export default podcast

name	podcast
description	Creates audio podcasts from text using browser text-to-speech. Use when user mentions podcast, audio conversation, dialogue, spoken content, voice narration, audio book, or text-to-speech generation. Supports multiple speakers with automatic language detection. Zero cost, no API keys, works in browser.
allowed-tools	Read, Write

Podcast Generator

Generates podcast-style audio that plays directly in the browser. Zero cost, no API keys needed.

Workflow Decision Tree

User provides formatted dialogue

→ Use existing dialogue as-is if quality is good → Refine structure only if flow needs improvement

User provides article, list, or text content

→ Create dialogue from content (see "Dialogue Creation Process")

User provides topic only

→ Request source material before proceeding

Dialogue Creation Process

Follow this two-phase workflow when creating podcast dialogue from content:

Phase 1: Analyze Source Content

1. Read source material completely
2. Detect language from content (en, de, fr, es, it, etc.)
3. Identify key information: facts, dates, names, numbers, details
4. Organize by theme: chronology, category, or logical grouping

Phase 2: Create Dialogue

1. Structure conversation:

   • Host (Speaker 1): ~20% - questions and transitions
   • Expert (Speaker 2): ~80% - factual responses from source
   • Length: as many exchanges as needed to cover all content (typically 10-50+ lines)

2. Apply TTS formatting: Read reference/tts-formatting.md for complete rules

3. Generate JSX file from template (see "Implementation Steps")

Information Accuracy

Convert text to audio. Use only source material facts.

Expert responses:

• Use only names, dates, numbers, details explicitly stated in source
• Never invent examples, context, or explanations not in source
• Never add interpretations, opinions, or evaluations
• Never explain WHY something happened unless source explains it

Host phrases:

• Use neutral transitions: "I see", "Tell me more", "Can you elaborate?"
• Reference previous statements: "You mentioned X - how does that connect to Y?" (when source shows connection)
• Never add new facts, context, or interpretations

Dialogue Format Guidelines

Host (Speaker 1) - ~20% of content

• Introduce topic with opening question
• Ask transition questions between topics
• Reference Expert's previous statements: "You mentioned X - can you elaborate?"
• Use conversational acknowledgments: "I see", "Tell me more"
• Never introduce facts not in source

Expert (Speaker 2) - ~80% of content

• Provide comprehensive factual responses from source material
• Include specific details: names, dates, numbers, locations
• Organize information logically by theme, chronology, or category
• Structure facts narratively using only source material
• Never repeat information already stated
• Never add context or examples not in source

Natural Conversation Techniques

Use these patterns without adding information:

• Vary question styles: "What happened next?" / "Can you explain that further?" / "Tell me about..."
• Ask follow-up questions based on Expert's previous response
• Expert elaborates when source provides multiple details about a topic
• Clear transitions between sections: "Moving to the next category...", "In the European context..."

Avoid

• Personal opinions: "I think...", "That's crazy..."
• Value judgments: "amazing", "fascinating", "interesting"
• Humor, irony, jokes
• Rapid back-and-forth after every sentence

Implementation Steps

When user requests a podcast:

1. Analyze source content and create dialogue following format above
2. Detect language from content
3. Read template from assets/podcast-template.jsx
4. Replace values in template:
   • PODCAST_SCRIPT - your generated dialogue
   • PODCAST_TITLE - descriptive title from content
   • PODCAST_LANGUAGE - detected language code
5. Save as JSX file - Use the Write tool to save the modified template as a .jsx file. The file will render as an interactive podcast player.
6. Recommend Microsoft Edge browser for best voice quality (250+ Natural voices vs Chrome's 19)

Technical Reference

Script Format

<speaker1>Host's question or statement.
<speaker2>Expert's response with factual information.

Voice Configuration (automatic)

• Speaker 1 (Host): Pitch 1.05, Rate 0.95
• Speaker 2 (Expert): Pitch 0.88, Rate 0.93

Platform-Aware Voice Selection

Platform detection:

• Automatically detects iOS, Android, Desktop Edge, or Desktop
• Selects best available voices based on platform

Desktop Edge:

• Priority: Microsoft Neural/Natural voices (Katja, Conrad, Aria, Guy, etc.)
• 250+ high-quality voices available

Desktop Chrome:

• Priority: Google voices (Google UK English Female, Google Deutsch, etc.)
• ~19 voices available (lower quality than Edge)
• Fallback: local system voices

iOS (Safari/Mobile):

• Priority: Native Siri voices (Samantha, Anna, Daniel, etc.)
• Best quality on iOS devices

Android (Chrome/Mobile):

• Priority: Google TTS voices (Google Deutsch, Google UK English Female, etc.)
• Wavenet voices preferred when available

Voice assignment:

• Automatically assigns different voices to Speaker 1 and Speaker 2
• Uses modulo distribution for 3+ speakers
• Ensures distinct voices even with limited availability

Player Features

• Play/Pause/Resume with full playback control
• Stop to reset to beginning
• Click any transcript line to resume from there
• Progress bar shows current position
• Auto-scroll follows current line

Technical Constraints

• Keep sentences under 14 seconds (Chrome limitation)
• 350ms pause between speakers
• Microsoft Edge browser provides 250+ high-quality Natural voices (best option)
• Chrome provides only 19 lower-quality voices with utterance bugs
• Firefox has very limited voice support

Quality Requirements

• Factual Accuracy: Expert responses use only source facts
• Natural Flow: Avoid rapid back-and-forth, value judgments
• TTS Compliance: All text must play without pronunciation errors
• Zero Hallucination: No invented examples or context
• Complete Coverage: Include all important facts from source
• No Duplicates: Each fact appears exactly once