Back to blog
NotebookLM Now Lets You Guide AI Podcasters With Custom Prompts

NotebookLM Now Lets You Guide AI Podcasters With Custom Prompts

September 21, 2026

NotebookLM's Audio Overviews Now Support Custom Prompts: Here's What It's Actually Like

Google just quietly dropped a major update to NotebookLM's most popular feature. You can now guide the AI hosts of your "Audio Overviews" with specific instructions before they start talking.

Instead of sitting back and hoping the two AI podcasters focus on the right details, you can now tell them exactly what to do. Want them to focus on a specific sub-topic? Want them to talk like they are explaining it to a middle schooler or preparing you for a technical job interview? Now you can.

Here is what this update actually means, how it performs under stress, and where it still falls flat.

---

How the Custom Guided Audio Overviews Work

The core mechanic of NotebookLM remains the same. You upload your sources—PDFs, Google Docs, YouTube URLs, or markdown files—and it builds a localized knowledge base.

Previously, clicking "Generate" under the Audio Overview section kicked off an automated, highly convincing, but completely uncontrollable conversational podcast between two AI hosts.

With this new update, a text box appears before generation. Google calls it "Customize."

You get up to 500 characters to steer the conversation. Here is how the workflow plays out:

  1. Upload your corpus: Feed the notebook your raw data.
  2. Access Audio Settings: Under the Audio Overview generation panel, click the customization option.
  3. Write your prompt: Give the hosts their marching orders (e.g., "Focus entirely on the architectural trade-offs in chapter 4 and keep the tone highly technical").
  4. Generate: The model rebuilds the audio file from scratch based on your instructions.

---

Real-World Use Cases: Testing the Limits

We threw three distinct scenarios at the updated generator to see if the AI hosts actually listen or just nod and do their own thing.

Scenario 1: The Technical Interview Prep

  • The Prompt: "Focus exclusively on the system design bottlenecks. Act as an interviewer grilling a candidate. Keep it highly technical, skip the basic definitions, and challenge each other's assumptions."
  • The Result: Surprisingly good. The hosts skipped their usual "Hey there, welcome back to the channel!" intro and jumped straight into database partitioning. They used realistic developer jargon and even staged a minor disagreement over horizontal vs. vertical scaling.

Scenario 2: The "ELI5" (Explain Like I'm 5)

  • The Prompt: "Explain the core concepts of Kubernetes to a non-technical marketing manager. Use analogies involving shipping containers and toy stores."
  • The Result: The hosts adopted an incredibly enthusiastic, simplified tone. They leaned heavily into the "toy store" analogy. However, they occasionally slipped back into complex terms like "container orchestration" without defining them, breaking the illusion.

Scenario 3: The Deep-Dive Focus

  • The Prompt: "Ignore the introduction and history of the company in the PDF. Focus only on the Q3 financial metrics and why the gross margin dropped. Spend 90% of the time on this."
  • The Result: This is where the feature shines. Normally, NotebookLM podcasts spend the first three minutes summarizing the broad context. With this guide, the hosts spent exactly 20 seconds on the intro before dissecting the margin squeeze.

---

The Hard Limitations

While this is a massive step forward for utility, the technology still has obvious boundaries.

  • No Real-Time Mid-Podcast Corrections: You cannot pause the podcast and tell them to "explain that last point again." You must re-generate the entire file, which takes several minutes.
  • Hallucination of Tone: If you ask them to sound "angry" or "sarcastic," they struggle. The emotional range of the underlying Gemini audio model is still tuned to a cheerful, NPR-style default.
  • The 500-Character Limit: You cannot feed them a complex, multi-stage script. Your instructions must be concise, or you will run out of character budget.
  • Pronunciation Quirks: Specialized industry acronyms or obscure developer tools still get mangled occasionally. They will confidently mispronounce library names while sounding entirely sure of themselves.

---

The Bottom Line

This update shifts NotebookLM's audio feature from a neat party trick to an actual productivity tool. If you are trying to digest complex documents on your morning commute, being able to force the hosts to skip the fluff and focus on the data that matters to you is a massive time-saver.

It is not perfect, and you still cannot control the output with surgical precision. But as a way to quickly parse dry technical documentation? It is incredibly useful.