book-to-podcast

name	book-to-podcast
description	Convert a non-fiction epub or pdf into a 45-minute two-host podcast (Empire-style co-hosts), generate audio via ElevenLabs or Sarvam, and publish to a private RSS feed on Cloudflare R2. Use when the user says "make a podcast from this book", "convert epub to podcast", "/book-to-podcast", or hands you an .epub/.pdf with intent to listen on commute.

book-to-podcast

End-to-end pipeline: book file → 45-min two-host script → ElevenLabs or Sarvam MP3 → private RSS feed on Cloudflare R2 → user subscribes once in any podcast app.

When to invoke

User says any of:

• "make a podcast from this book"
• "turn this epub into a podcast"
• "/book-to-podcast "
• drops an .epub or .pdf path with clear intent to consume it as audio

One-time setup (check before first run)

The skill expects ~/.claude/skills/book-to-podcast/.env to exist with:

TTS_PROVIDER=sarvam
DEFAULT_FORMAT=monologue

ELEVENLABS_API_KEY=...      # only if using elevenlabs
HOST_A_VOICE_ID=...
HOST_B_VOICE_ID=...

SARVAM_API_KEY=...
SARVAM_HOST_A_SPEAKER=anushka
SARVAM_HOST_B_SPEAKER=abhilash
SARVAM_NARRATOR_SPEAKER=arya

R2_ACCOUNT_ID=...
R2_ACCESS_KEY_ID=...
R2_SECRET_ACCESS_KEY=...
R2_ENDPOINT=https://<account>.r2.cloudflarestorage.com
R2_BUCKET=podcast
R2_PUBLIC_BASE=https://pub-<random>.r2.dev

If .env is missing, walk the user through copying config.example.env, picking two ElevenLabs voices, and choosing a random unguessable slug (e.g. python -c "import secrets; print(secrets.token_urlsafe(12))").

Python deps (install once into a venv at ~/.claude/skills/book-to-podcast/.venv):

ebooklib pypdf beautifulsoup4 elevenlabs pydub feedgen python-dotenv

No Quarto setup needed. MP3s and feed.xml go directly to Cloudflare R2 via publish.py. The subscribe URL is $R2_PUBLIC_BASE/feed.xml.

Workflow

1. Resolve input path. Fail clearly if not .epub or .pdf.
2. Ask the user three quick questions before proceeding:
   • Format: monologue (single narrator) or conversation (two-host Empire style)?
   • Provider: Sarvam (cheap, Indian voices) or ElevenLabs (premium)?
   • Language: English (default) / Hindi (हिंदी) / Kannada (ಕನ್ನಡ) / Tamil (தமிழ்) / Telugu (తెలుగు) / other Indian language?
     • Non-English + Kokoro → block: "Kokoro does not support Indian languages. Use Sarvam."
     • Non-English + ElevenLabs → warn: "ElevenLabs multilingual quality for Indian languages is inconsistent. Sarvam is recommended — continue anyway?"
     • Non-English + Sarvam + SARVAM_MODEL=bulbul:v2 → warn: "bulbul:v2 has limited speakers. Set SARVAM_MODEL=bulbul:v3 in .env for best quality."
3. Run scripts/extract.py <input> → prints title, author, chapter list, word count, and writes out/<slug>/book.txt.
4. Generate the script yourself (do NOT shell out — use Claude's own context):
   • Read out/<slug>/book.txt.
   • For conversation mode: read prompts/script-empire.md. Format every line as [Host A]: ... or [Host B]: ....
   • For monologue mode: read prompts/script-monologue.md. Format every line as [Narrator]: ....
   • If language ≠ English: apply the "Indian language output" section at the bottom of the relevant prompt. Write the entire script in the target language.
   • If word count > ~110k words, use the two-pass fallback described in the prompt template.
   • Write the script to out/<slug>/script.md. Target 6,500–7,200 words.
   • Run an anti-AI-tells pass. Use writing-anti-ai skill if available; otherwise self-edit.
5. Run scripts/chunk.py out/<slug>/script.md [--max-chars N]:
   • For ElevenLabs: default --max-chars 4500.
   • For Sarvam (English): --max-chars 500.
   • For Sarvam (Indian language): --max-chars 450 (extra margin for multibyte characters).
6. Run scripts/tts.py out/<slug>/chunks.json out/<slug>/episode.mp3 [--provider P] [--language L]:
   • Pass --language <code> when language ≠ English (e.g. --language hi for Hindi).
   • Language codes: hi Hindi · kn Kannada · ta Tamil · te Telugu · bn Bengali · ml Malayalam · mr Marathi · gu Gujarati · pa Punjabi · od Odia.
   • Cache is persistent and per-provider; safe to retry.
7. Run scripts/publish.py out/<slug>/episode.mp3 --title "<book title>" --author "<author>" --summary "<2-line blurb>". This:
   • Uploads MP3 to Cloudflare R2 bucket $R2_BUCKET.
   • Pulls existing feed.xml from R2, appends the new episode, pushes back.
   • Both MP3 and feed.xml live on R2 only — no GitHub commit needed.
8. Print the subscribe URL: $R2_PUBLIC_BASE/feed.xml. On first episode, instruct user to add it once in Pocket Casts / AntennaPod on Android.

Important constraints

• Do not ask before the TTS step. Sarvam runs are cheap (~₹40/book); ElevenLabs runs are pricier but the user picks the provider via env, not per-run.
• Output dir is $WORK_DIR/<slug>/ (from .env). Create if missing.
• If the book is fiction or under 30k words, warn the user — pipeline is tuned for non-fiction books of substantial length.

File layout

~/.claude/skills/book-to-podcast/
  SKILL.md
  config.example.env
  prompts/script-empire.md
  scripts/extract.py
  scripts/chunk.py
  scripts/tts.py
  scripts/publish.py

Output (per book):

$WORK_DIR/<slug>/
  book.txt
  script.md
  chunks.json
  episode.mp3