Latest / AI Fire Daily / #461 Neil: AI Presentation Hacks That Force NotebookLM To Generate Pro Decks
Transcript
- 0:00Imagine the promise. You just click one single
- 0:02button and a completely formatted professional
- 0:06slide deck just materializes in two minutes.
- 0:10Oh, man. It sounds like the absolute ultimate
- 0:12productivity dream. It really does. But then,
- 0:14well, then you look closely at the result. Yeah.
- 0:16And it's just this plain, generic, totally uneditable
- 0:20mess. Oh, totally. It's the classic trap of early
- 0:22automation. We've basically been conditioned
- 0:25by the tech industry to expect instant magic.
- 0:28Right. and we completely bypass the actual architectural
- 0:32work required to make that magic, you know. actually
- 0:36useful. Welcome to this deep dive. We're really
- 0:38glad you're joining us today. I've been thinking
- 0:41a lot lately about the friction between our creative
- 0:44intentions and the actual output we get from
- 0:46AI tools. It's a huge pain point for so many
- 0:48people right now. It is. So our mission today
- 0:51is to kind of step outside of that one -click
- 0:53illusion. We are going to break down a professional
- 0:57three -step presentation workflow inside Notebook
- 1:00LM. Which is such a game changer once you understand
- 1:02it. Exactly. So first, we'll look at how to build
- 1:06a walled garden of trusted sources to completely
- 1:09eliminate data hallucinations. Essential. Then
- 1:12we'll explore how to craft slide by slide prompts
- 1:16for really rigorous brand control. And finally,
- 1:20we're going to solve one of the most maddening
- 1:22bottlenecks in this entire process. Oh, I know
- 1:24what you're going to say. Yeah. The flat image
- 1:27export limitation. Yes. It is a critical shift
- 1:31in mindset. You really have to move from being
- 1:33just a passive consumer of a software feature
- 1:36to acting as an active art director. I like that,
- 1:39an active art director. Exactly. Because the
- 1:41tool is wildly powerful, but only if you constrain
- 1:44it first. It's all about setup. generation, and
- 1:48targeted revision. You know, whenever I sit down
- 1:50to build a presentation, the immediate temptation
- 1:52is to just jump straight to the design phase.
- 1:54Oh, we all do it. Right. We want to pick the
- 1:56fonts, we want to arrange the colors, and just
- 1:57make it look pretty. Yeah. But the foundation
- 2:00of any meaningful deck isn't the aesthetic layer.
- 2:02No, not at all. It's the data layer. If we skip
- 2:05straight to the generate button, we are basically
- 2:07just guaranteeing mediocrity. Because the prettiest
- 2:10typography in the world cannot save factually
- 2:13bankrupt content. Right. NoBookLM is fundamentally
- 2:16a source -based tool, and that makes its underlying
- 2:19architecture very different from a generalized
- 2:22model, like chat GPT or Gemini. Let's define
- 2:25that clearly for a second. A source -based AI
- 2:28strictly analyzes the specific files you upload.
- 2:31It doesn't scour the open internet unless explicitly
- 2:34told to. It's kind of like giving a student an
- 2:36open book test, but you only allow them to bring
- 2:39three very specific textbooks into the exam room.
- 2:42That is a perfect analogy. That constraint is
- 2:45exactly what gives it power. But, you know, human
- 2:48nature kicks in, right? Always. We're busy. We're
- 2:50on a deadline. We think, oh, I'll just use the
- 2:52auto web search feature and let the AI pull the
- 2:54background context for me. I mean, I have to
- 2:56be honest here. I still wrestle with prompt drisht
- 2:58myself when I don't give the AI rigid boundaries.
- 3:01Really? Oh, yeah. I get lazy. I'll just drop
- 3:03in one vague link, skip the deep source curation
- 3:06entirely, and just hope the model magically intuits
- 3:09my entire business strategy. We all do it. It's
- 3:12so tempting. But relying on live web searches
- 3:15for your core data is essentially just inviting
- 3:17chaos into your workflow. It's a mess. It's like
- 3:20a dog catching a random scent on a walk. the
- 3:22AI just wanders off. You're pulling in SEO spam,
- 3:25you're grabbing irrelevant sidebars, completely
- 3:28out of context quotes. It completely dilutes
- 3:30the core message of your presentation because
- 3:32the AI is trying to synthesize too many conflicting
- 3:35realities. So the solution is heavily curating
- 3:38that left -hand source panel. We can upload up
- 3:40to five trusted formats, local files, live links.
- 3:44PDFs, Google Drive documents, or even YouTube
- 3:47videos. Let's ground this in a real scenario.
- 3:50Say you are tasked with building a really complex
- 3:53deck on climate change mitigation. Oh, wow. OK,
- 3:56a massive topic. If you just hit generate. with
- 3:58an open web search on climate change, the AI
- 4:00is going to drown in contradictory noise and,
- 4:02you know, political opinion. Exactly. But if
- 4:04you purposefully combine a verified NASA PDF
- 4:07on atmospheric data, a really well curated Wikipedia
- 4:11summary and a specific YouTube explainer video
- 4:14breaking down carbon capture, you are essentially
- 4:16curating the entire universe the AI is allowed
- 4:19to perceive. Waldgarden. Yes. you're building
- 4:22a verifiable reality. And the fascinating part
- 4:25is just watching how it synthesizes those specific
- 4:27inputs. It cross -references the dense academic
- 4:32data tables in that NASA PDF with the highly
- 4:35accessible conversational transcript from the
- 4:38YouTube video. It creates this factual foundation
- 4:41that is completely bulletproof, but still really
- 4:44easy for a general audience to digest. And it
- 4:47gives you a quick summary in the center panel,
- 4:48which sort of acts as a confirmation that it
- 4:50has actually ingested the reality you've built.
- 4:53Exactly. So stepping back for a second, why exactly
- 4:56does pulling from live web searches dilute the
- 4:59quality of the slides so much? Basically because
- 5:01it introduces completely uncontrolled variables
- 5:04and just totally irrelevant context into the
- 5:06AI's synthesis engine. OK, so bad sources equal
- 5:10mixed results. Verified uploads equal laser -focused
- 5:13accuracy. That is the anchor of the whole workflow.
- 5:16You have to wall off your reality first. So once
- 5:19we've successfully walled off the AI's universe
- 5:22with our trusted PDFs and videos, The next logical
- 5:25hurdle is visual. Yep. How do we stop it from
- 5:28turning that great data into a boring gray wall
- 5:31of text? We have to move our attention from the
- 5:34source panel over to the prompt box. And this
- 5:37transition is where a lot of users just totally
- 5:40drop the ball. They do. They assume the AI inherently
- 5:43knows how to pace a narrative arc. But before
- 5:46you even hover over that generate button, you
- 5:49have to interact with the visual description
- 5:50box. You absolutely must define your structural
- 5:54goal. Let's unpack the formats Notebook LM offers
- 5:56first. I notice in the advanced options, there's
- 5:59a distinct split between presenter slides and
- 6:02a detailed deck. Yes, and it's a vital distinction.
- 6:05Presenter slides are lightweight visual layouts
- 6:07meant strictly for live public speaking. Right,
- 6:10so if I'm speaking on a stage, I want to force
- 6:12the AI into presenter mode. Exactly. I don't
- 6:14want it generating a massive paragraph that basically
- 6:17acts like a teleprompter on the screen behind
- 6:18me and a detailed deck. Those are heavy text
- 6:21sides designed for people to read alone. So think
- 6:24of an internal company memo or a really dense
- 6:27research report. That's just formatted as a presentation
- 6:29Okay, so you set those guardrails in the advanced
- 6:32options menu by clicking the small arrow next
- 6:35to the slide deck button You choose the format
- 6:38the language and a rough slide length Under 10
- 6:41slides is short, 10 to 15 is the default, and
- 6:44anything over 15 is long. But the real leverage,
- 6:47the actual design control, comes from the description
- 6:50box itself. And so many people leave it entirely
- 6:53blank. I know. Or they write something incredibly
- 6:56weak. A classic weak prompt is just typing something
- 6:59like, create a professional presentation about
- 7:02digital marketing. Which sounds like a recipe
- 7:04for a completely generic, soul -crushing corporate
- 7:07snooze fest. So it guarantees it. The AI will
- 7:10default to the safest, most statistically average
- 7:13interpretation of professional. You get slate
- 7:15gray backgrounds, standard bullet points, and
- 7:18incredibly dry pacing. So how do we force it
- 7:21to act like a high -end design agency? We have
- 7:24to be brutally specific. Contrast that weak prompt
- 7:27with a strong one. You don't just ask for a deck.
- 7:31You specify a 10 -slide deck on 2025 marketing
- 7:35trends. You demand a dark navy background, white
- 7:39text, vibrant orange accents, and bold sans serif
- 7:42fonts. You're stepping into the role of an art
- 7:45director. you are assigning a highly specific
- 7:48creative brief. And you should dictate the structural
- 7:51rhythm, too. How so? Well, instructed to use
- 7:53one main headline, exactly two to three short
- 7:56bullets, and one simple visual per slide. And
- 7:59you define the tone. You tell it, make this practical,
- 8:02confident, and written for non -technical business
- 8:04owners. Exactly. But if we connect this to the
- 8:07absolute ceiling of what the tool can do, the
- 8:09ultimate control method isn't just writing a
- 8:11global style prompt. Right. It's dictating the
- 8:14entire presentation. slide by slide. This is
- 8:16where it gets incredibly powerful. It obviously
- 8:18requires a lot more upfront labor, but the resulting
- 8:20polish is just staggering. It really is. You
- 8:23essentially write a script for the AI, pacing
- 8:26out the entire narrative, and paste that detailed
- 8:29brief directly into the visual description box.
- 8:31Let's walk through what that looks like in practice.
- 8:33Okay. Imagine we're building a deck on digital
- 8:35tools for remote teams. Yeah. You start by establishing
- 8:38that global style we talked about. clean, minimal,
- 8:41light gray background, flat icon illustrations,
- 8:44friendly professional tone. Then you literally
- 8:46script the sequence. You type, slide one, title,
- 8:50the remote team toolkit, visual, a simple minimalist
- 8:54desk illustration. Perfect. Then you move to
- 8:56the problem statement, slide two, title, why
- 8:59remote teams struggle, visual, an illustration
- 9:03of confused expressions on a video call grid.
- 9:05And then you get highly specific on the solution,
- 9:07slide three, loom. You explicitly tell the AI
- 9:11the key takeaway you want on screen, which is
- 9:13record a three minute video, and you demand a
- 9:16specific icon like a red record button. Whoa.
- 9:20Just imagine the scale of this for a second.
- 9:22Imagine generating 10 completely distinct, perfectly
- 9:24tailored presentation decks for 10 different
- 9:26audiences. Ranging from highly technical engineers
- 9:28to absolute beginners. Exactly. All pulling from
- 9:31the exact same source material just by swapping
- 9:34out that master prompt in a matter of minutes.
- 9:36Just tweaking the prompt, hitting Generate, and
- 9:38letting it run. That completely redefines how
- 9:41we think about presentation prep. Oh, entirely.
- 9:43Now, the generation process does take two to
- 9:45four minutes, and you have to keep the browser
- 9:47tab open while it works. Yeah, that's true. But
- 9:50because you've mapped the sequence, the story
- 9:52stays tightly connected. The visual language
- 9:55doesn't just wander off halfway through. Right.
- 9:57But wait. Doesn't writing a massive, highly detailed,
- 10:01slide -by -slide brief sort of defeat the purpose
- 10:05of using a fast AI tool to begin with? Not at
- 10:08all. Because it forces the AI into your specific
- 10:11story flow, saving hours of back -end revision.
- 10:14Right. Putting in the time up front saves hours
- 10:17of frustrating formatting later. Exactly. Front
- 10:20-load the creative thinking so you can actually
- 10:22enjoy the final product. Okay, so you followed
- 10:24the blueprint? You've walled off your sources,
- 10:27you've written a meticulous slide -by -slide
- 10:29prompt, and the AI has generated a gorgeous deck.
- 10:33You download the PPTX file, open it up because
- 10:36you spotted a tiny typo on slide 4. You double
- 10:40-click the text. And nothing happens. Nothing
- 10:43happens. The sheer panic of that moment. Oh,
- 10:45yeah. You've got a stakeholder meeting in 20
- 10:47minutes. You're furiously clicking the title
- 10:50to change your revenue number. And you realize
- 10:52you're just dragging the entire slide around
- 10:55the screen like a photograph. It's the most critical
- 10:57limitation in Notebook LM right now. And you
- 10:59absolutely must understand it before you build
- 11:01a workflow around it. Notebook LM exports. slides
- 11:05as flat images. The system relies on a backend
- 11:09rendering process called NanoBanana. Let's define
- 11:12that quickly for everyone. It's the backend software.
- 11:15rendering your slides into flat, single -layer
- 11:18images. Which means it makes text natively uneditable.
- 11:22It does not create editable text layers or vector
- 11:24shapes or individual elements the way a standard
- 11:27PowerPoint or Apple keynote file does natively.
- 11:30It's kind of like someone handing you a beautiful,
- 11:32perfectly framed painting of a spreadsheet instead
- 11:34of the actual Excel file. Wow. That is a painful
- 11:39but incredibly accurate analogy. You can admire
- 11:42the painting. of the Q3 revenue numbers on the
- 11:46wall. But you cannot click into the cells and
- 11:48update them. So we are stuck with a painting.
- 11:51How do we fix a typo on a painting? We have three
- 11:54distinct paths for a revision. Let's start with
- 11:56option A, which is revising entirely inside Notebook
- 11:59LM. This is your best strategic move if you need
- 12:02to make large content updates that affect multiple
- 12:05slides. You just hit the revise button inside
- 12:07the tool to open the slide editing view. But
- 12:09again, the AI is not a mind reader. If you use
- 12:11a vague prompt like, make this slide better,
- 12:15you're going to get a mess. Because better is
- 12:17a highly subjective human concept. Right. The
- 12:19AI has absolutely no idea what better means for
- 12:21your specific audience. You have to give it surgical
- 12:23structural instructions. Exactly. You define
- 12:26an update goal, make the slide feel more modern.
- 12:29than a specific title update. Replace the current
- 12:31title with, how AI automation saves teams 10
- 12:34hours every week. And you must constrain the
- 12:37content changes. Reduce the long paragraphs into
- 12:40three short bullet points. Crucially, you have
- 12:43to remind it of your original design instructions
- 12:45so it doesn't lose the plot and generate something
- 12:48completely mismatched. Right, like keep the dark
- 12:50background and blue accent colors. Yep. It will
- 12:54re -render the slide as a brand new image. Now,
- 12:57the backgrounds might shift slightly because
- 12:59it's generating a fresh picture, but it creates
- 13:02a new version without overwriting your original
- 13:04deck, which is a really nice safety net. That's
- 13:06good. Yeah. But what if my content is actually
- 13:09perfect and I just need to change the structural
- 13:11order of the presentation? That's where we move
- 13:13to option B. Exporting the PPTX file. OK, so
- 13:16this is ideal for basic layout adjustments. You
- 13:19download the file and you open it in Google Slides
- 13:21or standard PowerPoint by going to File, Import
- 13:24Slides, and Upload. Right. The slides themselves
- 13:27are still completely flat images, but now they
- 13:29exist in traditional presentation timeline. You
- 13:32can drag slide four to the front. You can delete
- 13:35slide seven entirely. And if you're truly desperate
- 13:38and in a rush, you can always just draw a new
- 13:41editable text box. right over the top of the
- 13:44flat image to cover something up. Oh yeah, it's
- 13:46a bit of a hack, but it definitely works for
- 13:48quick fixes. But what if I don't want to hack
- 13:51it with layered text boxes? What if I just need
- 13:53to fix one single word on one slide and I need
- 13:57it to look perfectly native? Then you need option
- 13:59C. Gemini Visualize. Walk me through how this
- 14:03actually works. It is absolutely brilliant for
- 14:06lightning fast single slide edits. You don't
- 14:08even need to leave Google Slides. You just use
- 14:10the Visualize side panel. right there in the
- 14:12interface. So I click on the flat image slide
- 14:14I want to change. I open that visualize panel
- 14:17and I just describe the correction. You literally
- 14:20just type change the title of this slide to three
- 14:23habits of highly productive workers. Gemini analyzes
- 14:27the original image, matches the font style, matches
- 14:29the background aesthetics, and generates a new
- 14:32corrected image version of the slide right there.
- 14:34Wow. It seamlessly rewrites the title and swaps
- 14:38the image, saving me the massive headache of
- 14:40going all the way back into Notebook LM, writing
- 14:42a new prompt, waiting for the render, and re
- 14:45-exporting. Exactly. It's a surgical strike.
- 14:48It fixes the pain point without abandoning the
- 14:50visual quality. I keep coming back to this design
- 14:53choice though. Why would the developers engineer
- 14:55it to export as uneditable flat images in the
- 14:59first place? It feels so counterintuitive to
- 15:01how we've worked for the last 30 years. Basically,
- 15:04rendering them as images ensures a much cleaner,
- 15:07highly polished visual output. without formatting
- 15:10breaking across different devices. So it trades
- 15:12easy text editing for a much more polished, consistent
- 15:16visual design. It's a highly calculated trade
- 15:18-off. Think about it. If it exported a bunch
- 15:20of portally aligned, messy text boxes that broke
- 15:23when you opened them on a different operating
- 15:25system, users would immediately complain that
- 15:27the AI was terrible at design. Fair point. Okay,
- 15:30let's step back and look at the whole picture
- 15:32here. What does this all mean for our workflow?
- 15:34Let's recap the big idea from today's deep dive.
- 15:37The overarching philosophy is that you must treat
- 15:40Notebook LM as a rapid visual framing tool. You
- 15:44cannot treat it like a traditional granular slide
- 15:46editor. Right. The heavy lifting, the actual
- 15:49secret to a perfect deck, does not happen in
- 15:51the editing phase. It is heavily weighted in
- 15:54the upfront prompting and relentlessly curating
- 15:57exact sources. If you put the intellectual work
- 16:00in before you click generate, the output feels
- 16:03incredibly precise. So remember the best practices.
- 16:06Wall off your reality. Always use verified sources
- 16:09when accuracy matters. Bring your own material
- 16:12and do not rely on random web searches. Make
- 16:14sure you're matching your format to the room,
- 16:16too. Use detailed decks if the audience is reading
- 16:19it alone. and present your slides if you're speaking
- 16:21on a stage. Keep your slide instructions in plain,
- 16:23direct language. No flowery descriptions needed.
- 16:27Just clarity. Break complex revisions into single
- 16:29steps so the AI doesn't get overwhelmed and drop
- 16:32details. And don't be afraid to create multiple
- 16:34presentation versions within a single notebook.
- 16:37You can test completely different styles without
- 16:39starting from scratch every single time. Thank
- 16:41you for joining us on this deep dive. You now
- 16:43have the exact blueprint to take scattered notes,
- 16:46build a walled garden of truth, and turn them
- 16:48into professional polished presentations without
- 16:51falling into the one -click trap. It really is
- 16:53a paradigm shift once you adapt to the boundaries
- 16:55of the tool, sources, structure, and targeted
- 16:58revision. Before we go, I want to leave you with
- 17:01a final thought to mull over, something that
- 17:03builds on everything we just discussed. I love
- 17:05these. Let's hear it. We've spent this whole
- 17:07time figuring out how to precisely manipulate
- 17:10an AI to output static flat images that mimic
- 17:14legacy software like PowerPoint. But if AI can
- 17:18already perfectly synthesize, brand, and visualize
- 17:21our dense data in mere minutes, how long until
- 17:25the traditional slide deck becomes entirely obsolete,
- 17:28replaced by dynamic AI -generated visual dashboards
- 17:31that we interact with in real time?