Novel to Multimedia Pipeline — From Chapter Draft to Douyin in One Session

Writing an AI novel is only step one. The real leverage comes when a single Claude Code session takes 150 chapters of raw markdown and converts them into multi-voice audiobooks, vertical video, and published Douyin content — with you clicking nothing but “publish.” This entry documents the complete pipeline built during a live session with the 探花书房·逆袭 project.

The Pipeline at a Glance

Chapter .md (raw text)
    │
    ├──[1]──→ Parse: split narration vs dialogue, detect speakers
    │
    ├──[2]──→ TTS: Edge TTS multi-voice (旁白/Qing/林晚/妈/爸)
    │
    ├──[3]──→ Audiobook MP3: concatenate + 1.5x speed
    │
    ├──[4]──→ Douyin Video: text frames + fade transitions (1080×1920)
    │
    └──[5]──→ Publish: Playwright fills form, you click "发布"

One command per step. No manual editing. No switching between apps.

Case Study: 探花书房·逆袭

Step 1: Multi-Voice Audiobook with Edge TTS

Why Edge TTS?

Edge TTS wins for prototyping — zero cost, zero setup, good enough quality to validate the concept before investing in premium voices.

Voice Configuration (The Key Decision)

Finding the right voice required iteration. We generated 15+ demo clips before landing on this configuration:

Key insight: Edge TTS voices are all “young professional” by default. To age a voice, drop pitch 5-8Hz and slow rate 10-12%. To differentiate characters on the same base voice, vary both pitch AND rate — changing only one makes them sound like the same person at a different speed.

The Parsing Challenge

Chinese webnovel dialogue uses "" (curly quotes). Speaker detection uses a simple heuristic:

# Look 30 chars before the quote for character names
def detect_speaker(context_before):
    for name, voice in CHAR_MAP.items():
        if name in context_before[-30:]:
            return voice
    return 'narrator'  # fallback

This works ~90% of the time for the novel’s writing style (speaker name almost always precedes dialogue). The 10% misattribution goes to narrator voice, which sounds natural enough.

Generation Pipeline

# Install
python3 -m venv /tmp/tts-env
/tmp/tts-env/bin/pip install edge-tts Pillow

# Generate chapters 1-20
/tmp/tts-env/bin/python batch_audiobook.py 1 20

Each chapter: parse → generate segments (8 parallel) → add silence gaps → concatenate → apply 1.5x speed → output MP3.

Output: ~10 minutes per chapter at 1.5x speed. 20 chapters in ~15 minutes.

Step 2: Douyin Vertical Video

Format Iteration (What We Learned)

The lesson: Start simple, iterate fast. v1 took 30 minutes to build, revealed the wrong assumption (short = better for Douyin), and v2 took another 30 minutes to pivot.

Video Design

┌──────────────────────┐
│  探花书房·逆袭        │  ← Faint watermark (opacity 20%)
│                      │
│                      │
│  十月一号傍晚，苏州到  │
│  太仓的末班大巴晃晃悠  │  ← White text, centered
│  悠地停在了浏河路汽车  │     44pt PingFang SC
│  站。                 │     Fades in 0.3s, out 0.3s
│                      │
│                      │
│                      │
│                      │
└──────────────────────┘
  1080×1920 / near-black bg (#0F0F19)
  Dialogue: slightly brighter bg (#161423)
  Speaker name below dialogue text

Generation

# Generate full-chapter videos for chapters 1-20
/tmp/tts-env/bin/python douyin_fullchapter.py 1 20

Each chapter: parse → generate audio per segment → create text frame images (Pillow) → render video segments with fade filter (ffmpeg) → concatenate → apply 1.5x speed.

Output: ~9-12 MB per chapter. 20 chapters total ~255 MB.

Step 3: Semi-Automated Publishing via Playwright

Why Semi-Automated?

Semi-auto is the sweet spot: Playwright handles the tedious form-filling, but you maintain control over when each video goes live.

How It Works

1. First time: Playwright opens a browser, you scan QR code with Douyin app to log in
2. Per chapter: Script uploads video, fills title + description + hashtags
3. You: Review in the browser, click “发布”

// Per-chapter upload (runs in Playwright MCP)
async (page) => {
  await page.goto('https://creator.douyin.com/creator-micro/content/upload');
  await page.waitForTimeout(1500);
  const fileInput = await page.locator('input[type="file"]').first();
  await fileInput.setInputFiles('<VIDEO_PATH>');
  await page.waitForTimeout(4000);
  const titleInput = page.locator('input[placeholder*="标题"]').first();
  await titleInput.fill('探花书房·逆袭｜第X章 <Title>');
  const editor = page.locator('.editor-kit-container').first();
  await editor.click();
  await page.keyboard.press('Meta+a');
  await page.keyboard.type('<Hook description> #小说推荐 #有声小说 ...');
  return 'ready';
}

Publishing speed: ~30 seconds per chapter (upload + fill + click). 20 chapters in ~10 minutes.

Description Strategy

Each chapter’s Douyin description follows a formula:

<One-line hook from chapter's most dramatic moment> + hashtags

Examples:

• “东北穷小子回到太仓，妈炖了排骨，爸的手指肿成紫色却不敢去医院…”
• “系统第一次过载。Qing的鼻子流出血来，视线模糊了三秒。”
• “她是谁？Qing在平江路的巷口再次遇到林晚，这次她没有笑。”

The hook should create curiosity without spoiling — make viewers start the video.

The Skill: /webnovel-publish

The entire pipeline is packaged as a Claude Code skill. In any future session:

/webnovel-publish          # Triggers the full pipeline
"发抖音"                    # Also triggers it
"做有声书 21-50"            # Generate audiobooks for chapters 21-50
"publish chapters 21-50"   # Full pipeline for a chapter range

The skill contains:

• Voice configuration table
• Chapter parsing logic
• batch_audiobook.py — audiobook generation script
• douyin_fullchapter.py — video generation script
• Playwright publishing workflow
• Description template

How LearnAI Team Could Use This

• Convert LearnAI tutorials or case studies into audio/video formats for mobile-first learners.
• Package repeatable publishing workflows as Claude Code skills for consistent media asset production.
• Prototype short-form course promotion clips from existing markdown lessons.

Real-World Use Cases

• Webnovel authors turning chapters into audiobooks and vertical videos.
• Educators repurposing lesson notes into narrated shorts or long-form mobile videos.
• Content teams batch-producing platform-specific posts from one canonical markdown source.
• Solo creators using semi-automated publishing while keeping manual approval before release.

What’s Next

Key Takeaways

1. Edge TTS is underrated — free, zero-setup, 6 Chinese neural voices with pitch/rate control. Good enough for MVP audiobooks.
2. Iterate format fast — our v1 (1-min clips) was wrong. We pivoted to full-chapter in 30 minutes. Ship, test, adjust.
3. Semi-automation is the sweet spot — Playwright fills forms, you click publish. Zero ban risk, 95% effort reduction.
4. One session, three outputs — the same chapter text becomes an MP3, a video, and a Douyin post. Write once, distribute everywhere.
5. Parallel agents for writing — 4 chapters at a time with detailed outlines = 48 chapters in ~2 hours. The outline is the bottleneck, not the writing.