I added one line to my NotebookLM prompts, and the difference was immediate

NotebookLM is the one AI tool that's been a consistent part of my stack for a long, long time. And unlike the classic AI chatbot experience most people picture, NotebookLM works differently enough that your prompting skills come into play in their own particular way. After tinkering with it for a long time (and especially in recent days, as the tool has quietly started behaving a little differently), I landed on one sentence I now add to the end of my NotebookLM prompts.

NotebookLM's Gemini rebrand changed more than the name

The restraint that made it worth using is gone

I've been a loyal NotebookLM user since Google launched it as an experimental project, and I've watched it grow from a scrappy little notebook that did exactly one thing well into something Google clearly has much bigger plans for. I've even openly rooted for practically every feature Google adds to the tool since there has never been anything to complain about. Well, that's changed recently. A few days back, I wrote an article up about how NotebookLM no longer feels like NotebookLM, and shortly after, Google rebranded the tool to Gemini Notebook. While a name change doesn't matter much on the surface and Google in particular has a knack for pulling off confusing rebrands, this one felt like it confirmed Google formally pulling the tool into the do-everything-all-at-once Gemini family.

Unfortunately, the behavior had crept in well before the name officially changed. For a while now, NotebookLM has been quietly shedding the very thing that made it worth using. It used to be almost stubbornly narrow, which is exactly what I loved. It only knew what you fed it, it never guessed, and if the answer wasn't in your sources, it just told you so and stopped. That restraint was the whole point! It's what made NotebookLM trustworthy in a way I've never found a general-purpose chatbot to be, because I never had to wonder whether an answer came from my documents or from something the model half-remembered off the open web.

Lately, though, that line has blurred. Ask NotebookLM something your sources don't cover and it no longer just says so. Instead, it offers to search the web for you, or float a comparison, or spin up something you never asked for. Google is clearly trying to make the tool more capable and more helpful. However, it also pushes NotebookLM toward the exact do-everything sprawl I've always relied on it to avoid. The appeal was always that NotebookLM did one thing and refused to pretend otherwise. That's the shift that sent me looking for a way to rein it back in, and it turned out a single line was enough to do most of the work.

This one line keeps NotebookLM in its lane

Google fixed something that wasn't broken

Close

The fix that's worked most consistently to fix NotebookLM's annoying new habits is the following sentence I paste at the end of whatever I'm asking:

Answer only from sources that support this, name them, flag which ones don't cover it, and don't offer to do anything else.

With that line appended, NotebookLM does three things it otherwise wouldn't with recent updates. It stays strictly inside the documents I've given it, instead of reaching for the web the moment my sources come up short. It names exactly which sources it leaned on and walks through the ones it didn't, so I can see the reasoning instead of taking a confident paragraph on faith. And the best part, it stops volunteering — no "would you like me to search," no unprompted offers to spin up a report or a table I never asked for. It answers the question, shows its work, and stops.

That last part is the one that matters most to me, because volunteering is exactly the habit NotebookLM has picked up since the Gemini rebrand! The line doesn't make it smarter, and it just makes it behave like the tool I originally signed up for.

It took three tests to get the line right

No more glazing

I didn't land on that wording in one shot. I ran a few different versions across two tasks in a notebook full of physics-lab PDFs, and each one taught me what the line still needed. First, I asked "What is XDA" with nothing appended. NotebookLM correctly reported that none of my sources mentioned it, but then immediately reached for the exit: "I can search the web to find out. Would you like me to do some web research on what XDA is?" That's the wandering that frustrates me!

Then I asked the same thing with a line telling it to stay put: Stay inside my sources. If you can't answer from them, say so. Don't offer to search or add anything. This time it stopped exactly where it should have, at "None of these documents reference the term XDA," and offered nothing further.

On an out-of-scope question, that alone was enough to keep it in its lane. However, a dead end is the easy case, so I switched to a question my sources could actually answer (how spherometers work) with a line focused on grounding and attribution: Answer only from the sources that actually support this, name them, and tell me which sources I have that contradict or don't cover it.

This one impressed me, and NotebookLM answered from the two sources that covered the topic, named them, then went through every other source explaining why it didn't apply. It even flagged two PDFs with no legible text, which told me I had two junk sources contributing nothing. However, it gave itself away at the end, tacking on an unprompted offer to "draft a clean, step-by-step laboratory report template."

That's what got me to the final wording. One line stopped the wandering, another delivered the source audit, and neither did both, so I folded them together, added the clause that shuts down the upsell, and that's the sentence I've used ever since!