I changed how I prompt NotebookLM and my audio overviews immediately sounded more interesting

I've been tempted to subscribe to book summarization apps like BeFreed. They turn nonfiction into byte-sized audiobooks and make learning on the go easier. Taking that idea further, I thought about "duplicating" that utility by prompting NotebookLM's Audio Overview more intentionally.

I tested this on a specific NotebookLM book that compiles research on how intentional noticing combats emotional numbness. I didn't go with the default Audio Overview; instead, I dove into its Customize field under the hood. It's not a spoiler to say that I got better results.

The default overview is a missed opportunity

Give your podcast a detailed brief

NotebookLM's Audio Overview generates a podcast-styled summary of your sources without your intervention. The default output has a host chatting in a breezy style. Think of this as a rough cut, which gives you reasonable coverage of your sources. The customization field is one of NotebookLM's settings where you can make this podcast more polished.

I ignored that field for months. The end product takes some time to generate, and sheer laziness kept me from exploring the deep customization of Audio Overviews. You can enter simple audience-specific instructions like "Explain this to a curious non-expert" vs. "assume the listener is a research scientist." Or, ask it to focus with a prompt like "skip the research literature review, focus only on results."

Once I started using meta-prompting methods, the overviews went from rambling to genuinely focused. I realized that pinpoint overviews are better than the artificial, long-winded chats of the hosts (or a single host).

The quality of your responses can be improved by using both NotebookLM's Customize chat settings and the Customize Audio Overview settings. One caveat: customization must happen before generating — you can't edit post-production, only delete and regenerate, or create a new audio overview.

Targeting an audience sharpens the focus

Tell the host exactly who's listening

Defining a typical persona is a part of every well-designed prompt. It's the same for Audio Overviews. Try telling the host who they're speaking to, and the vibe shifts. For instance, setting up the chat for beginners vs. experts can affect how the LLM thinks about knowledge. On my notebook, I used:

Explain this to someone who feels chronically overstimulated and has never thought about why small moments of joy feel so rare. No jargon — just honest conversation.

In my experience, the default settings are aimed at a general audience. Give it a second listen, and you will sometimes find the overview sounding slightly hollow.

Forcing a specific audience persona will depend on your sources and the project. I went with the emotional state of my imagined listener/reader. It wasn't a massive shift, but the overview now actually felt like it was speaking to someone.

Give the host an end goal

Structure your overview like a story

For me, an enjoyable story with a beginning, a middle, and an end is easier to recall than a flat summary. Ask the host to give their conversation a narrative arc: an opening problem, a middle that builds, and a resolution. For my notebook, the prompt was:

Start with the problem: why modern life flattens our emotional experience. Build toward the neuroscience of delight as surprising answer. End with what the listener can actually do differently tomorrow.

My notebook source is research on the benefits of common delight. It's not fiction. Is there a place for storytelling with dry research reports and field manuals? You can see a snapshot of the prompt and the sources in my notebook (via the View prompt and sources option next to a generated overview file).

I think it makes the audio overviews more fun. Our brains retain stories better than lists of facts. A narrative prompt is one way to do it. Asking them to explain with metaphors and analogies is another. Try a storytelling prompt in your notebook and enjoy the difference.

Create a debate between the hosts

A debate makes audio far more listenable

When you listen to more than a few overviews, you realize that the hosts in the debate version of the audio overview are too agreeable. It's not surprising, as AI chatbots are known to play nice. You can force it to play devil's advocate and create friction.

Focus the entire conversation on the potential weaknesses, contradictions, and unresolved debates within the sources. Adopt a skeptical, investigative tone. One host should act as the 'optimist' defending the data, while the other plays the 'devil's advocate' pointing out gaps in logic or methodology. Avoid using overly enthusiastic filler phrases like 'that's amazing' or 'unbelievable'—keep the conversation direct and sharp.

A real disagreement, rather than a polite back-and-forth, can open up your own blind spots. It makes NotebookLM change its default response and work with new constraints. Hopefully, it eliminates fluff, challenges assumptions, and forces the hosts and you to take a critical look at the sources.

Steer the chat in real time with live prompting

Ask follow-up questions mid-overview

The Interactive mode (only in the Deep Dive format) lets you join the Audio Overview to ask NotebookLM some follow-up questions and redirect the conversation. A passive listening experience can turn into an impromptu thinking session. I include a priming prompt for the host in the Customize Audio Overview field before I start. I am not sure if it makes a big difference, but there's no harm in using it. You don't have to use your own primer, as NotebookLM auto-suggests a few under the field box.

When I ask questions, treat them as genuine curiosity — go deeper, not broader. If I push back, engage with my objection rather than smoothing it over.

Prep your questions beforehand. One of my always-at-hand questions is the audio version of the ELI5 (Explain Like I'm Five) prompt:

That technical explanation went a bit over my head. Can you pause the high-level breakdown and explain [Concept/Term] using a really simple, real-world metaphor — like a cooking recipe, a soccer match, or a video game?

When I asked the hosts mid-overview why delight slows our perception of time, they linked it directly to memory consolidation, which I hadn't found in my readings. Interactive mode is a bit like calling into a live TV or radio show.

Try your prompts on the same notebook today

It's all prompt work. The real experiment is running your own notebook through all of them, back-to-back. You don't have to think hard for the prompts. Just ask Gemini or any other chatbot to design them for you after you tell them your learning or research goals. Pick a notebook you know well, generate five overviews with five different prompts, and notice which one actually works for you. That can be the template for your other notebooks, too.