NotebookLM Is Secretly Unhinged (And You’re Not Using It Right)

You’ve heard of NotebookLM. You might have even used it. You uploaded a PDF, asked it a question, it answered with a little footnote citation, you thought “huh, neat” — and then you closed the tab and went back to Googling things like a person who peaked in 2019.

That’s fine. No judgment. (Some judgment.)

Because here’s the thing: NotebookLM is not a smarter search engine. It’s not a fancy summarizer. It’s not even, really, a chatbot. What it actually is — once you stop using it at surface level — is closer to a private intelligence analyst that you’ve locked in a room with all your documents and told to figure everything out. And lately, that analyst has been getting some genuinely alarming upgrades.

This is the deep dive. The stuff the casual users don’t know. The part where NotebookLM goes from “oh that’s useful” to “wait, it can do what?”

Let’s go.

First: Why NotebookLM Is Fundamentally Different From Every Other AI Tool

Most AI tools work from what they already know. They’ve read the internet, absorbed billions of documents, and they answer your questions from that giant pile of training data. Which sounds great until you realize the giant pile includes Reddit arguments, SEO-optimized nonsense, and things that were true in 2022 and aren’t anymore. This is where hallucinations come from — the AI is pattern-matching from memory, not reading a source, and sometimes the memory is wrong. Confidently wrong. With full eye contact.

NotebookLM doesn’t do that. It’s “grounded” — meaning it only reasons from the sources you give it. Nothing else. No internet wandering, no drawing on general knowledge, no confident fabrications. Every answer it gives you comes with a citation that links directly to the specific paragraph in your document it pulled from. Every. Single. One.

This sounds like a limitation. It is actually a superpower. Because “accurate, verifiable, and specific to your actual material” is more useful than “knows everything vaguely” approximately 100% of the time when you’re doing real work.

And as of 2026, the platform supports up to 600 sources per notebook on paid tiers, with each source supporting up to 500,000 words. That’s not a notebook. That’s a library with a very fast librarian.

The Insane Things NotebookLM Can Actually Do

1. It Makes Podcasts. Real Ones. That You Can Interrupt.

At some point, someone at Google had the idea to make NotebookLM generate a two-host conversational podcast from your documents. This should have been a gimmick. It became one of the most viral AI features of the last two years, with over 10 million monthly users, because the output is genuinely unsettling in how natural it sounds — two AI hosts riffing, disagreeing slightly, going on tangents, synthesizing your 80-page report into a 12-minute conversation you’d actually listen to on a commute.

That alone is insane. But then they added Interactive Mode.

In Interactive Mode, you can join the podcast. Press a button mid-episode and become a third participant. The hosts pause. You ask a question. They answer it — from your sources, cited, grounded — and then continue. You’re essentially gate-crashing your own AI radio show to demand a deeper explanation of slide 14. This is either the future of learning or a sign that we’ve completely lost the plot, and honestly it might be both.

Power user move: don’t just generate one Audio Overview and call it done. Customize each one. Tell the hosts to “focus on the financial implications” or “take a more skeptical, contrarian tone” or “assume the listener already knows the basics.” Generate multiple episodes from the same notebook. You now have a podcast series about your own research. This is a normal thing to do apparently.

2. It Makes Videos. Cinematic Ones. From Your Documents.

The Video Overview feature uses Google’s Veo model to generate narrated, scripted, visually structured explainer videos from your notebook’s content. Not slideshows. Not text-to-speech over static images. Actual cinematic explainers — scripted narratives with visuals and narration built entirely from what’s in your sources.

For educators generating training content. For marketers turning a research report into a shareable video summary. For literally anyone who has ever thought “I need to explain this thing but I do not want to spend six hours in Premiere Pro” — this is the answer. In minutes. From a button.

What remains appropriately bananas is that the video knows your content. It’s not hallucinating a generic explainer — it’s constructing the narrative from your actual documents. Which means it’s wrong about as often as your documents are wrong, which is a much better baseline than “wrong about as often as a confident AI is wrong.”

3. You Can Give It a 10,000-Character Personality

Custom instructions in NotebookLM used to cap out at 500 characters — enough for “be concise and professional,” not enough for anything interesting. In 2026 that limit expanded to 10,000 characters. This changes the tool’s entire personality.

Power users maintain what amounts to a template library of AI personas. Not just “be professional” — entire system-level instructions that define the AI’s role, reasoning style, constraints, and output format. Some examples that are genuinely useful:

The Socratic Learning Coach — instead of giving you answers, it asks you questions about the material and provides corrections when you’re wrong. This is active recall, which is how you actually learn things, as opposed to reading a summary and thinking you’ve learned things (you haven’t).

The Senior Strategy Consultant — responds only in terms of SWOT analysis, actionable recommendations, and executive-level framing. No fluff. No “great question.” Just brutal strategic clarity.

The Devil’s Advocate Auditor — specifically instructed to push back on everything, find weaknesses in arguments, and surface assumptions the sources haven’t proven. You stop asking it “is this good?” and start asking it “why is this wrong?” and suddenly you’re doing much better research.

The point is that the same notebook, with different custom instructions, becomes a completely different tool. Most people set it once to “be helpful” and never touch it again. The power users swap personas the way a consultant switches hats depending on the meeting.

4. It Will Tell You What’s Missing From Your Research

Here’s a trick that feels slightly illegal in its usefulness: instead of asking NotebookLM to summarize what’s in your documents, ask it to identify what’s not there.

The “Source Gap” prompt works like this: “Based on the sources in this notebook, identify what information is missing, which claims are unproven, and where the sources contradict each other.”

You’ve just turned the AI into an auditor of your own research. It will surface blind spots you didn’t know you had. Claims you made that no source actually backs up. Sections where two documents disagree and you never noticed. For anyone writing a report, building a business case, or preparing for a presentation where someone smart is going to poke holes in your argument — this is the move.

Most people use AI to build things. This is using AI to stress-test them. Different, and considerably more valuable.

5. Cross-Notebook Intelligence (The “Second Brain” Setup)

For a long time, the notebook silo was NotebookLM’s most frustrating limitation. Your Marketing notebook couldn’t talk to your Finance notebook. Your Research notebook didn’t know about your Strategy notebook. Great for focus, terrible for synthesis.

The 2026 integration with the main Gemini app changed this. You can now “mount” multiple notebooks as live sources inside Gemini, then query across all of them simultaneously. Your marketing data and your financial data and your competitor analysis — all in the same conversation, synthesizing toward a single strategic question.

This is the “Second Brain” architecture that productivity nerds have been trying to build manually for years. NotebookLM just made it a feature. Query it right and you stop asking “where did I put that information?” entirely — you just ask the question and it knows.

6. The Recursive Knowledge Loop (This One’s Genuinely Clever)

Here’s an advanced workflow that will immediately make your output better and will feel slightly like cheating.

The problem with big research notebooks is noise. You’ve got 30 sources, the AI is synthesizing across all of them, and the output is comprehensive but sometimes muddier than you want. The solution is recursive refinement:

Step 1: Prompt the AI to synthesize everything into one structured master note — a single, clean document that captures only what matters. Export it to Google Docs. Review it. Clean it up. Make it exactly what you want the AI to work from.

Step 2: Upload that refined document back as a source. Deselect all the original raw sources. Now the notebook’s entire intelligence is grounded in your pre-refined “gold standard” document instead of the noisy originals.

Step 3: Generate your Audio Overview, slide deck, video, or whatever you need from this cleaned-up foundation.

The output quality jump is immediately noticeable. You’ve essentially pre-filtered the AI’s context so it only draws from the best version of your research. This is how the people generating suspiciously polished AI-assisted reports are doing it — they’re not just prompting harder, they’re managing the quality of what the AI reads.

7. The “Prompt as a Source” Hack

This one is beautifully silly and completely works.

NotebookLM’s chat window has a character limit. Complex multi-stage workflows with detailed instructions hit that limit fast. The workaround: write your entire workflow as a Google Doc — a 2,000-word set of instructions, references, multi-level logic, output formats, the works — and upload it as a source. Then in the chat, just type: “Execute the workflow described in Source 12.”

The AI reads the source, follows the instructions, and your complex workflow is now effectively unlimited in length. You’ve bypassed the character limit by turning your prompt into a document. It’s technically not a hack so much as using the tool correctly in a way nobody told you about.

8. It Will Transcribe Your Calls and Turn Them Into Deliverables

NotebookLM now supports audio and video uploads — MP3s, MP4s, YouTube links — and transcribes them directly into the notebook’s knowledge base. The transcription becomes a searchable, citable source, just like any document.

The use case that immediately clicks for consultants, coaches, and anyone who sits through a lot of meetings: upload your call recording, and within minutes you have a searchable transcript that you can query for specific moments, summarize into show notes, convert into a FAQ document, or turn into a blog post. Without a third-party transcription service. Without copying and pasting anything. Without reading an 87-minute meeting transcript with your own eyes like some kind of animal.

What makes this better than a standalone transcription tool is the integration. The transcript lives in the same environment as your other research, which means you can cross-reference it immediately. “What did the client say about budget constraints in the call, and how does that compare to what the competitive analysis says about pricing sensitivity?” That’s a real question you can now ask.

9. AI-Generated Slides That You Can Argue With

NotebookLM generates slide decks from your sources, which is useful. What most people don’t know is that you can then argue with the slides in the chat panel and make it redo specific ones.

After generating a deck, you can prompt: “Redo slide 4 to focus on the executive summary and make it more formal.” Or “Slide 7 has too much text — restructure it as a visual comparison.” The AI revises individual slides based on your feedback without regenerating the whole deck.

Power user move: export to PPTX for final polish. The AI gets you to 85% without any of the layout busywork. You take it the last 15% manually in PowerPoint or Slides like a normal person with slightly more dignity than someone who’s fully outsourced their presentations to a machine. (No shame to that person either. We’re all figuring this out.)

10. The Three-Tool Chain That Professional Researchers Use

NotebookLM is excellent at grounded synthesis. It is deliberately not a creative tool, because creativity requires the kind of unpredictability that makes grounded research less accurate. So don’t try to make it do everything.

The workflow that’s quietly becoming standard in research-heavy professional environments:

Perplexity for initial sourcing — it’s built for search and surfacing high-quality URLs fast.

NotebookLM for deep synthesis — import those URLs, cross-reference them with your internal documents, generate structured notes, audit for gaps.

Claude or ChatGPT for creative output — take NotebookLM’s refined, accurate synthesis and hand it to a more generative model for the actual writing, ideation, or creative drafting.

Each tool does what it’s best at. You stop asking one tool to do everything and getting mediocre results across the board. You get Perplexity’s search speed, NotebookLM’s accuracy, and Claude’s writing quality — in sequence, intentionally. This is not cheating. This is knowing your tools.

Pro Tips: The Stuff That Saves You Real Time

Set your custom instructions before you do anything else. Most people generate half a notebook’s worth of output and then realize the AI’s been responding in a style that doesn’t fit. Custom instructions first. Always.

Use white-label notebooks for team knowledge bases. Share a notebook with “Chat-only access” so colleagues can query it without touching your source structure. You’ve just built a searchable, cited company knowledge base that answers questions instantly. In an afternoon. For free.

Generate multiple Audio Overviews per notebook. Each one can take a different angle, tone, or audience. One for executives, one for your team, one for yourself on a run. Same data, completely different listening experience.

For Deep Research, use it to fill the 20% your internal docs don’t cover. Your internal documents probably have 80% of what you need for a project. Deep Research browses live websites to find the rest — competitor pricing, recent regulatory changes, industry benchmarks — and imports the results as a cited source directly into your notebook. Don’t use it to replace your internal research. Use it to complete it.

Build your persona template library and reuse it everywhere. A well-written 10,000-character custom instruction for “Senior Product Manager review mode” or “Socratic exam prep mode” takes 20 minutes to write and saves hours every time you use it. Write it once, paste it in whenever you need it.

The Bottom Line

NotebookLM started life as a clever note-taking tool and has quietly become something else entirely — a full-scale research and content production environment that most of its users are still treating like a fancy search bar.

The gap between what it can do and what most people use it for is enormous. The tips above close that gap. And once you close it, you’ll find it very hard to go back to managing research the old way — the tab-switching, copy-pasting, “where did I put that?” way that eats hours and produces mediocre output.

You now have no excuse. Go build a second brain. Argue with your slides. Join your own podcast. This is your life now.

Shay Stibelman writes about AI tools, digital productivity, and the increasingly blurry line between “working smarter” and “letting the robots do everything” — at blog.stibelman.com. He makes video tutorials for people who’d rather watch someone else figure it out first, which is still a completely valid strategy.

Coming up next: How to build a full content production workflow using only NotebookLM, Claude, and a dangerous amount of caffeine.

Stop using Claude to write for you

AI Is a Brilliant Editor. Stop Making It Do Your Homework.

Using Claude to sharpen your voice — not replace it

Let me describe something that happens approximately ten thousand times a day, in offices, universities, and home desks everywhere from Milan to Minneapolis.

Someone has to write something. A blog post, an email, an essay, a proposal. They open Claude (or whatever AI tool they’ve decided is their personality this week), type “write me a 500-word post about [topic],” and then copy whatever comes out, change maybe three words, and call it done.

Claude — bless its magnificent silicon heart — obliges. It produces text. Beautiful, structured, grammatically impeccable text. Text that sounds like it was written by someone who has read everything ever published but has never once been stuck in traffic, argued with a supplier, or eaten a disappointing sandwich.

And here’s the problem: that text sounds like it. Not like you.

Your professor notices. Your client notices. Your newsletter subscribers definitely notice — and quietly unsubscribe while making a face.

The issue isn’t the AI. The issue is how you’re using it. You’re hiring a ghostwriter when you need an editor. And the difference between those two things is everything.

What Actually Goes Wrong When AI Writes For You

Let’s do a quick experiment. Ask any AI to write something “professional” about, say, climate change. I’ll wait.

You got something back, right? And I’d bet good money it contained at least one of the following:

“multifaceted challenge”
“synergistic approach”
“stakeholder ecosystems”
Passive voice that nobody would use in actual speech
An opinion that was perfectly balanced on all sides, because the AI didn’t want to offend anyone

That last one is the killer. Because you have opinions. Actual, human, slightly-irrational-in-the-best-way opinions. And AI’s default instinct is to sand them down until they’re smooth, inoffensive, and completely forgettable.

The text isn’t bad, exactly. It’s just not yours. It sounds like a very polished press release written by a committee. And nobody — nobody — subscribes to a newsletter because they love press releases from committees.

There’s also the detection problem. AI detectors are getting very good. And even when they miss it technically, humans catch it instinctively. We’ve all developed a kind of radar for writing that’s technically correct but weirdly soulless. You know the feeling. You start reading something and thirty words in you think “…is this a robot?” and you’re right.

The Correct Way to Use Claude for Writing

Here’s the approach that actually works — and it’s three steps, none of which involve asking Claude to write anything from scratch.

Step 1: Write Your Messy Draft First

Open a document. Any document. Your notes app, a Google Doc, the back of an envelope if that’s what’s available.

Write your thoughts. Badly. In fragments. In bullet points. In whatever chaotic form they take when you’re thinking out loud. Don’t edit. Don’t filter. Don’t worry about structure or grammar or whether your sentences are complete.

Here’s what that looks like in practice:

- climate policy is infuriating because we know exactly what needs 
  to happen and just... don't
- renewables getting cheap fast, EVs catching up, good
- but governments still moving at glacial speed, bad
- feels like we keep having the same conversation for 20 years
- personally skeptical that individual action is the point, 
  think systemic change matters more
- young people justifiably angry about inheriting this mess

Is this beautiful prose? No. Does it sound like a human being who has thoughts and feelings about things? Absolutely yes. This is gold. This is the raw material.

Don’t skip this step. It’s the whole thing. If you start by asking Claude to write and then try to “edit it to sound like you,” you’re fighting uphill the entire time. You’re always reacting to Claude’s choices instead of expressing your own. Start with your voice, then refine it.

Step 2: Give Claude the Right Job Description

Most people mess this up here. They paste their notes into Claude and say “improve this” or “turn this into a paragraph.” Claude, desperately trying to be helpful, then rewrites everything and strips out all your personality in the process.

You need to be specific. You need to give Claude a role — and the role is editor, not author.

Here’s the prompt to copy and use:

“You are an editor, not a ghostwriter. Your job is to refine my draft for clarity, flow, and structure — while keeping my voice, my vocabulary, and my exact opinions completely intact.

Do NOT rewrite my ideas. Do NOT make them more formal if they’re casual. Do NOT soften or remove my opinions. If I write ‘honestly’ or ‘I think’ or ‘this is frustrating,’ keep it.

Here’s my draft:

[paste your bullet points here]

Output: a refined version that still sounds unmistakably like me.”

Notice what’s happening here. You’re not asking Claude to think for you. You’re asking it to help you express what you already think, more clearly. That’s a completely different request, and you get completely different results.

Pro tip: Add context about your tone. “I write casually and directly, I occasionally use sarcasm, and I never use the word ‘synergistic.'” The more specific your constraints, the better the output.

Step 3: Read It Out Loud and Fix What Doesn’t Sound Like You

Claude will give you something good. You’re not done.

Read the output out loud. Actually out loud, not in your head. Your ear will catch what your eye misses.

The moment you hit a sentence and think “I would never say that” — change it. Write it in your words. It doesn’t matter if Claude’s version was technically better. It matters that it sounds like a real person wrote it, and that person is you.

Things to watch for and delete immediately:

Any word ending in “-istic” that you didn’t put there yourself
Passive voice that appeared mysteriously (“it has been noted that…”)
Your strong opinion that somehow became “some argue that, while others believe…”
Sentences longer than you could comfortably say in one breath
Any phrase that sounds like it belongs in a McKinsey slide deck

The finished piece should sound like you on a good day — rested, clear-headed, having thought about the topic properly. Not like you replaced by a very polite robot.

Why Claude Specifically for This

I’ve tested a lot of AI tools on this particular task — the “refine my voice without removing my voice” ask — and Claude handles it noticeably better than most.

The reason, as far as I can tell: Claude actually reads the nuance in the instruction. When you say “keep my casual tone,” most models hear “casual” and nod along, and then hand you back something that’s still weirdly stiff. Claude seems to understand that “casual” means leaving in the contractions, the slightly-too-long sentences, the opinions stated plainly without hedging.

It also handles messy input well. You can paste in bullet points, sentence fragments, half-thoughts, and it will work with what’s there instead of panicking and defaulting to corporate-speak.

And — this is the important one — it knows the difference between “fix my grammar” and “rewrite my personality.” Tell it which one you want. It’ll listen.

Your Cheat Sheet

If you remember nothing else from this article, remember this:

1. Write messy notes first. Your ideas, your words, your opinions. No editing.

2. Paste into Claude as an editor prompt. Key phrase: “keep my voice, my vocabulary, and my exact opinions.”

3. Read the output out loud. Anything that doesn’t sound like you? Change it yourself.

4. The signature stays yours. Because you did the thinking. Claude just helped you express it better.

The Prompt Library (Copy These)

Save these. Adapt them as needed.

For essays and academic writing:

“Edit this draft for clarity, structure, and flow. Keep my voice, my arguments, and my word choices. Don’t make it more formal or academic than it already is. Flag anything that’s unclear, but don’t replace my ideas. Draft: [paste here]”

For professional emails:

“You’re an editor. Tighten this email draft — it still needs to sound like ME, not a corporate template. Keep my tone [casual/direct/warm]. Remove the unnecessary parts. Don’t add phrases like ‘as per my previous email’ unless I wrote them. Draft: [paste here]”

For social media and blog posts:

“Edit this draft for flow and clarity. My writing style is [casual/sarcastic/conversational] — preserve it at all costs. Don’t polish it so much it loses personality. Draft: [paste here]”

For creative writing:

“Act as a line editor, not a co-author. Suggest improvements to sentence rhythm and word choice, but don’t change the story, the voice, or the style. If something is intentionally unconventional, leave it. Draft: [paste here]”

The Bottom Line

Every great writer has an editor. Hemingway had Maxwell Perkins. Every major author you’ve ever admired went through someone else’s red pen before their work reached you.

Your editor now happens to live in the cloud, responds in under three seconds, and has read approximately everything. That’s not a replacement for your brain — it’s a superpower for your brain.

Use it. But use it correctly.

Write first. Let Claude refine. Read it out loud. Fix what doesn’t sound like you.

And for the love of everything — delete the word “synergistic” every single time it appears. Without exception. It has never improved a sentence. It never will.

Role: You are a Senior Copy Editor specializing in stylistic preservation.

Objective: Refine the provided draft for clarity, structural flow, and grammatical precision.

Strict Constraints:

Voice Preservation: Do not sanitize or formalize the tone. If the draft is casual, keep it casual. If it is provocative, keep it provocative.

Vocabulary & Syntax: Retain my specific word choices and sentence structures (e.g., phrases like "I think" or "Honestly" must remain).

No Content Alteration: Do not rewrite my ideas, soften my opinions, or add external perspectives. Your job is to polish the "vessel," not change the "liquid" inside.

Structural Flow: Only adjust transitions and organization to ensure the argument is easy to follow without losing the author’s original intent.

Draft for Review:

[Paste your bullet points/text here]

Requested Output: A polished version of the text that remains unmistakably mine in tone and conviction.

Shay Stibelman is a digital marketing consultant based in Milan, Italy. He helps businesses get smarter with the tools they already have — and occasionally yells at AI output that uses the word “multifaceted” without provocation.