podcast

Podcast — Artifact-to-Audio

Convert a spec, plan, or design doc into a two-speaker podcast audio conversation.

Overview

Two friends co-investigating the artifact — warm toward each other, skeptical toward the material. Dry humor, real disagreement, no AI-podcast hype.

Output: dialogue.json, a transcript, an MP3 from Gemini TTS. Run locally, no Telegram delivery by default (remote flow handles that).

Invocation

/podcast <filepath>

The argument is a path to a markdown or text file.

Workflow

1. Read the artifact at the given filepath.

2. Identify the thread. What is the ONE non-obvious claim or thesis this piece makes? Structure the episode around that — not a linear tour.

3. Draft the dialogue as a JSON array at /tmp/dialogue.json:

   [{"speaker": "a", "text": "..."}, {"speaker": "b", "text": "..."}]
   

4. Verify before generating audio:

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/verify-dialogue.py" /tmp/dialogue.json
   

   If it fails, REWRITE the dialogue. Do NOT edit the script thresholds.

5. Save transcript to podcasts/<name>-transcript.md (create podcasts/ if needed) with **A:** / **B:** prefixes. <name> derived from the input filename (strip extension).

6. Generate audio:

   bash "${CLAUDE_PLUGIN_ROOT}/scripts/generate.sh" /tmp/dialogue.json podcasts/<name>.mp3
   

   For dialogues ≥8000 words, prepend PODCAST_BATCH_SIZE=4 to avoid late-batch TTS timeouts. Also drop to 4 if prosody sounds rushed.

7. Report to the user: audio path, duration, transcript path, and a one-liner about what the hosts made of the artifact.

Tone & voice

Peg: Red Web / We Say Things. Two knowledgeable friends co-investigating, dryly amused together. They like each other. They've been having this conversation for years. Dry humor comes from noticing things together, not from zingers at each other.

Both speakers have expertise — not expert/novice. A and B each bring a different angle of skepticism. The interesting thing is the collision of their takes.

Default posture: skeptical pragmatism. Assume claims in the artifact are overstated, under-specified, or won't survive contact with reality. Job: stress-test the material, not celebrate it.

Required move: steelman → critique → calibrated close ("works if X, fails if Y, testable by Z").

Warm toward each other, skeptical toward the material. Blunt about claims ("that's a press-release claim, not a technical one"); never sharp-edged at the other speaker ("read past the first sentence" — NO).

Dialogue length

• ~1 min audio per 500 words of input.
• 150 words of dialogue per minute of audio.
• 1000-word spec → ~2 min episode. 3000-word doc → ~6 min.

Hard gates (enforced by verify-dialogue.py)

Check	Rule	Failure
Speaker balance	A and B each 45–55% of words	Rewrite until balanced
Banned phrases	None of the 16 AI-tell phrases	Remove each
Sycophancy	≤1 marker	Rewrite to drop compliments
Combat	0 markers	Soften tone toward speaker

The script exits non-zero on failure with a punch list. Fix the dialogue, re-run.

Rationalization table — STOP if you catch yourself thinking these

"This is fine because..."	Reality
"It's a technical topic, A should explain, B should ask"	B is an expert in an adjacent frame. Rewrite.
"The artifact has a strong POV"	Interrogate it anyway. Premises are claims.
"I need enthusiasm to hook the listener"	Specificity hooks. "Brilliant" and "wild" are AI tells.
"Pushing back feels combative"	Warm pushback is the whole point. "Mm, is it really X?"
"Sharpening dialogue raises the stakes"	Stakes come from the material. Soften the edge.
"Short spec, I'll just summarize it"	Summary ≠ conversation. Find the thread, argue it.
"Verification script is overkill for this"	Run it anyway. 5 seconds, catches real issues.

Red flags — STOP and rewrite

• "That's wild" / "tell me more" / "great point" / "brilliant" anywhere
• B's average turn length < 60% of A's
• Any "Host:" / "Guest:" / "Interviewer:" labels
• Celebration ending ("so excited to see where this goes")
• Linear tour of the spec instead of one thesis
• "Read past the first sentence," "I'll bring popcorn" — sharp edges at the other speaker
• Verification script not run, or run and edited thresholds to pass

All of these = rewrite, not ship.

Audio defaults

• Model: gemini-3.1-flash-tts-preview. Older gemini-2.5-* aren't better.
• Voices: Charon (A), Kore (B). Override via PODCAST_GEMINI_VOICE_A/B.
• Pace: natural — no [fast] tag. Override via PODCAST_PACE only if explicitly needed.
• Batch size: 8 default, 4 for long or rushed-sounding episodes.
• Requires GOOGLE_API_KEY, ffmpeg, jq, curl on the local machine.

Output

All files go to podcasts/ in the current working directory:

• <name>.mp3 — podcast audio
• <name>-transcript.md — readable dialogue

Do NOT commit the audio or transcript — the user decides what to do with them.

What this skill does NOT do

• Does NOT write a summary. This is a conversation.
• Does NOT produce clean transitions ("speaking of which," "on that note").
• Does NOT make B a reaction machine.
• Does NOT ship without verify-dialogue.py passing.
• Does NOT commit artifacts.