podcast

Podcast Generator

Analyze sources, generate a Korean podcast script, produce audio via OpenAI TTS, and auto-upload to YouTube.

Pipeline

[Source Collection] → [Analysis/Fusion] → [Script Writing] → [TTS Generation] → [MP4 Conversion] → [YouTube Upload]

Step 1: Source Collection & Analysis

Collect and analyze user-provided sources. Processing by type:

• URL/Article: WebFetch or subagent for full text
• Tweet/X post: Use WebFetch with api.fxtwitter.com (replace domain in X/Twitter URL)
• PDF: Read tool directly
• GitHub repo: Clone and analyze structure (use subagent)
• Conversation context: Reuse content already analyzed in current session

When 2+ sources are provided, always spawn parallel subagents for each.

Step 2: Script Writing

Structure (8-12 min, 3000-5000 chars)

# [Episode Title]

> [Duration] podcast script | [Date]
> Sources: [source list]

---

## Opening (1 min)
- Hook: one sentence on why this topic matters
- Introduce sources
- Lead with conclusion (state core message upfront)

## Body Part 1 (3 min)
- Deep analysis of first source/perspective

## Body Part 2 (3 min)
- Deep analysis of second source/perspective

## Fusion/Intersection (3 min)
- Emergent insights from combining sources
- Patterns, commonalities, contrasts
- Generalizable implications

## Closing (30 sec)
- One-sentence summary of core message
- Sign-off

Script Writing Principles

• Write as you speak: conversational Korean ("~입니다", "~거죠", "~인데요")
• Numbers in Korean: "267K" → "이십육만", "$75,000" → "칠만오천 달러"
• English names in Korean pronunciation: "Garry Tan" → "개리 탄"
• No tables or code blocks: TTS cannot read them. Convert table content to sentences
• Shift tone for quotes: "개리 탄 본인이 이렇게 말합니다." to create distinction
• Short sentences: keep each sentence under 50 characters

File Layout

<output-dir>/
├── script.md       ← Script
├── episode.mp3     ← Audio
├── episode.mp4     ← Video (for YouTube)
└── metadata.json   ← Title, description, tags, YouTube URL

The output directory can be any user-specified path. A sensible default is podcast/YYYY-MM-DD-[slug]/ relative to the current working directory.

Step 3: TTS Generation

Convert script to audio using scripts/generate_tts.py:

python3 <plugin-path>/skills/podcast/scripts/generate_tts.py \
  --input <script.md path> \
  --output <episode.mp3 path> \
  --api-key <OpenAI API key>

Replace <plugin-path> with the actual path where this plugin is installed (use ${CLAUDE_PLUGIN_ROOT} if available, or the resolved plugin installation path).

OpenAI API Key

Check OPENAI_API_KEY environment variable first. If not set, ask the user.

TTS Settings

Setting	Value	Note
Model	gpt-4o-mini-tts	Latest model with instructions support
Voice	marin	Best for Korean. cedar as alternative
Chunk size	1500 chars	2000 token limit, Korean ~1.5 char/token
Instructions	Auto-generated per script	See default below

Default TTS instructions:

"따뜻하고 친근한 한국어 팟캐스트 호스트. 명확한 발음으로 또박또박 읽되, 자연스러운 억양과 적절한 감정을 담아서. 중요한 포인트에서는 약간 힘을 주고, 인용구에서는 톤을 살짝 바꿔서 구분감을 준다. 전체적으로 지적이면서도 편안한 분위기."

If the user specifies a tone, customize via --instructions.

Step 4: MP4 Conversion

Convert MP3 to MP4 with a static title card:

python3 <plugin-path>/skills/podcast/scripts/convert_mp4.py \
  --input <episode.mp3 path> \
  --output <episode.mp4 path> \
  --title "Episode Title" \
  --subtitle "Subtitle"

Generates a 1920x1080 video with dark background (#1a1a2e) and Korean title/subtitle overlay.

Step 5: YouTube Upload

python3 <plugin-path>/skills/podcast/scripts/upload_youtube.py \
  --video <episode.mp4 path> \
  --title "Episode Title" \
  --description "Description" \
  --privacy unlisted

OAuth Setup

• Google OAuth client secret: auto-discovers ~/Downloads/client_secret_*.json or ~/.config/google/client_secret_*.json
• Token: stored alongside the video file by default (override with --token-path)
• First run requires browser-based Google authentication
• Ask user which YouTube account to use if multiple are available
• Never copy scripts to the episode directory. Always run from the plugin's original path

Upload Defaults

• Privacy: unlisted (unless user specifies otherwise)
• Category: People & Blogs (22)
• Language: ko

Step 6: Completion Report

After upload, report to user:

Done!
- Script: <path>/script.md
- Audio: <path>/episode.mp3
- Video: <path>/episode.mp4
- YouTube: https://youtu.be/VIDEO_ID (unlisted)

Play episode.mp3 with afplay so the user can listen immediately.

Partial Execution

Users may request only part of the pipeline:

• "Just write the script" → Steps 1-2 only
• "Generate TTS from this script" → Step 3 only
• "Upload to YouTube" → Step 5 only (requires existing MP4)
• "Make it public" → Update YouTube privacy via API

Requirements

• ffmpeg: required for audio merging and MP4 conversion. On macOS, homebrew-ffmpeg/ffmpeg tap may be needed for full codec support
• OpenAI API key: for TTS generation (OPENAI_API_KEY env var or provided by user)
• Google OAuth client secret: for YouTube upload (download from Google Cloud Console)
• macOS font: uses /System/Library/Fonts/AppleSDGothicNeo.ttc for Korean text overlay. On other platforms, adjust FONT_PATH in convert_mp4.py
• Python 3.10+: all scripts use standard library only (no pip install needed)