# Podcast > Research a topic and produce a podcast episode with AI-generated voices. Use when user wants to create a podcast, audio episode, narrated discussion, or audio content from a topic or document. Triggers include "create a podcast", "make a podcast episode", "podcast about", "audio episode", "narrated discussion", "turn this into a podcast". - Skill: `rebyteai/podcast` (Agent Skill) - Install (CLI): `npx skillmds@latest add rebyteai/podcast` - Raw SKILL.md: https://api.skillmd.com/api/skills/rebyteai/podcast/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: rebyteai (https://skillmd.com/u/rebyteai) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/rebyteai/podcast --- # Podcast Produce podcast episodes from scratch or from source material. This skill handles content preparation, preview, and audio production end-to-end. ## Sub-Skills - `rebyteai/internet-search` — Quick web search for facts, quotes, and current data - `rebyteai/text-to-speech` — TTS synthesis (voices, style, dialogue) - `rebyteai/show-me-how` — Interactive widgets for the episode preview ## Workflow ### Step 1: Understand the Episode Parse what the user wants: - **Topic or source** — A topic to research, or a document/article to convert? - **Format** — Solo narration, two-host discussion, interview style, news roundup? - **Length** — Short (5 min, ~750 words), medium (10 min, ~1500 words), long (15+ min, ~2250+ words) - **Tone** — Conversational, educational, debate, storytelling, professional? - **Audience** — Technical, general, executive? ### Step 2: Research (if needed) Skip if the user provides source material (uploaded document, pasted text, etc.). - **News/current events** — Use `internet-search` for 3-5 targeted searches. - **Deep topic** — Use `internet-search` for comprehensive multi-source coverage. - **Debate/discussion** — Research both sides with `internet-search`. Organize findings into an outline: group by segment, note quotes/stats, identify narrative arc. ### Step 3: Write the Script Write a complete, natural-sounding script. Script quality determines podcast quality. **Script rules:** - Write for the **ear**, not the eye. Short sentences, contractions, conversational phrasing. - Avoid jargon unless the audience is technical. - Include transitions between segments. - Use `[SPEAKER NAME]` markers for each speaker on their own line. **Format by episode type:** **Solo narration:** ``` [HOST] Welcome to the show. Today we're diving into... [HOST] That's it for today. If you found this useful... ``` **Two-host discussion:** ``` [HOST A] So I've been reading about this new trend in... [HOST B] Yeah, I saw that too. What surprised me was... ``` **Interview:** ``` [INTERVIEWER] Tell us about your experience with... [GUEST] Well, it started when... ``` **Structure every episode with:** 1. **Intro** — Welcome, topic intro, what listeners will learn 2. **Body** — Main content in 2-4 segments with transitions 3. **Outro** — Summary, key takeaway, sign-off ### Step 4: Show Episode Preview (REQUIRED) **Before generating any audio, show the user a preview widget for approval.** Audio generation is expensive (TTS API calls, ffmpeg processing). The preview lets the user catch issues early. Generate a `show-me-how` widget that displays the full episode plan. The widget should include: 1. **Episode header** — Title, estimated duration, format (solo/discussion/interview/news) 2. **Cast** — Each speaker with their assigned voice and a short voice description 3. **Sound design** — What music/ambience will be used (e.g., "Intro: downloaded lo-fi track from Pixabay, Background: ocean waves, Outro: same as intro") 4. **Full transcript** — The complete script, styled with: - Speaker names as colored labels (different color per speaker) - The actual dialogue text - Structural markers (`[INTRO MUSIC]`, `[TRANSITION]`, `[OUTRO MUSIC]`) shown as visual dividers - Estimated timestamp for each segment **Widget template:** ```` ```widget

🎙️ Episode Preview: TITLE HERE

Review the episode plan before generating audio

~10 min
Duration
2
Speakers
Discussion
Format
3
Segments

Cast

HOST A marin — Female, warm, confident
HOST B cedar — Male, calm, authoritative

Sound Design

Intro MusicLo-fi podcast intro (Pixabay, 6s)
BackgroundSoft coffee shop ambience (0.2x volume)
TransitionsGenerated tonal sting (3s)
Outro MusicSame as intro (8s, fade out)

Transcript

🎵 Intro Music (6s)
HOST A 0:06
Welcome back to the show. Today we're looking at...
HOST B 0:32
Yeah, this is a fascinating topic because...
🔀 Transition (3s)
🎵 Outro Music (8s)
``` ```` **After showing the preview, ask the user:** > Here's the full episode plan. You can: > - **Continue** — I'll generate the audio now > - **Change voices** — e.g., "Make Host B use ash instead of cedar" > - **Edit the script** — tell me what to change > - **Change music/ambience** — e.g., "Use rain instead of coffee shop" or "No background ambience" > - **Adjust length** — e.g., "Make segment 2 shorter" **Only proceed to Step 5 after the user approves.** ### Step 5: Produce Audio Follow the **Audio Production Engine** section below. It handles: - Voice selection and pairing (Gemini multi-speaker TTS as primary) - Script chunking and synthesis (one API call per chunk, not per line) - Music download, ambience mixing, mastering - Fallback to per-line OpenAI `gpt-audio-mini` if Gemini fails ### Step 6: Deliver 1. Upload the final MP3 to the Artifact Store 2. Provide: - The audio file - The full script (so the user can review/edit) - Episode metadata: title, duration, segment breakdown, voices used - Sources cited (if research was done) 3. Ask if the user wants: - A different voice or pacing - Script edits before regenerating - Additional segments or a follow-up episode - A web player app (can build with `rebyte-app-builder`) ## Decision Points - **"Research or use provided content?"** — If the user uploads a document or pastes text, use that. If they give a topic, research it. Some need both. - **"How many voices?"** — Solo = 1, Discussion/debate/interview = 2. Default to solo unless specified. - **"How long?"** — Default ~10 minutes (~1500 words). News = 5 min. Deep dives = 15 min. - **"User wants a web player"** — Build with `rebyte-app-builder` and deploy to rebyte.pro. Only if asked. --- # Audio Production Engine Turn a script into a finished podcast episode. Uses Gemini multi-speaker TTS as primary (natural dialogue in one call), falls back to per-line OpenAI if needed. **Requires Rebyte API auth** — `$AUTH_TOKEN` and `$API_URL` from the system prompt. ## Pipeline ``` Script → format as "Speaker: line" dialogue → chunk at ~3500 chars → Gemini synthesize_dialogue per chunk (2 speakers per call) → concat chunk WAVs → download music/ambience → mix → master → MP3 ``` ### Setup ```bash for cmd in ffmpeg curl jq base64; do command -v "$cmd" >/dev/null || { echo "FATAL: $cmd not found"; exit 1; } done WORKDIR=$(mktemp -d /tmp/podcast-XXXXXX) mkdir -p "$WORKDIR/chunks" "$WORKDIR/assets" ``` ## Voice Selection ### Gemini Voices (primary — used with synthesize_dialogue) | Voice | Character | Best For | |-------|-----------|----------| | **Kore** | Female, warm, firm | Primary host, narration | | **Charon** | Male, deep, informative | Co-host, expert segments | | **Puck** | Male, upbeat, playful | Casual, comedy, chat | | **Aoede** | Female, breezy, light | Soft segments, intimate | | **Fenrir** | Male, excitable | High-energy, sports, hype | | **Leda** | Female, youthful | Bright explainer | ### Voice Pairing | Format | Recommended | Why | |--------|-------------|-----| | Two-host discussion | Kore + Charon | Warm female + deep male — distinct | | Interview | Kore + Puck | Firm host + upbeat guest | | Debate | Fenrir + Kore | Energetic vs. measured | | News roundup | Kore + Fenrir | Confident anchor + energetic reporter | ### OpenAI Fallback Voices (per-line synthesis if Gemini fails) | Voice | Character | |-------|-----------| | **openai:marin** | Female, warm | | **openai:cedar** | Male, authoritative | | **openai:ash** | Male, energetic | | **openai:coral** | Female, professional | ## TTS Synthesis ### Primary: Gemini Multi-Speaker Dialogue Format the script as `Speaker: line` text and call `synthesize_dialogue`: ```bash RESPONSE=$(curl -s -X POST "$API_URL/api/data/tts/synthesize_dialogue" \ -H "Authorization: Bearer $AUTH_TOKEN" \ -H "Content-Type: application/json" \ -d '{ "dialogue": [ {"speaker": "Host A", "text": "Welcome to the show. Today we are diving into..."}, {"speaker": "Host B", "text": "Yeah, this is a fascinating topic because..."}, {"speaker": "Host A", "text": "Let me share some numbers that surprised me."} ], "voices": { "Host A": "gemini:Kore", "Host B": "gemini:Charon" } }') echo "$RESPONSE" | jq -r '.audio.base64' | base64 -d > "$WORKDIR/chunks/chunk_001.wav" ``` **Constraints:** - Max **2 speakers** per call. For 3+ speakers, split by speaker pairs and concat. - Max **4096 chars** combined text. Chunk scripts at ~3500 chars at sentence boundaries. - Always returns WAV (24kHz PCM). Normalize: `ffmpeg -i in.wav -ar 44100 -ac 2 -sample_fmt s16 out.wav` **For solo narration**, use single-speaker `synthesize` instead: ```bash curl -s -X POST "$API_URL/api/data/tts/synthesize" \ -H "Authorization: Bearer $AUTH_TOKEN" \ -H "Content-Type: application/json" \ -d '{"text": "Welcome to the show...", "voice": "gemini:Kore"}' \ | jq -r '.audio.base64' | base64 -d > "$WORKDIR/chunks/chunk_001.wav" ``` ### Chunking Split scripts > 3500 chars at: 1. Scene/segment breaks (best) 2. Paragraph boundaries (`\n\n`) 3. Sentence boundaries (`.` `!` `?` followed by space + uppercase) Keep speaker pairs consistent across chunks. Concat chunks with ffmpeg. ### Fallback: Per-Line OpenAI If Gemini `synthesize_dialogue` fails (returns `no_audio` or API error), fall back to per-line OpenAI synthesis: 1. Split script into individual lines 2. Call `synthesize` per line with `openai:` and `gpt-audio-mini` 3. Generate 0.6s silence WAVs between speakers 4. Concat all with ffmpeg This is slower and more expensive but always works. ### Error Handling | Error | Action | |-------|--------| | `no_audio` from Gemini | Retry once. If still fails, fall back to per-line OpenAI | | `text_too_long` | Re-chunk smaller, retry | | `rate_limit` | Wait 5s, retry (max 3) | | TTS fails after all retries | **FATAL** — report error | ## Music & Ambience Download royalty-free audio from Pixabay (CC0): ```bash curl -L -o "$WORKDIR/assets/intro_raw.mp3" "" ffmpeg -i "$WORKDIR/assets/intro_raw.mp3" -ar 44100 -ac 2 -sample_fmt s16 "$WORKDIR/assets/intro.wav" ``` | Element | Duration | Fade In | Fade Out | |---------|----------|---------|----------| | Intro music | 5-10s | 1s | 2s | | Outro music | 5-10s | 2s | 3s | | Ambience | Full episode | 2s | 5s | If download fails, skip music/ambience and produce speech-only. Warn user. ## Assembly & Mastering ### 1. Concat speech chunks ```bash # speech_list.txt: file 'chunks/chunk_001.wav' \n file 'chunks/chunk_002.wav' ... ffmpeg -f concat -safe 0 -i "$WORKDIR/speech_list.txt" -c copy "$WORKDIR/all_speech.wav" ``` ### 2. Mix ambience (if available) ```bash DURATION=$(ffprobe -v error -show_entries format=duration -of csv=p=0 "$WORKDIR/all_speech.wav" | cut -d. -f1) ffmpeg -i "$WORKDIR/all_speech.wav" -i "$WORKDIR/assets/ambience.wav" \ -filter_complex "[0:a]volume=1.0[speech];[1:a]volume=0.25,afade=t=out:st=$((DURATION-5)):d=5[bg];[speech][bg]amix=inputs=2:duration=shortest" \ -ac 2 -ar 44100 "$WORKDIR/episode_with_ambience.wav" ``` ### 3. Add intro/outro ```bash # episode_list.txt: intro.wav, silence, episode_with_ambience.wav, silence, outro.wav ffmpeg -f concat -safe 0 -i "$WORKDIR/episode_list.txt" -c copy "$WORKDIR/episode_raw.wav" ``` ### 4. Two-pass loudnorm mastering (-16 LUFS) ```bash # Pass 1: measure STATS=$(ffmpeg -i "$WORKDIR/episode_raw.wav" -af "loudnorm=I=-16:TP=-1.0:LRA=7:print_format=json" -f null /dev/null 2>&1) INPUT_I=$(echo "$STATS" | grep '"input_i"' | grep -o '[-0-9.]*') INPUT_TP=$(echo "$STATS" | grep '"input_tp"' | grep -o '[-0-9.]*') INPUT_LRA=$(echo "$STATS" | grep '"input_lra"' | grep -o '[-0-9.]*') INPUT_THRESH=$(echo "$STATS" | grep '"input_thresh"' | grep -o '[-0-9.]*') TARGET_OFFSET=$(echo "$STATS" | grep '"target_offset"' | grep -o '[-0-9.]*') # Pass 2: apply ffmpeg -i "$WORKDIR/episode_raw.wav" \ -af "loudnorm=I=-16:TP=-1.0:LRA=7:measured_I=${INPUT_I}:measured_TP=${INPUT_TP}:measured_LRA=${INPUT_LRA}:measured_thresh=${INPUT_THRESH}:offset=${TARGET_OFFSET}:linear=true" \ "$WORKDIR/episode_mastered.wav" ``` ### 5. Encode MP3 ```bash ffmpeg -i "$WORKDIR/episode_mastered.wav" -codec:a libmp3lame -b:a 192k -ar 44100 "podcast-episode-${SLUG}.mp3" ``` Upload to Artifact Store. Clean up `$WORKDIR`.