v2.3.0)

CRITICAL OPTIMIZATION: Use parallel Read tool calls to fetch ALL files in ONE MESSAGE.

Files to read (make 4-5 parallel Read calls):

1. ref_video/video_synthesis.md (CRITICAL)

   • Hook patterns, selling points, replication strategy
   • Target audience, production patterns
   • DO's and DON'Ts

2. product_images/image_analysis.md (if exists)

   • Visual hooks (Section 10)
   • German text from packaging
   • Color/variant recommendations

3. tabcut_data.json (product metadata)

   • Product name, price, shop
   • Sales data, conversion rate

4. fastmoss_data.json (fallback if tabcut data is missing/unknown)

   • Use if product_name is "Unknown Product" or key sales metrics are null

5. Individual video_N_analysis.md files (optional, for deep dive)

   • Use if video_synthesis.md lacks detail

Example parallel execution:

<Read file_path="product_list/YYYYMMDD/{product_id}/ref_video/video_synthesis.md" />
<Read file_path="product_list/YYYYMMDD/{product_id}/product_images/image_analysis.md" />
<Read file_path="product_list/YYYYMMDD/{product_id}/tabcut_data.json" />

All 3 files read simultaneously in ~2-3 seconds (was 6-9 seconds sequential).

Extract these key elements:

• Top 3 hook patterns from synthesis
• Top 5 selling points (ranked)
• 3 recommended script angles
• German terminology from packaging
• Visual filming instructions

Step 2-3: Generate and Write All Scripts (BATCHED v2.3.0)

CRITICAL OPTIMIZATION: Generate all 3 scripts + Campaign Summary in ONE response, then make 4 parallel Write tool calls.

Old workflow (sequential - SLOW):

Generate Script 1 → Write Script 1 → Generate Script 2 → Write Script 2 → Generate Script 3 → Write Script 3 → Generate Summary → Write Summary
Total: ~5-8 minutes per product

New workflow (batched - FAST):

Generate all 4 files (Scripts 1-3 + Summary) → Write all 4 in parallel
Total: ~2-3 minutes per product ⭐ 2x faster

Implementation:

1. After reading analysis files, generate complete content for all 4 files
2. In a SINGLE MESSAGE, make 4 Write tool calls in parallel:
   • Write Script 1
   • Write Script 2
   • Write Script 3
   • Write Campaign Summary

Each script must have:

Required YAML Frontmatter (Obsidian)

See templates/frontmatter_template.md for complete specification.

Quick reference:

• duration target: 00:40–00:50
• tags max 5 and meaningful for commerce/interest
• caption uses YAML block scalar caption: >- with hashtags appended (space-separated)
• Always include source_notes to exact files used

TikTok Policy Compliance (DE Market) - ENHANCED v2.0

CRITICAL: These rules prevent TikTok ad rejection and "false advertising" complaints in the German market. All scripts MUST pass compliance validation before publication.

POLICY 1: No Exact Low Price Bait 🚫

Prohibited: Specific € amounts combined with urgency, comparison, or scarcity framing.

Safe Price Language:

• German: "preiswert", "erschwinglich", "für kleines Geld", "fairer Preis", "gutes Preis-Leistungs-Verhältnis"
• Chinese: "价格很友好", "很划算", "性价比高", "不贵"

Rule: Avoid exact €, comparison math ("statt/vs"), or urgency tied to price.

POLICY 2: No Absolute Effect Claims 🚫

Prohibited: "100%", "pure", "perfect", "genauso gut", "never", "always", "instant" without qualification.

Safe Comparison Language:

• German: "hilft bei", "unterstützt", "kann", "für... geeignet", "stark genug"
• Chinese: "有助于", "支持", "适合", "功能强大", "能满足需求"

Rule: Use qualified claims with "hilft bei", "unterstützt", or specific context instead of absolutes.

POLICY 3: No Exaggerated Promotions 🚫

Prohibited: Unverifiable superlatives, false urgency, hyperbolic claims without proof.

Safe Promotion Language:

• German: "praktisch", "clever", "hilfreich", "empfehlenswert", "wertvoll", "besonders"
• Chinese: "实用", "方便", "值得推荐", "有价值", "特别"

Rule: Avoid hyperboles without proof. Use descriptive, specific benefits instead.

POLICY 4: Medical & Health Claims 🚫

Prohibited: Therapy, healing, pain relief, medical treatment framing.

Safe Wellness Language:

• German: "Entspannung", "Wohlbefinden", "lockert Verspannungen", "angenehm", "komfortabel", "entspannend"
• Chinese: "放松", "舒适", "缓解紧绷感", "享受"

Rule: Never promise medical outcomes. Use wellness/comfort framing only.

POLICY 5: Waterproof & Tech Specs 🚫

Prohibited: Absolutes without IP rating proof, ambiguous tech claims.

Safe Tech Language:

• Waterproof: Use IP rating ONLY if sourced (e.g., "IP67" from packaging)
• Performance: Use qualified terms ("flüssig", "schnell", "zuverlässig")
• Avoid: "instant", "zero", "100%", "komplett", "total"

Compliance Validation Checklist

Before finalizing scripts, verify:

# Run compliance validator on generated scripts
for script in product_list/YYYYMMDD/{product_id}/scripts/*.md; do
  [[ "$(basename "$script")" == "Campaign_Summary.md" ]] && continue
  python3 scripts/validate_compliance_flags.py "$script"
done

Manual Review Points:

• No € symbols in captions or voiceovers
• No "100%", "pure", "perfect", "genauso gut", "besser als"
• No "unbezahlbar", "genial", false urgency ("weg ist", "letzte Chance")
• Medical terms replaced with wellness language
• Tech specs qualified (not absolute)

If validation fails:

1. Check the specific violation flagged
2. Find the compliant alternative in the table above
3. Replace and re-validate
4. Never bypass compliance checks

Required Sections

## Scripts

[1–2 sentence concept description]

### Structure (35s)
- Hook: [Description] (0-3s)
- Problem/Reveal: [Description] (3-8s)
- Solution/Demo: [Description] (8-20s)
- Benefits: [Description] (20-30s)
- CTA: [Description] (30-35s)

### Visual Strategy - ENHANCED TEMPLATE

**Required format:** Seconds | Visual | Purpose (psychological function)

**Sample-quality example:**

| Seconds | Visual | Purpose |
|:--------|:-------|:--------|
| 00-03 | Screenshot of negative comment: "Breaks after 1 day" with creator looking annoyed | Hook: Relatable objection creates curiosity |
| 03-07 | Hands dropping LARGE ice cubes + frozen berries into transparent cup | Build tension: "Can it really handle this?" |
| 07-10 | Close-up: LED display showing battery at 95% before starting | Trust: Tech credibility with real data |
| 10-15 | Blender upside down, ice crushing visible through transparent cup | Proof: Visual confirmation of power |
| 20-25 | Final smoothie texture shot (creamy, no chunks) | Result proof: Delivers on promise |

**Purpose categories (choose 1 per shot):**
- **Hook:** Attention/Relatability/Pattern Interrupt
- **Problem:** Pain amplification/Anchoring
- **Proof:** Evidence/Trust building/Demonstration
- **Objection Handling:** Removes barrier (noise, battery, cleaning)
- **Social Proof:** FOMO/Aspiration
- **Conversion:** CTA/Urgency

**Minimum:** 8-10 shots for 40-50s, each with clear psychological purpose

Voiceover Sections

See templates/voiceover_format.md for complete specification and all audio tags.

Critical requirements:

• INLINE cues: [emotion] Text. on SAME LINE
• 1-2 audio tags per line (MANDATORY)
• Both German (DE) and Chinese (ZH) sections required
• Target length: 40-50s (100-150 German words, 130-180 Chinese characters)

Script Naming Convention

Product_Model_KeyAngle.md   # e.g., HTC_NE20_AI_Uebersetzer_Earbuds.md

Campaign Summary (Reference-Based - Included in Step 2-3)

See templates/campaign_summary_structure.md for complete template.

Key Change from v1: DO NOT duplicate content from analysis files. REFERENCE them.

MANDATORY: Use inline Chinese throughout (English text (中文翻译))

Required sections:

1. Product Overview | 产品概述
2. Campaign Strategy | 活动策略
3. Key Selling Points | 核心卖点
4. Target Audience | 目标受众
5. Scripts Summary | 脚本摘要
6. Production Notes | 制作说明
7. Source Files | 源文件

Step 4: Quality Gate (Post-Write Validation)

product_id="{product_id}"
date="YYYYMMDD"
scripts_dir="product_list/$date/$product_id/scripts"

echo "=== QUALITY GATE: $product_id ==="

# Check script count (exclude Campaign_Summary.md)
script_count=$(ls -1 "$scripts_dir"/*.md 2>/dev/null | grep -v 'Campaign_Summary.md' | wc -l | tr -d ' ')
if [ "$script_count" -lt 3 ]; then
    echo "❌ FAIL: Only $script_count scripts (need 3)"
    exit 1
fi
echo "✅ Scripts: $script_count"

# Check Campaign Summary
if [ ! -f "$scripts_dir/Campaign_Summary.md" ]; then
    echo "❌ FAIL: Campaign_Summary.md missing"
    exit 1
fi
echo "✅ Campaign Summary exists"

# Check script quality (not stubs)
for script in "$scripts_dir"/*.md; do
    if [ "$(basename "$script")" = "Campaign_Summary.md" ]; then
        continue
    fi
    lines=$(wc -l < "$script" | tr -d ' ')
    if [ "$lines" -lt 40 ]; then
        echo "⚠️ WARNING: $(basename $script) may be a stub ($lines lines)"
    fi

    # Check for placeholder content
    if grep -q 'product: "Product"' "$script"; then
        echo "❌ FAIL: $(basename $script) has placeholder product name"
        exit 1
    fi
    if ! grep -q '### ZH' "$script"; then
        echo "❌ FAIL: $(basename $script) missing Chinese translation section"
        exit 1
    fi

    # Caption should use YAML block scalar so hashtags don't break YAML parsing
    if ! awk 'BEGIN{in=0} /^---$/{in=!in} in && /^caption: >-/{ok=1} END{exit !ok}' "$script"; then
        echo "⚠️ WARNING: $(basename $script) caption is not using 'caption: >-' (hashtags may break YAML)"
    fi
done
echo "✅ Script quality verified"

# CRITICAL: TikTok Policy Compliance Validation
echo ""
echo "=== TIKTOK POLICY COMPLIANCE CHECK ==="
echo "Validating scripts against TikTok advertising policies..."
compliance_fail=0
for script in "$scripts_dir"/*.md; do
    [[ "$(basename "$script")" == "Campaign_Summary.md" ]] && continue
    [[ -e "$script" ]] || continue

    if ! python3 scripts/validate_compliance_flags.py "$script" >/dev/null 2>&1; then
        echo "❌ FAIL: $(basename "$script") has compliance violations"
        echo "   Run: python3 scripts/validate_compliance_flags.py \"$script\""
        compliance_fail=1
    fi
done

if [ "$compliance_fail" -eq 1 ]; then
    echo ""
    echo "⚠️ COMPLIANCE VIOLATIONS DETECTED"
    echo "Fix violations before proceeding. See 'TikTok Policy Compliance' section for safe alternatives."
    exit 1
fi
echo "✅ All scripts pass TikTok policy compliance"

echo ""
echo "=== QUALITY GATE PASSED ==="

Batch Processing (v2.3.0 Optimized)

For multiple products, process sequentially with batched writes:

Product A: [Read all files in parallel] → [Generate + Write 4 files in parallel] → [Gate] = 2-3 min
Product B: [Read all files in parallel] → [Generate + Write 4 files in parallel] → [Gate] = 2-3 min
Product C: [Read all files in parallel] → [Generate + Write 4 files in parallel] → [Gate] = 2-3 min

Performance Improvement:

• Old (sequential writes): 5-8 min per product × 8 products = 40-64 min
• New (batched writes): 2-3 min per product × 8 products = 16-24 min
• Savings: ~25-40 minutes for 8-product batch ⭐ 2x faster

Why sequential across products (not parallel)?

• Claude Code produces better quality with focused attention
• Each product needs full context window for analysis
• Quality over speed for creative content
• BUT: We optimize WITHIN each product via batched writes

Time estimate (v2.3.0): ~2-3 minutes per product for script generation

Common Mistakes to Avoid

❌ Mistake 1: BROKEN LINE FORMAT (Most Common Error)

Symptoms: Writing emotion cues on separate lines from text

[frustrated]
Du kennst das?

Why it fails: Validation script expects [cue] Text. on SAME LINE. Separate lines create "orphan" text without cues. Fix: ALWAYS write inline: [frustrated] Du kennst das? Validation Error: "Low cue density" / "No German voiceover section found"

❌ Mistake 2: Starting before analysis complete

Fix: Always run Step 0 Pre-Check first

❌ Mistake 3: Generic placeholder content

Symptoms: "You won't believe...", product="Product" Fix: Read analysis files, use specific product details

❌ Mistake 4: Duplicating analysis content in Campaign Summary

Fix: Reference files with quotes, don't copy-paste

❌ Mistake 5: Missing ElevenLabs v3 markers

Fix: Every VO line needs [tone] marker at start (inline format)

❌ Mistake 6: Skipping Chinese translations

Fix: Chinese translation is mandatory for every script (DE + ZH sections required)

File Structure After Completion

product_list/YYYYMMDD/{product_id}/
├── tabcut_data.json                    # From scraper
├── product_images/
│   ├── *.webp
│   └── image_analysis.md               # From analysis skill
├── ref_video/
│   ├── video_*.mp4
│   ├── video_*_analysis.md             # From analysis skill
│   └── video_synthesis.md              # From analysis skill (CRITICAL)
└── scripts/                            # FROM THIS SKILL (new workflow)
    ├── Product_Model_KeyAngle.md
    ├── Product_Model_KeyAngle.md
    ├── Product_Model_KeyAngle.md
    └── Campaign_Summary.md

Integration with Other Skills

Workflow order:

1. tiktok-product-scraper → Product data + videos
2. tiktok-product-analysis → Image + video analysis (Gemini)
3. tiktok-script-generator → Scripts + summary (THIS SKILL - Claude)

Handoff verification:

# Before running this skill, verify:
date="YYYYMMDD"
ls product_list/$date/$product_id/ref_video/video_synthesis.md  # Must exist

Version: 2.3.0 Last Updated: 2026-01-31 Changes from v2.2:

• MAJOR PERFORMANCE OPTIMIZATION: Batched Write calls ⭐ 2x faster
• Parallel Read tool calls (all analysis files fetched at once)
• Parallel Write tool calls (4 files written simultaneously in one message)
• Workflow reduced from 5 steps to 4 steps (Steps 2-3 merged)
• Time per product: 2-3 min (was 5-8 min)
• 8 products: 16-24 min (was 40-64 min)
• Template extraction: Moved frontmatter, voiceover format, and campaign summary to separate template files
• Added template references throughout skill for better maintainability