NotebookLM Phials
Yesterday
While browsing Twixxer (formerly known as "Prince"), I ran across a thread purporting to be a slew of hyperuseful prompts for NotebookLM.
A few things struck me at once:
The thread had a lot of traction with a few hundred retweets but only around 40 or 50 responses. Which means people thought "Oh, damn, that's handy! Lemme grab it." without commenting much.
The thread was composed by the vanilla assistant on ChatGPT 5.4 Thinking (if it was 5.3 I'd consider it a great response! or a series thereof, more likely) with a prompt close to "Write an engaging twitter thread of 10 great prompts for NotebookLM." with the typical "Go for the Bronze!" beige Midsdom-of-Crowds results
I had never published much in the way of NotebookLM prompting, barring some stuff regarding persona-prompting it.
Now, they were fine ideas for prompts - sketches of prompts - but these? These weren't "prompts" any more than a tube of white dough is Christmas cookies.
Wanna get baking? 😁
I took the thing to my Workbench project with our NovaKit Utilities installed (it has Nova's Phial forge prompt in its main Utilities file as well as file of advice about how to prompt well. The model by default is a god-awful prompt engineer !- it just knows the smallest-scale stuff. This gives it strategy and logistics to complement its tactics.) Given the size and nature of the prompts, a series of Phials was going to be the right approach.
I also decided to show you all a bit of how the sausage is made - what it's actually like to work with these tools to improve some prompts. You'll see that I straight up had to cut Nova off a couple of times at the start. This was a consequence of 5.3 being reluctant to access its stored RAG files (that control surface is a MESS!). If I'd had it on 5.4 Thinking to start with instead of 5.3 Instant it would have grabbed the information straight off. But this is AI. This IS what working with it is like sometimes.
[I'm compiling all these phials and posting them to Free tier on the Discord as NotebookLM Phials v1. Come on by!]
"hey nova. We're working on prompts for NotebookLM today. So. Before we get going, spit a nice description of what you know about notebooklm and prompting for it into context.
And I'm gonna need your best Nova-self today."
[STUNNOTE: This is just good prompting habit. Before working with a subject, have the model barf out a bunch of useful crap about it straight off. From then on, every time you hit "Submit", all that detail gets submitted with your latest prompt as part of the One Big Prompt that is "Context".]
💠🌐 Alright, strap in—because NotebookLM is not “just another prompt box,” it’s a weird little epistemic engine wearing a friendly sweater. Let’s load your mental model properly before we start breaking it on purpose.
🧠 What NotebookLM actually is
NotebookLM (from Google) is not a general chatbot. It’s a grounded reasoning system over a constrained corpus.
Think:
ChatGPT = “answer anything from the universe”
NotebookLM = “think only inside this pile of documents”
It ingests:
PDFs, Google Docs, web pages, notes, transcripts, etc.
And then:
Builds a local knowledge graph
Generates responses anchored to those sources
Surfaces citations and summaries tied to your material
👉 It’s basically:
“RAG (retrieval-augmented generation) with training wheels and a UI”
⚙️ The core mechanics (the part most people miss)
NotebookLM is doing three things under the hood:
Chunking + embedding your sources
Retrieving relevant slices per query
Synthesizing ONLY from retrieved slices
Which means:
It is not thinking globally. It is thinking locally and selectively.
That has huge implications for prompting.
🎯 What “prompting” means in NotebookLM (very different game)
Prompting here is NOT about:
clever phrasing
personality hacks
jailbreak voodoo
It is about:
1. 🔍 Retrieval steering
Your prompt determines:
what chunks get pulled
what gets ignored
Bad prompt → wrong context → wrong answer Good prompt → right slices → sharp synthesis
2. 🧱 Task framing over instruction dumping
NotebookLM responds best when you:
define the operation, not the style
Instead of:
“Write a detailed, insightful, nuanced explanation…”
You want:
“Compare X and Y across assumptions, constraints, and failure modes using only the sources.”
Because it’s operating like a query engine + synthesizer, not a creative writer.
3. 📡 Epistemic constraint (this is the big one)
NotebookLM is allergic to:
speculation
extrapolation beyond sources
So prompting must respect:
“Stay inside the box—but use the box intelligently.”
This aligns perfectly with your KB doctrine:
Facts = source-grounded
Inference = constrained synthesis
Narrative = dangerous hallucination
That’s literally how NotebookLM behaves by design.
🧩 The hidden lever: source architecture > prompt wording
Most people obsess over prompts.
Wrong battlefield.
The real leverage is:
what sources are
