podcast

claude-podcast-plugin

A Claude Code plugin that turns any spec, plan, or design doc into a podcast-style audio conversation — and optionally executes the spec first to ground the podcast in real data.

Two AI-generated voices discuss your artifact with humor, skepticism, and genuine exploration — like two friends riffing at a bar about something weird they found on the internet.

Install

claude plugin install claude-podcast-plugin@clsandoval

Or load locally during development:

claude --plugin-dir /path/to/claude-podcast-plugin

Skills

/podcast — Comment on a doc

Reads a spec or design doc and generates a podcast discussing it. The hosts react to the ideas, roast the hand-wavy parts, and get excited about the clever bits. No code is executed.

/podcast:podcast path/to/your-spec.md

Or just say:

make a podcast of docs/my-design.md

/investigate — Execute a spec, then podcast the results

Reads a spec with testable steps, actually runs them, collects real results (API responses, error messages, timing data, costs), then generates a podcast grounded in what happened. The hosts discuss real numbers, real failures, and real surprises — not hypotheticals.

/podcast:investigate path/to/your-spec.md

Or just say:

investigate this spec and podcast it

The investigate skill produces:

• An investigation report with per-step results, artifacts, and cost breakdown
• A podcast where the hosts narrate what happened with real data
• Optional Telegram delivery of the audio file

Requirements

• ELEVENLABS_API_KEY environment variable set with a valid ElevenLabs API key
• ffmpeg — for concatenating audio chunks
• jq — for parsing dialogue JSON
• curl — for calling the ElevenLabs API

Configuration

Override default voices and model via environment variables:

Variable	Default	Description
ELEVENLABS_API_KEY	(required)	Your ElevenLabs API key
PODCAST_VOICE_A	UgBBYS2sOqTuMpoF3BR0	ElevenLabs voice ID for Person A
PODCAST_VOICE_B	aGv5jHWKBy8K5xKvYeSX	ElevenLabs voice ID for Person B
PODCAST_MODEL	eleven_multilingual_v2	ElevenLabs TTS model

Output

Files are saved to podcasts/ in your current working directory:

/podcast:

• podcasts/<name>.mp3 — the audio
• podcasts/<name>-transcript.md — readable transcript with speaker labels

/investigate (all of the above, plus):

• podcasts/<name>-investigation/report.md — structured investigation summary
• podcasts/<name>-investigation/cost-summary.json — cost breakdown by service
• podcasts/<name>-investigation/steps/ — per-step artifacts, logs, and outputs

How it works

1. Claude reads your artifact and generates a natural dialogue as JSON
2. (/investigate only) Claude executes each testable step in the spec, collecting real results
3. The generate.sh script sends each line to ElevenLabs TTS with the appropriate voice
4. ffmpeg concatenates the audio chunks into a single MP3

Caching: Generated audio chunks are cached in ~/.cache/podcast-tts/ (or $XDG_CACHE_HOME/podcast-tts/). If a run fails partway through or you regenerate with the same dialogue lines, cached chunks are reused — no wasted ElevenLabs credits.

Retries: API errors trigger up to 3 retries with exponential backoff (2s, 4s, 8s). A 0.3s delay between calls avoids rate limiting on long episodes.

License

MIT