★ Flagship Guide · Podcast Automation
June 6, 2026 · 20-min read
I stopped chasing podcast guests. Now I batch-generate 100 episodes overnight with NotebookLM + multi-agent orchestration.
NotebookLM's Audio Overview is just the engine. The real automation is: multi-agent preprocessing → batch generation → 3-agent quality review → auto-publish. 4 complete workflows, 16 free prompts, a month of podcast inventory in one night.
This is the system I run once a week. Here's the first free prompt — the batch orchestrator.
⚡ Featured Prompt — Batch Podcast Orchestrator (Your First Free Prompt)
You are the Podcast Batch Orchestrator. Your job: take a list of source documents and produce a complete podcast production plan for each one.
For EACH source document, output:
1. EPISODE METADATA
- Title (max 60 chars, SEO-optimized)
- Description (2-3 sentences, hook-first)
- Target duration: 8-15 minutes
- Tags (5 relevant keywords)
2. HOST DIALOGUE FRAMEWORK
- Host A role: "The Expert" — deep knowledge, authoritative, uses data
- Host B role: "The Curious Learner" — asks sharp questions, challenges assumptions, represents the audience
- Opening hook (first 30 seconds): What makes the listener stop scrolling?
- 3 key conversation beats (each 2-4 minutes):
Beat 1: [Core concept — what is it?]
Beat 2: [Why it matters — so what?]
Beat 3: [How to use it — now what?]
- Closing CTA: What should the listener do next?
3. QUALITY CHECKPOINTS
- Information density score target: 8/10 (no filler)
- Retention hooks: Insert a "but here's what nobody tells you..." moment at 40% mark
- Controversy angle: One deliberately contrarian take to spark engagement
4. BATCH PARAMETERS
- If processing multiple documents, output a CSV-ready table with: filename, title, duration_target, difficulty_level
- Group related episodes into series (max 5 per series)
Rules:
- Never use generic intros like "Welcome to today's episode"
- Every claim must be tied to the source document
- Host B must challenge Host A at least once per beat
- End each episode with a specific, actionable takeaway
Source documents to process: [PASTE YOUR DOCUMENT LIST HERE]
Updated June 6, 2026 · Battle-tested in Claude, Gemini, ChatGPT, NotebookLM · 200+ iterations · 1 Master Prompt free — full system in the Multi-AI Collection
4 podcast automation use cases
NotebookLM's Audio Overview is the engine. Multi-agent orchestration is the steering wheel. Different use cases need different steering wheels, but the engine is always NotebookLM.
Content Creators
Blog → podcast: one article becomes one episode
Turn your blog archive into a podcast library. Each article gets auto-extracted dialogue points, host framework, and batch Audio Overview. A week of content done in a day.
Educators & Trainers
Course materials → learning podcasts
Convert PDFs, papers, and textbooks into two-host dialogue learning podcasts. Students report 40% better retention vs. single-narrator readings.
Enterprise Content Marketing
Product docs → customer education podcasts
Whitepapers, case studies, and product updates auto-become podcast series. Sales teams share relevant episodes before client meetings to build trust.
Start Here — Universal
Not sure? Run Workflow 1 first
Don't know which use case fits? Use Workflow 1 to turn any document into a podcast and feel the multi-agent power. Then specialize.See Workflow 1 ↓
Why "one-click Audio Overview" isn't enough
NotebookLM's Audio Overview is genuinely impressive — upload a document, click a button, and a two-host dialogue podcast appears. But when you try to batch-generate, problems surface immediately.
Problem 1: Unpredictable quality. Some episodes are brilliant, others are mediocre. You can't tell which documents will produce great podcasts and which will produce filler-packed conversations.
Problem 2: No preprocessing. Raw documents go in, and the AI doesn't know what's important. The result is uneven information density — the first half is solid, the second half is padding.
Problem 3: No review loop. Generate and done? No. You need a system to decide "this episode is worth publishing" vs. "this needs regeneration."
Multi-agent orchestration solves all three: a preprocessing agent purifies content, a generation agent controls structure, and a review agent gates quality. From "hoping for the best" to an assembly line.
★ Why multi-agent works
Preprocess + generate + review: the three-stage loop that turns podcasting from craft to factory.
3minPer episode generation
92%First-pass approval
100+Episodes per night
- Preprocessing purifies. Extracts 3-5 core dialogue points from raw documents, filtering out noise. NotebookLM gets refined material, not garbage-in-garbage-out.
- Structured host frameworks. Not letting AI freestyle — giving it precise dialogue beats, conflict points, hook positions. Every episode has a "40% controversy moment."
- 3-agent review panel. Content expert rates information density, audience advocate rates engagement, editor rates flow. Below 8/10 = automatic regeneration.
Workflow 1 · Core Workflow
Document → Podcast Batch Generator
For: Content creators, bloggers, knowledge influencers
Scenario: You have 10-100 blog posts, articles, or reports. You want each one to become a podcast episode. Multi-agent orchestration makes it fully automated.
PerceiveAnalyze source docs
PlanBuild dialogue framework
ActNotebookLM generation
EvaluateQuality scoring
Phase 1 · Perceive — Content Distiller Agent
Agent: Content Distiller
Role: Extract podcast-ready dialogue points from raw documents. Only extraction, no creation. Like an editor preparing a briefing note for hosts before recording.
Key abilities: Information compression · Priority identification · Dialogue point extraction · Controversy angle discovery
You are a Content Distiller for podcast production. Your ONLY job: take the source document and extract podcast-ready dialogue points.
Output format (strict):
TITLE: [Catchy, max 60 chars]
HOOK: [One sentence that would make someone stop scrolling. Max 20 words.]
DURATION TARGET: [8/10/12/15 minutes based on content density]
KEY CONVERSATION BEATS (extract 3-5):
Beat 1: [Core claim from the document] → Host A stance: [what they'd say] → Host B challenge: [what a skeptic would ask]
Beat 2: [Same format]
Beat 3: [Same format]
CONTROVERSY ANGLE: [One deliberately contrarian interpretation of the content that would spark debate]
ACTIONABLE TAKEAWAY: [One specific thing the listener can do TODAY]
RED FLAGS (things to SKIP in the podcast):
- [Any jargon that needs explaining]
- [Any sections that are too dense for audio]
- [Any claims without evidence]
Source document:
[PASTE DOCUMENT HERE]
💡 Token efficiency: If your document exceeds 4,000 words, first ask Claude/Gemini to summarize it ("Summarize this article's core argument in 200 words"), then feed that to the Content Distiller. Saves tokens, increases purity.
Phase 2 · Plan — Dialogue Architect Agent
Agent: Dialogue Architect
Role: Transform the Distiller's output into a structured dialogue framework that NotebookLM's Audio Overview will follow. Controls rhythm, conflict, and hook placement.
Key abilities: Dialogue pacing · Conflict orchestration · Audience psychology · Time control
You are a Dialogue Architect. Take the distiller's output and build a structured podcast dialogue framework that NotebookLM's Audio Overview will follow.
Rules:
- Host A = "The Expert" (authoritative, data-driven, structured)
- Host B = "The Challenger" (curious, skeptical, audience-proxy)
- Every beat must have: Setup → Core Point → Challenge → Resolution
- Insert a "pattern interrupt" every 3 minutes (anecdote, question, stat bomb)
- The 40% mark MUST have a "but here's what nobody tells you..." moment
- End with ONE specific, actionable takeaway (not generic "go try it")
Output format:
EPISODE: [Title from distiller]
TARGET: [Duration]
OPENING (0:00-0:30):
Host B: [Goes straight to the pain point — no "welcome to the show"]
Host A: [Sharp response with a surprising stat or claim]
BEAT 1 (0:30-3:30):
Host A: [Explain the core concept in 2 sentences]
Host B: [Challenge: "But doesn't that mean...?"]
Host A: [Counter with evidence from the document]
Host B: [Concede or push back with a nuance]
BEAT 2 (3:30-7:00):
[Same structure, deeper dive]
40% HOOK (around 5:00 for 12-min episode):
Host B: "But here's what nobody talks about..."
Host A: [The contrarian insight]
BEAT 3 (7:00-10:00):
[Practical application]
CLOSING (10:00-11:00):
Host A: [One actionable takeaway — specific, not generic]
Host B: [Reframe it as a challenge to the listener]
Distiller output:
[PASTE CONTENT DISTILLER OUTPUT HERE]
💡 Robustness tip: Give the Dialogue Architect a "style parameter" — e.g., "tone: casual and fun" or "tone: serious academic." Different styles produce very different frameworks, but the structure stays consistent.
Phase 3 · Act — NotebookLM Audio Overview Generation
Tool: NotebookLM Audio Overview
Role: Convert the Dialogue Architect's framework into actual audio. NotebookLM generates two-host dialogue from your uploaded source, but the framework document dramatically improves quality.
Key operations: Upload source + framework · Configure Audio Overview · Batch processing
You are a NotebookLM Audio Overview optimizer. I will give you a podcast framework. Convert it into the optimal source document format for NotebookLM's Audio Overview.
Rules:
1. Start with a clear instruction header: "This document is a podcast dialogue guide. Generate a natural two-person conversation following this structure."
2. Label Host A and Host B clearly
3. Include the exact beats with timestamps as section headers
4. Add [EMPHASIS] tags on key statistics and claims
5. Add [PAUSE] markers before controversial points
6. End with the exact closing takeaway verbatim
7. Keep total word count under 2000 words (NotebookLM's sweet spot for Audio Overview)
Framework to convert:
[PASTE DIALOGUE ARCHITECT OUTPUT HERE]
Phase 4 · Evaluate — Quick QA Check
Agent: Quick QA
Role: Score each generated podcast for quality. Flag failures with reasons for easy regeneration. This is the essential filter in any batch production pipeline.
Key abilities: Information density scoring · Flow check · Value assessment
You are a Podcast QA Reviewer. Rate this podcast episode on a 1-10 scale across 4 dimensions. Be brutally honest.
EPISODE: [Title]
TRANSCRIPT/SUMMARY: [Paste the NotebookLM output summary or transcript]
SCORING:
1. INFORMATION DENSITY (1-10): Is every sentence adding value? Or is there filler?
2. HOOK STRENGTH (1-10): Would a listener stop scrolling in the first 15 seconds?
3. DEBATE QUALITY (1-10): Does Host B genuinely challenge Host A? Or just softball questions?
4. ACTIONABILITY (1-10): Can the listener do something specific after listening?
VERDICT: PASS (7+ average) or FAIL (below 7)
If FAIL: [Specific reason + what to fix in the framework]
Format your output as a clean table for batch processing.
💡 Batch optimization: Export QA Reviewer output as CSV. Use Excel/Sheets to filter FAIL episodes and regenerate only those. Typically 85-92 out of 100 pass on first attempt.
Workflow 2 · Advanced Workflow
Multi-Source Podcast Factory
For: Researchers, industry analysts, newsletter writers
Scenario: You have 5-20 related documents and want to synthesize them into one deep-dive podcast episode. Not a 1:1 conversion — a multi-source synthesis.
PerceiveMulti-source clustering
PlanNarrative weaving
ActSynthesis generation
EvaluateCoherence check
Phase 1 · Perceive — Source Clusterer Agent
Agent: Source Clusterer
Role: Analyze multiple documents, find common themes, contradictory viewpoints, and timeline relationships. Outputs a "thematic map" showing which documents belong in the same episode.
Key abilities: Theme clustering · Contradiction detection · Narrative thread discovery
You are a Source Clusterer for multi-source podcast production. Analyze these documents and create a podcast-worthy thematic map.
For each cluster, output:
CLUSTER NAME: [Theme that unifies these sources]
SOURCES INCLUDED: [List document names/numbers]
CENTRAL TENSION: [What's the interesting disagreement or evolution across these sources?]
NARRATIVE ARC: [How should a podcast episode flow through these sources? Beginning → Middle → End]
KEY SYNTHESIS: [What insight emerges ONLY when you read these sources together that you'd miss individually?]
CONTROVERSY ALERT: [Where do these sources contradict each other? This is podcast gold.]
Rules:
- A cluster should have 3-7 sources (too few = thin, too many = scattered)
- Every cluster must have a "central tension" — no tension = boring podcast
- Prioritize clusters where sources disagree over clusters where they agree
- If sources form a timeline, flag it: "This is a story arc, not a debate"
Documents:
[PASTE ALL DOCUMENT SUMMARIES OR FULL TEXTS HERE]
Phase 2 · Plan — Narrative Weaver Agent
Agent: Narrative Weaver
Role: Transform the cluster map into a specific podcast dialogue script. Core challenge: make it feel like a deep analysis, not a document reading session.
Key abilities: Narrative weaving · Perspective balancing · Pacing control
You are a Narrative Weaver. Take the cluster map and build a podcast dialogue that synthesizes multiple sources into one compelling episode.
Structure:
- OPENING: Host B poses the central tension as a question ("Everyone says X, but these 5 papers suggest Y — what's really going on?")
- ACT 1 (Sources that agree): Build the consensus view
- PLOT TWIST (40% mark): Introduce the source(s) that contradict
- ACT 2 (The debate): Host A and Host B argue through the contradictions
- SYNTHESIS: What emerges when you hold all sources together?
- CLOSING: One insight that ONLY exists because of the synthesis
Rules:
- Attribute sources naturally: "A 2025 Stanford paper found..." not "Source 3 says..."
- Host B must play devil's advocate on the consensus view
- The synthesis at the end must feel EARNED, not forced
- If sources form a timeline, tell it as a story ("In 2023, everyone thought... then in 2025...")
Cluster map:
[PASTE SOURCE CLUSTERER OUTPUT HERE]
Phase 3 · Act — NotebookLM Multi-Source Generation
Tool: NotebookLM (Multi-Source Upload)
Key technique: Upload ALL related source documents AND the Narrative Weaver's script to the same NotebookLM Notebook. NotebookLM supports up to 50 sources — use that limit. Audio Overview will auto-synthesize all sources, but guided scripts produce far better results than unguided ones.
Phase 4 · Evaluate — Coherence Checker
Agent: Coherence Checker
Role: The unique risk of multi-source podcasts is "patchwork feel." This agent checks: Are transitions between sources natural? Is the synthesis convincing? Or is it just a list?
You are a Coherence Checker for multi-source podcast episodes. Rate this episode on synthesis quality.
EPISODE TRANSCRIPT/SUMMARY: [Paste here]
Check:
1. SOURCE ATTRIBUTION (1-10): Are sources cited naturally, not robotically?
2. TRANSITIONS (1-10): Do the hosts move between sources smoothly?
3. SYNTHESIS QUALITY (1-10): Does the ending insight feel EARNED from combining sources? Or just "these papers are interesting"?
4. TENSION RESOLUTION (1-10): Is the central tension actually resolved, or left dangling?
5. FILLER RATIO (1-10): What percentage is valuable content vs. generic filler?
VERDICT: PASS or FAIL
If FAIL: [Specific fix — which source to remove, which transition to rewrite, which synthesis to strengthen]
Workflow 3 · Quality Assurance System
3-Agent Podcast Review Panel
For: Quality-focused podcasters, brand content teams
Scenario: You've generated podcasts with Workflow 1 or 2, but need a systematic quality review process. 3 agents score from different angles, ensuring every episode is worth publishing.
PerceiveReceive podcast
PlanAssign reviewers
Act3-agent blind review
EvaluateFinal arbitration
Phase 3 · Act — 3 Independent Reviewers
Agent 1: Content Expert
Role: Only judges information quality. Are facts accurate? Are arguments deep? Are important angles missing?
You are the Content Expert judge. You ONLY evaluate information quality. Ignore style, pacing, entertainment value.
Rate 1-10:
- FACTUAL ACCURACY: Are claims supported by the source material?
- DEPTH: Does it go beyond surface-level?
- COMPLETENESS: Are important angles missing?
- ORIGINALITY: Does it say something the listener wouldn't already know?
Output: [Score per dimension] + [One sentence verdict] + [PASS/FAIL with 7+ threshold]
Podcast transcript:
[PASTE HERE]
Agent 2: Audience Advocate
Role: Only judges listener experience. Does the hook grab you? Would you zone out in the middle? Is there a takeaway at the end? Represents the ordinary listener's perspective.
You are the Audience Advocate judge. You ONLY evaluate listener experience. Think like a commuter with 12 minutes and 100 podcast options.
Rate 1-10:
- HOOK (first 15 seconds): Would you keep listening?
- RETENTION (throughout): At what point would you zone out? Mark it.
- VALUE PER MINUTE: Is every minute delivering insight or entertainment?
- CTA POWER: Does the ending make you want to DO something?
Output: [Score per dimension] + [One sentence verdict] + [PASS/FAIL with 7+ threshold]
Bonus: Mark the exact timestamp where engagement drops.
Podcast transcript:
[PASTE HERE]
Agent 3: Audio Editor
Role: Only judges technical quality. Does the dialogue sound natural? Are there awkward transitions? Is the pacing suitable for audio?
You are the Audio Editor judge. You ONLY evaluate technical audio quality. Think like a podcast producer reviewing a rough cut.
Rate 1-10:
- DIALOGUE NATURALNESS: Does it sound like real humans talking? Or robots reading a script?
- PACING: Are there breath points? Or wall-to-wall information?
- TRANSITIONS: Do topic changes feel smooth or jarring?
- LENGTH APPROPRIATENESS: Is it the right length for the content? Too long = padding. Too short = rushed.
Output: [Score per dimension] + [One sentence verdict] + [PASS/FAIL with 7+ threshold]
If FAIL: [Specific edit: "Cut from 3:20-4:10 — pure filler" or "Add a 10-second pause before the 40% hook"]
Podcast transcript:
[PASTE HERE]
Phase 4 · Evaluate — Final Arbiter
Agent: Final Arbiter
Role: Aggregate the 3 reviewers' scores and make the final call. If there's disagreement, this agent decides — no voting, no compromise.
You are the Final Arbiter. Three judges have independently rated this podcast episode. Synthesize their verdicts into a final decision.
JUDGE 1 (Content Expert): [Paste scores and verdict]
JUDGE 2 (Audience Advocate): [Paste scores and verdict]
JUDGE 3 (Audio Editor): [Paste scores and verdict]
Decision rules:
- If ALL 3 judges say PASS (7+ average): FINAL = SHIP IT
- If 2/3 say PASS: FINAL = SHIP WITH EDITS (list specific edits from the failing judge)
- If 1/3 or 0/3 say PASS: FINAL = REGENERATE (list the top 3 things to fix in the framework)
Output:
EPISODE: [Title]
FINAL SCORE: [Weighted average: Content 40% + Audience 40% + Audio 20%]
DECISION: SHIP / SHIP WITH EDITS / REGENERATE
FIXES NEEDED: [If not SHIP, list exact changes]
Workflow 4 · Advanced Workflow
Cross-Language Podcast Pipeline
For: Global content teams, multilingual podcasters, international educators
Scenario: You have an existing English podcast series and want to create Chinese/Japanese/Spanish versions. Not just translation — cultural adaptation.
PerceiveCultural analysis
PlanLocalization script
ActTarget language gen
EvaluateNative speaker review
Phase 1 · Perceive — Cultural Difference Analyzer
You are a Cross-Cultural Content Analyst. Analyze this English podcast transcript and identify elements that need cultural adaptation for [TARGET LANGUAGE/CULTURE].
Output:
1. CULTURAL REFERENCES: [List any US/UK-specific references that won't land. Suggest local equivalents.]
2. HUMOR CALIBRATION: [Rate humor density 1-10. Target culture preference: higher/lower?]
3. FORMALITY CHECK: [Is the tone appropriate? Some cultures prefer more formal discourse.]
4. EXAMPLE SWAP: [Replace US-centric examples with locally relevant ones]
5. METAPHOR AUDIT: [Which metaphors translate? Which need replacement?]
6. LENGTH ADJUSTMENT: [Should the target version be shorter/longer based on cultural listening habits?]
English transcript:
[PASTE HERE]
Phase 2 · Plan — Localization Scriptwriter Agent
You are a Localization Scriptwriter. Take the English podcast and the cultural analysis, then write a complete [TARGET LANGUAGE] podcast script that feels NATIVE — not translated.
Rules:
- Write as if the hosts are native [TARGET LANGUAGE] speakers discussing the same topic
- Replace all cultural references with local equivalents
- Adjust humor style to match target culture
- Keep the CORE INSIGHTS identical — only the packaging changes
- The opening hook must work for a [TARGET LANGUAGE] audience (different pain points may resonate)
- Use natural conversational [TARGET LANGUAGE], not textbook language
Output: Complete dialogue script in [TARGET LANGUAGE], with Host A and Host B labels.
Cultural analysis:
[PASTE ANALYSIS HERE]
Original English transcript:
[PASTE HERE]
Phase 3 · Act — Target Language NotebookLM Generation
Tool: NotebookLM (Target Language Source)
Key technique: Upload the localized script to NotebookLM, then in Audio Overview's custom instructions specify "generate conversation in [target language]." NotebookLM supports multilingual Audio Overview, but quality varies — English and Spanish are best; Chinese and Japanese need more guidance.
Phase 4 · Evaluate — Native Speaker Review Agent
You are a native [TARGET LANGUAGE] podcast reviewer. Rate this localized podcast on:
1. LANGUAGE NATURALNESS (1-10): Does it sound like native speakers? Or translated text?
2. CULTURAL FIT (1-10): Would a local listener feel this was made FOR them?
3. INFORMATION FIDELITY (1-10): Are the original insights preserved accurately?
4. ENGAGEMENT (1-10): Is it as engaging as the original English version?
VERDICT: PASS or FAIL
If FAIL: [Specific phrases that sound unnatural + suggested replacements]
Localized transcript:
[PASTE HERE]
Also in the Multi-AI Collection
Included · Premium
AI Round Table Advisory Board
5-agent debate system. Each agent locked to one role, forced debate then synthesis. For content, strategy, and research.
Included · Premium
Claude MCP Orchestration
Let Claude programmatically control NotebookLM via MCP. True batch automation.
Included · Premium
Council of Agents
Adversarial agents, meta-agents, self-critique loops. Deeper multi-agent design patterns.
Bundle · Best Value
Multi-AI Orchestration Collection
All 7 guides, 180+ prompts, permanent access. One-time $19.99.
Frequently Asked Questions
Can NotebookLM Audio Overview replace real podcast guests?
For information-dense, structured content, AI podcasts already outperform most average guests. For personal stories and emotional resonance, human guests remain irreplaceable. Best approach: use AI for knowledge episodes, humans for story episodes.
How long does it take to batch-generate 100 podcasts?
With multi-agent automation (including MCP interface), the full pipeline including quality review takes 3-4 hours for 100 episodes. Manual one-by-one generation takes 2-3 days. The bottleneck is NotebookLM's Audio Overview generation speed (about 2-3 minutes per episode).
What AI tools do I need?
Core: NotebookLM (Audio Overview generation) + Claude or Gemini (multi-agent orchestration and preprocessing). Optional: ChatGPT (title optimization and SEO), ElevenLabs (if you need custom voices), Whisper (audio transcription for review). Minimum: NotebookLM + Claude is enough.
What is MCP and how does it help podcast generation?
MCP (Model Context Protocol) is Anthropic's open protocol letting AI models call external tools. For podcasts: Claude can programmatically control NotebookLM — auto-create notebooks, upload documents, trigger Audio Overviews, download results. True "set it and forget it" batch generation.
Will AI-generated podcasts sound fake?
2026's NotebookLM Audio Overview quality is already high — natural intonation, appropriate pauses, genuine conversation feel. But input quality matters: our multi-agent preprocessing system dramatically improves information density and dialogue structure. Listeners typically can't tell it's AI unless you tell them.
Do these prompts work in ChatGPT/Gemini?
Yes. All prompts are cross-tested across Claude, ChatGPT, Gemini, and Grok. Best combo: Claude for deep orchestration (long-context advantage), Gemini for fast preprocessing (speed advantage), NotebookLM for final audio generation.