Latest / AI Fire Daily / #462 Neil: NotebookLM Tutorial Creates Reusable Master Prompts
Transcript
- 0:00There is this strange modern trap that we all
- 0:03kind of fall into. Oh, the viral prompt trap.
- 0:06Yeah, exactly. You find this viral master prompt
- 0:08online. You copy it, paste it into an AI. And
- 0:12for like exactly five minutes, you feel like
- 0:15an absolute genius. Right, because it looks so
- 0:17complex. But then you read the output. I mean,
- 0:19it's grammatically perfect. It's highly structured.
- 0:22But you realize it sounds absolutely nothing
- 0:24like you. Yeah, it sounds like a machine doing
- 0:27a really cheap impression of a marketer. It's
- 0:29the ultimate trap, right? We're effectively outsourcing
- 0:32our thinking. We use a template built for someone
- 0:34else's brain, and then we wonder why the result
- 0:37feels so hollow. Well, welcome to today's Deep
- 0:40Dive. We are taking a completely different approach
- 0:43to AI today. A much better approach, honestly.
- 0:45I think so. The mission is simple, but pretty
- 0:48ambitious. We want to move past those generic
- 0:51plastic sounding outputs. So we're going to deconstruct
- 0:54how to build a personalized, highly specific
- 0:57master prompt. from the ground up. I am really
- 1:00excited to get into this. If you follow this
- 1:02process, it genuinely becomes a build once use
- 1:04forever system. It completely solves the biggest
- 1:07headache in modern content creation, which is
- 1:10just endlessly staring at a blinking cursor,
- 1:13trying to make the AI sound human. It really
- 1:16is a massive headache, and I'll be completely
- 1:19vulnerable here. I mean, I still wrestle with
- 1:21prompt drift myself. Oh, we all do. Even with
- 1:23custom instructions, I'll be working on a script,
- 1:25and suddenly the AI just kind of wanders off
- 1:27into this generic internet speak. Using words
- 1:30like delve or game changer. Right. So to fix
- 1:33this, we need to completely rethink think our
- 1:35tools. We all think we know what a master prompt
- 1:38is. Usually it's just a giant block of text we
- 1:40bought from some guru. But how would you actually
- 1:43define it, practically? I'll give you the absolute
- 1:45tightest definition. A master prompt is a reusable
- 1:49template that guides AI consistently. So it's
- 1:51not a magic spell. It's an architecture. Exactly.
- 1:55A strong master prompt has a really rigid six
- 1:58-part architecture. If you miss even one of these
- 2:00parts. The AI loses the plot. It just starts
- 2:03making things up. Yeah, it starts guessing. And
- 2:05when an AI guesses, it defaults to the mathematical
- 2:08average of the internet, which is boring. So
- 2:11let's build this architecture. What is the absolute
- 2:13foundation, the very first piece? Part one is
- 2:16the role. You have to define the specific specialist
- 2:19the AI is supposed to be. OK. Without a role,
- 2:22the model just defaults to a polite, hyper -generic
- 2:25assistant. But wait, I've tried giving it a persona
- 2:28before? I'll say, you know, act like a professional
- 2:31writer and the output still sounds like a caricature.
- 2:35It gets overly dramatic. Because professional
- 2:37writer isn't a specific role. I mean, it's way
- 2:40too broad. A weak role... creates a weak output.
- 2:44A strong role gives the AI a very specific worldview.
- 2:48So instead, you tell it, you are a YouTube script
- 2:50writer. You specialize in tutorial hooks for
- 2:53non -technical audiences. And these audiences
- 2:55are highly skeptical of big flashy promises.
- 2:58Oh, wow. That's fascinating. You're not just
- 3:00giving it a job title. You're giving it a philosophy.
- 3:03Right. You're giving it real boundaries. But
- 3:04even if it knows exactly who it is, it's still
- 3:06sort of shouting into a void if it doesn't know
- 3:08who's listening. Yeah. Which I assume brings
- 3:10us to part two. It does. Part two is context.
- 3:14This is where you describe the audience's mindset.
- 3:17And this is where most people fail, actually.
- 3:20Because they just list demographics. Exactly.
- 3:22They say, my audience is millennials who live
- 3:24in cities, which gives the AI almost nothing
- 3:28to work with emotionally. True context is about
- 3:31mindset. You need to list their distrusts. You
- 3:34need to list their actual desires. How does that
- 3:36look in practice, though? You'd say something
- 3:38like... You're writing for solo business owners
- 3:41age 30 to 50. They have tried AI tools before
- 3:45and were deeply disappointed. That's very specific.
- 3:48Yeah. And you add, they are extremely short on
- 3:51time. They want a specific actionable result,
- 3:54not a vague promise. That really grounds the
- 3:56whole interaction. No fluff allowed because the
- 3:59audience just won't tolerate it. So we have the
- 4:01role. We have the context. What's the third pillar?
- 4:03Part three is the objective. You have to state
- 4:05the exact format, the exact length, and the precise
- 4:08purpose. You never just say, like, help me write
- 4:11a video intro. No, never. Because it'll give
- 4:13you a three -page essay. Exactly. You say, write
- 4:16the opening hook for a 10 -minute YouTube tutorial.
- 4:19It must be spoken in under 30 seconds. You leave
- 4:22absolutely zero room for the AI to guess the
- 4:25container it's filling. Which leads perfectly
- 4:27into part four. And the source material highlights
- 4:30this as the most critical step. constraints.
- 4:34This tells the AI what to actively avoid. Yes.
- 4:37You have to block the bad patterns explicitly.
- 4:41Language models are basically prediction engines,
- 4:42right? They predict the next most likely word.
- 4:46If you don't build a fence, they will wander
- 4:47into the most highly probable and therefore overused
- 4:50vocabulary. I have a theory about this actually.
- 4:53Copying someone else's master prompt. without
- 4:56understanding these constraints is it's like
- 4:59wearing someone else's prescription glasses.
- 5:01Oh, that's a good way to put it. It's perfectly
- 5:02focused. It's just not focused for your eyes.
- 5:05You're adopting their boundaries, not your own.
- 5:07But let me ask you this. Why are constraints
- 5:10the part that most people completely skip, yet
- 5:12they're the most vital? Because without them,
- 5:14the model defaults to those hyped up words like
- 5:17game changer or revolutionary. Explicitly blocking
- 5:21bad patterns forces good ones to appear naturally.
- 5:24So telling the AI what not to do is the real
- 5:28secret. Absolutely. It clears out the weeds.
- 5:30When you forbid the AI from using generic buzzwords,
- 5:34it actually has to think. computationally, I
- 5:37mean, to find a more creative, precise way to
- 5:40express the idea. So we've blocked the bad habits.
- 5:43What is part five? Part five is examples. And
- 5:45this is your biggest lever for quality control.
- 5:48People think they need to feed an AI 50 examples
- 5:50for it to learn. Yeah, they just dump data into
- 5:53it. You really don't. Quality matters so much
- 5:55more than volume here. Two incredibly strong,
- 5:58perfectly calibrated examples are all you need.
- 6:00We'll dig into exactly how to find those examples
- 6:02in a minute. But let's finish the architecture.
- 6:04What's the sixth? and final piece. Part six is
- 6:07the output format. You tell the AI exactly how
- 6:10you want to read its response. If you don't do
- 6:12this, it will just dump a huge wall of text on
- 6:14you. Right. It gives you that standard bulleted
- 6:16list with emojis that screams, I use chat GPT.
- 6:19Exactly. So you control the format. You say,
- 6:22provide three distinct hook options. For each
- 6:25option, give me the full script. Below the script,
- 6:28identify the specific emotional trigger you used.
- 6:31And finally, state the specific promise made
- 6:33to the viewer. That is brilliant. It forces the
- 6:35AI to show its work. Like, it has to justify
- 6:39its own writing against the psychology of the
- 6:41audience. And it makes evaluating the output
- 6:43so much faster for you. Okay, so we have this
- 6:46beautiful pristine framework role context objective
- 6:50constraints examples output format But right
- 6:53now it's just an empty shell a framework needs
- 6:56data to actually run How do we fill this empty
- 6:59architecture with reality? This is where we introduce
- 7:01the tool stack and the best part is it's completely
- 7:04free The core engine we're gonna use is notebook
- 7:06LM We'll use Google Docs just to store our text
- 7:09and we'll use a platform like medium for psychological
- 7:12theory I want to stop right there Why notebook
- 7:14LM specifically? I think a lot of people listening
- 7:17are probably wondering, why not just build this
- 7:18in chat GPT or Claude? Why learn a whole new
- 7:21interface? Because of how notebook LM is engineered,
- 7:25it operates as a closed system. It uses a technique
- 7:28called source grounding. Meaning what? Exactly.
- 7:31When you ask it a question, it doesn't search
- 7:33its entire massive generic training data. It
- 7:37only reads the specific documents you upload
- 7:40to it. So it literally cannot wander off into
- 7:42the broader internet. Exactly. It completely
- 7:45eliminates that generic internet drift. It is
- 7:48legally bound, so to speak, to only use the reality
- 7:52that you provide. That is a really crucial distinction.
- 7:55OK, so we need to provide it with reality. How
- 7:57do we actually gather this data? Walk me through
- 8:00step one. Step one is manual data collection.
- 8:02You have to get your hands a little dirty here.
- 8:04You go to YouTube. You search for your specific
- 8:06niche. Let's say your niche is email productivity
- 8:09tips. OK, I search that. I get a million results.
- 8:11Right. So you filter those results by this year.
- 8:13So the data is fresh. Then you sort them by you
- 8:16count. But here is the critical part. You are
- 8:19not just looking for the biggest channels. What
- 8:21are we looking for? You are hunting for outliers.
- 8:24Define an outlier in this context. An outlier
- 8:27is a video that vastly outperformed its channel's
- 8:30baseline average. Imagine a channel with 2 ,000
- 8:33subscribers. Usually they get 500 views a video.
- 8:37But suddenly one video has 50 ,000 views. That
- 8:40means the algorithm didn't push it based on channel
- 8:42reputation. Yeah. It pushed it purely because
- 8:44the thumbnail and the hook were mathematically
- 8:47undeniable. Spot on. That specific hook broke
- 8:50through the noise. It is pure concentrated signal,
- 8:53and it's deeply worth studying. So you find 10
- 8:56to 15 of these true outliers. And what's the
- 8:59actual process of capturing them? It's very analog.
- 9:02You literally watch the first 30 seconds of each
- 9:04outlier video, you transcribe the spoken hook,
- 9:07and you paste that text into to a Google Doc
- 9:10right next to the video's title. So now I have
- 9:12a document filled with the raw proven text of
- 9:15what actually works in my niche. Yes. But here's
- 9:18the crucial twist that makes the system brilliant.
- 9:20You don't stop at the YouTube hooks. You open
- 9:23a new tab, go to Medium or a similar platform.
- 9:26You find two or three highly rated articles on
- 9:28human psychology or advanced copywriting principles.
- 9:31Like articles about the curiosity gap or how
- 9:33fear of missing out works in the brain. Exactly.
- 9:36You grab those URLs. So now you have your raw
- 9:39data. Step two is preparation. You open a new
- 9:42notebook in Notebook LM. You upload your Google
- 9:45Doc full of those 15 successful successful hooks,
- 9:48you paste in the URLs for the psychology articles,
- 9:50and finally you paste in the blank text of that
- 9:54six -part master prompt framework we just went
- 9:57over. I really love this approach. It anchors
- 9:59the language model entirely in empirical evidence.
- 10:02It's not sitting there guessing what might work
- 10:04based on old forum posts. It's being fed proven
- 10:07reality. Exactly. But I want to clarify something
- 10:10here. If we already have the successful YouTube
- 10:13hooks, which is the raw proof of what gets views,
- 10:15Why do we need to upload the psychology articles,
- 10:18too? Because the hooks show the AI what worked,
- 10:21but the psychology articles teach the AI why
- 10:23it works, giving it the theoretical layer. It
- 10:25gives the AI the underlying theory, not just
- 10:27the raw data. Exactly. It maps the fast -paced
- 10:31action of a YouTube hook to the deep, slow psychology
- 10:35of human behavior. It learns that a specific
- 10:38pause in a sentence isn't just a pause, it's
- 10:41a tension -building mechanism. Okay, so let's
- 10:47visualize this. We have our raw materials locked
- 10:50inside Notebook LM. We have 10 to 15 outlier
- 10:53hooks. We have deep psychology articles. We have
- 10:56this six -part framework. It's a massive wall
- 10:58of text. It is a lot of text. How do we synthesize
- 11:01all this without spending an entire week analyzing
- 11:03it ourselves? This is where the magic really
- 11:05happens. Step three is pattern analysis. We give
- 11:09Notebook LM a highly specific prompt. We command
- 11:12it to read all the uploaded sources and cross
- 11:14-reference them. What exactly are we instructing
- 11:16it to extract. We tell it to identify the five
- 11:18most common patterns across all those successful
- 11:21hooks. We want it to analyze the underlying sentence
- 11:23structure. We want it to identify the exact emotional
- 11:26triggers being used and link them to the psychology
- 11:28articles. Like recognizing when a creator is
- 11:31agitating a specific pain point before introducing
- 11:33the solution. Exactly. We also ask it to analyze
- 11:36the literal rhythm of the words. How long are
- 11:38the sentences? Where do they pause? We ask for
- 11:41a deeply analytical breakdown of the mechanics.
- 11:43That's incredible. It's essentially reverse engineering
- 11:46the charisma of a viral video. It is. And then...
- 11:49Based purely on those extracted patterns, we
- 11:52ask it to draft five strict constraints to help
- 11:55us avoid sounding hyped or generic. And finally,
- 11:58we ask it to pull the two absolute best examples
- 12:01from our raw data that perfectly embodied these
- 12:03successful patterns. Wow. Whoa. Just imagine
- 12:06scaling to a billion queries or just extracting
- 12:09the exact DNA of viral hooks in under a minute.
- 12:12Doing that by hand would take half a day. Right.
- 12:15I mean, you'd be staring at transcripts forever.
- 12:17You are distilling pure signal from the noise.
- 12:19here. Yeah. So notebook LM spits out this incredible
- 12:22pattern analysis. It gives us the patterns, the
- 12:24constraints, and the two perfect examples. What
- 12:26is step four? Step four is assembly, generating
- 12:30the actual master prompt. You immediately run
- 12:33a second prompt in notebook LM. You say, now
- 12:36use the patterns, the constraints, and the two
- 12:38examples you just identified. Fill out a complete
- 12:41master prompt using our six -part framework.
- 12:43So it acts as its own architect. It writes out
- 12:47the role. It defines the context based on what
- 12:49it learned about the audience from the hooks.
- 12:51Right. And it locks in the objective, lists the
- 12:54new constraints, inserts the examples, and sets
- 12:56up the output format. It writes the whole thing
- 12:59for you. That's amazing. It gives you a finalized,
- 13:02ready -to -use prompt that you can just copy
- 13:04and paste directly into your daily driver, whether
- 13:06that's Claude or ChatGPT. It's like stacking
- 13:08Lego blocks of data. You're letting the AI build
- 13:11a machine that will eventually build your content.
- 13:14And because it's built inside Notebook LM, the
- 13:16constraints are infinitely sharper because they
- 13:18come from hard data, not just our vague guesses
- 13:21about what sounds good. And the examples are
- 13:23bulletproof. They weren't just pulled from the
- 13:25ether. They were isolated as the mathematical
- 13:27best performers from your own curated research.
- 13:30I do have a question about that curation, though,
- 13:33about the examples. What happens if we get a
- 13:35bit lazy during the manual collection phase?
- 13:38Say we just feed it five average hooks instead
- 13:40of hunting down those two incredibly strong outliers.
- 13:43That will ruin the entire system. The AI learns
- 13:46from the patterns it's given. If you feed it
- 13:49mediocre examples, it will perfectly consistently
- 13:52replicate mediocrity. Quality is a far bigger
- 13:56liver than volume. Right. Garbage in means garbage
- 13:58out. So always choose absolute quality. Always.
- 14:02You want to feed it the gold standard because
- 14:03that becomes its new baseline. OK. So let's say
- 14:06we did the work. We have this beautifully crafted
- 14:08data -backed master prompt sitting in our clipboard.
- 14:12But I have to play the skeptic for a minute.
- 14:15Theory is great. It looks amazing on paper. But
- 14:17does it actually survive contact with the real
- 14:19world? That brings us to step five, the crucible,
- 14:23the test. You take your shiny new master prompt
- 14:25out of Notebook LM. You open your primary AI,
- 14:28let's say Chai GPT. You paste the master prompt
- 14:31in. And what do we ask it to generate? Do we
- 14:33ask it for an email productivity hook, since
- 14:35that was our research? No. That is the one crucial
- 14:38rule of the testing phase. You must test the
- 14:41prompt on a completely new topic, something that
- 14:43was absolutely not mentioned in your original
- 14:45source data. Wait, why? Wouldn't you want to
- 14:48test it on what it knows? Because if you test
- 14:51it on the same topic, you don't know if the AI
- 14:54is relying on the structural framework or if
- 14:56it's just regurgitating the specific facts it
- 14:59just read. You have to isolate the variable.
- 15:01OK, that makes perfect sense. Yeah. So give me
- 15:03an example of a good stress test. Let's stick
- 15:05with our scenario. All your source data was about
- 15:08email productivity tips. For the test, you give
- 15:11it the topic. Write a YouTube hook about using
- 15:14the Notion app to manage freelance client projects.
- 15:17A completely different software, different workload,
- 15:19a different target user. Exactly. You run the
- 15:22prompt, and then you aggressively audit the output.
- 15:25You look for three specific things. First, does
- 15:27the tone match those strict constraints we set?
- 15:31Second, does the structure actually follow the
- 15:33psychological patterns notebook LM identified?
- 15:36And third, and most importantly, does it sound
- 15:38like a real breathing human being? Rather than
- 15:41an AI trying to sound human. I mean, we all know
- 15:45that slightly too enthusiastic AI voice. We do.
- 15:49If the output sounds too formal, or if it slips
- 15:52in a dreaded word like powerful or crucial, you
- 15:56know your framework has a leak. I actually want
- 15:59to push back slightly here on how we fix those
- 16:01leaks. Measuring constraints is notoriously difficult.
- 16:04If the AI sounds too formal, a lot of people
- 16:07will just add a constraint that says don't be
- 16:09overly formal. But the source material has a
- 16:11great tip about this. Constraints must be strictly
- 16:14measurable. Walk me through what you mean by
- 16:17measurable. Well, you cannot give an AI a subjective
- 16:19command. Don't be wordy is a vague constraint.
- 16:22Vague constraints yield vague outputs because
- 16:25the AI doesn't know your personal definition
- 16:27of wordy. You have to give it a mathematical
- 16:28boundary. Ah, right. You have to say word cap
- 16:31of exactly 75 words or absolutely no sentences
- 16:35longer than 15 words. That is a brilliant clarification.
- 16:39The AI needs a hard boundary to bounce off of.
- 16:42So if it fails your test, You don't throw the
- 16:44prompt away. You go back to Notebook LM, you
- 16:47say, the output was too wordy. Update the constraint
- 16:50section to include a heart maximum of 12 words
- 16:53per sentence. Usually, one round of structural
- 16:56feedback completely fixes the leak. And once
- 16:59it finally passes the test, once you get that
- 17:02perfect, punchy, human -sounding hook. That is
- 17:05step six. You reuse it forever. You save that
- 17:09finalized text to a prompt library, keep it in
- 17:11Google Doc or a Notion dashboard, wherever you
- 17:13work, and you use this exact same workflow for
- 17:16everything you do. Gathering outliers, extracting
- 17:18patterns, building a framework. Exactly. You
- 17:21could build one for YouTube thumbnails, one for
- 17:23email subject lines, one for landing page headlines.
- 17:26One build infinite reuses. Next week, when you
- 17:29need a new hook for a video, you don't sit there
- 17:31staring at a blank screen. No, you don't start
- 17:33from scratch at all. You just grab your master
- 17:34prompt, paste it in, add your new topic, and
- 17:37you get high -quality, perfectly calibrated output
- 17:39in about five minutes. But how do we know for
- 17:42absolutely sure that this master prompt is genuinely
- 17:45a reusable system and not just a one -trick pony
- 17:48that got lucky once? By testing it on a topic
- 17:51that is wildly outside your usual range, if the
- 17:54tone and quality hold up without the AI hallucinating
- 17:57or breaking character, the framework is mathematically
- 18:00sound. If it works perfectly on a totally unrelated
- 18:03topic, your system is solid. It proves the architecture
- 18:07works independently of the subject matter. It
- 18:09proves you've captured the physics of good communication.
- 18:12Let's take a step back and look at the big picture
- 18:14here. The journey we just went on is deeply fascinating
- 18:17to me. We started with the frustration of copying
- 18:20other people's templates. And what we discovered
- 18:23is that the best prompt isn't found. It isn't
- 18:25a secret code you copy from a viral post. It
- 18:28is built. Built from the ground up, with intention.
- 18:31Yes. It is built from existing, proven evidence
- 18:34in your specific niche. It's constructed strictly
- 18:37around your specific audience's fears, their
- 18:39deep distrusts, and their actual desires. It
- 18:42fundamentally changes our relationship with the
- 18:44technology. It really does. It transforms AI
- 18:46from this generic, unpredictable slot machine
- 18:49where you pull the lever and just sort of pray
- 18:51for a good result into a reusable, highly calibrated
- 18:55engine. An engine that actually respects your
- 18:58time and your intelligence. It stops being a
- 18:59parlor trick and becomes a reliable piece of
- 19:01daily infrastructure. Which leaves me with a
- 19:03thought I cannot quite shake. It's a bit philosophical,
- 19:07but bear with me. If we can use this technology
- 19:10to so perfectly isolate, analyze, and automate
- 19:13the exact psychology of our most authentic human
- 19:17communication, what happens to our own voice
- 19:19in the long run? Does having a perfect digital
- 19:22clone, a framework that speaks exactly like us
- 19:25on our best day, force us to become even more
- 19:28deeply human, more unpredictable in our actual
- 19:31daily lives? It's something to think about as
- 19:33these tools get sharper. That is a profound question,
- 19:36and honestly, the best way to explore it is to
- 19:37get your hands dirty. While you ponder that,
- 19:39I highly recommend you take just 10 minutes today.
- 19:41Go to YouTube, search your niche, gather your
- 19:44first five outlier examples, run them through
- 19:46Notebook LM, test this out. See the mechanics
- 19:48for yourself. It is incredibly eye -opening once
- 19:51you see it work. Thank you for joining us on
- 19:54this deep dive. Take care, and we will catch
- 19:56you next time.