Most people moved to Claude Design for carousels this year. I didn't.
I'm still on NotebookLM. Same job: pitch decks, social carousels, infographics. Free, no tokens, no quality loss. And a workflow my paid-tool friends don't have.
This year I made 20+ decks in NotebookLM. About 20 minutes each. 2 languages. None of them look like a PowerPoint template. Some of them, like the legal-tech pitch attached at the bottom of this post, look like they came from a design studio.
Here's the workflow.
Why I'm still on NotebookLM
Three reasons. They apply whether you're making a 12-slide pitch deck or a 9-image carousel for LinkedIn.
- Free. Claude Design and similar tools eat tokens. NotebookLM doesn't charge per generation. For someone testing 5–10 visual ideas per deck, that adds up fast.
- Speed. A full deck renders in about 11 minutes. Sometimes 8. I've never waited longer than 15.
- Control. The visual brief is a paragraph you write yourself. Not a template, not a preset style, not "modern minimalist" with stock icons. You tell NotebookLM exactly what you want: color palette, typography, layout density, what to avoid.
The contrarian part: most AI presentation tools try to be everything. NotebookLM stays focused on your sources. That's the strength.
The 5-step workflow
This is the part that took me 20+ decks to figure out. The order matters. Skipping step 3 breaks everything.
- Load everything as sources. Text, brand book, even a 200-word post. NotebookLM reads all of it.
- Add visual references. Screenshots, magazine spreads, brand book pages. NotebookLM uses these too, not just text.
- Write the visual brief in ChatGPT. Same sources, same references. Ask: "write a structural prompt and a slide-by-slide breakdown of where the text goes." ChatGPT handles nuance better than NotebookLM does alone.
- Paste the brief into NotebookLM. Set language and format in slide-creation settings. Generate.
- Iterate one variable at a time. Don't regenerate from scratch. Change color, then layout, then headline style. Each tweak is faster and teaches you what actually matters.
About 5 minutes for the brief. 5 for references. 10 for NotebookLM to render. The deck is done.
Step 3 — the unlock
I called this "the unlock" before. After 20+ decks, I'd call it the architecture.
Without step 3: clean slides, generic feel. The visual style lives separately from the content. You upload a brand book, NotebookLM makes something on-brand but bland. The structure doesn't match the message.
With step 3: the visual brief and the content are shaped in the same conversation. ChatGPT sees your sources and references at the same time. The brief it writes knows what slides you actually need.
What I send ChatGPT:
- the raw source material I'm presenting
- the same visual references I gave NotebookLM
- an instruction like "write a structural prompt and a slide-by-slide breakdown of where the text goes"
What comes back: a 1,200–1,500 token visual brief that NotebookLM doesn't generate on its own.
The actual prompt I use
This is what I wrote for the legal-tech deck attached below. It's longer than most people would write. That's the point.
Create a premium editorial presentation based on the article. The presentation should feel like a modern design magazine, luxury publication, or creative studio case study, not a traditional PowerPoint. Use a minimalist Swiss-inspired layout with generous white space, strong typography, asymmetrical composition, subtle paper textures, geometric construction lines, and collage-style cut-out elements. The overall mood should be sophisticated, architectural, and design-first. The color palette should consist of deep burgundy (#650F14), warm light gray (#ECECEC), black, white, with burgundy used only as an accent. Use bold sans-serif typography for impact words and elegant serif typography for key statements. Make oversized numbers and short headlines the visual focus. Visually incorporate legal-themed objects such as scales of justice, gavels, law books, document fragments, newspaper textures, classical statues, technical diagrams, and editorial labels. Avoid stock icons, gradients, corporate templates, and default presentation layouts. Structure the story visually rather than with text-heavy slides.
Feel free to save it and adapt it for yourself.
Let me show you on a real example
Last month I needed a 9-page pitch deck for Zakonnik BY, the AI legal bot I'm building for Belarusian legislation. The bot runs on StepFun step-3.7-flash with ChromaDB retrieval. 18 national laws indexed, 3,794 chunks. A Validator Agent that catches hallucinations before they reach the user. 177 tests, all passing.
Brand palette: deep burgundy, warm gray, editorial typography. No stock icons. No gradients. No corporate template.
Here's what I actually loaded into NotebookLM:
- my text notes (raw thoughts, half-baked ideas, written for myself)
- the LinkedIn post about Zakonnik BY I had already published
- 4 magazine-spread screenshots as visual references (Swiss layout, asymmetric composition, oversized numerals)
- a text file with the main ideas I wanted the deck to communicate
Then in the slide-creation settings, I pasted the visual brief ChatGPT wrote for me.
Output: 9 slides. On-brand, no template vibe, no text hallucinations in the legal copy. Two iterations to tighten slide 4, which was too dense for a pitch context.
Total time: 25 minutes including iteration.
The deck is attached at the bottom of this post. And if you want the full story behind Zakonnik BY — why I built it, the hallucination problem, how the Validator Agent works, what "177 tests, all passing" actually means in production — I wrote about it in this post:
Building a Belarusian Legal AI Bot with Explicit Bans and Validator Agent | Tory Kovdya posted on the topic | LinkedIn
Last month, I decided to create a legal AI bot for Belarusian legislation. I couldn't afford a lawyer. But I needed answers. That's why I built Zakonnik – a Telegram bot. StepFun step-3.7-flash, ChromaDB, 3,794 chunks from 18 national laws. 9 categories: labor, family, housing, consumer, civil, criminal, medical, social, administrative. Just a bit boring info, I know 😅 The first tests were a real challenge. The bot kept mixing two legal systems in one answer. I'm convinced the model was trained primarily on Russian legal texts and treats RB legislation as the same thing. So I rebuilt the system prompt. Added explicit bans. Made pravo.by the only authoritative source. The model still slipped through. That's when I added a validator agent. Before the answer reaches the user, a second check runs – it looks for hallucinations, wrong citations, references to Russian law. If it flags something → the bot either corrects itself or escalates to a human. 177 AI agent tests. All passing. The bot cites specific articles from the source documents, not just a general answer. And it shows a confidence score, so the user knows how much to trust it. I'm still testing. The bot answers questions in Telegram. It cites specific articles. It gives confidence scores. And now it adds a short disclaimer: "⚠️ This answer was generated by Zakonnik BY AI. For informational purposes only. We recommend contacting a licensed lawyer in the Republic of Belarus." Has anyone else built AI for a legal system that's not the dominant one on the internet? I'm still figuring out how to make it reliable enough to actually trust it. #LegalAI #BuildInPublic #CreateImpact
linkedin.com
Iteration is part of the process
First generation is rarely the best one.
Helpful rule: don't regenerate from scratch. Tweak the prompt with one variable at a time. Color. Layout density. Headline style. Each small change is faster and tells you more about what matters in the design.
Who this workflow is for
Not everyone.
If you need a 50-slide corporate deck with quarterly financials, stick to PowerPoint.
If you need a 9–12 slide editorial-style pitch or a LinkedIn carousel that doesn't look like everyone else's, this works.
If you're testing visual ideas before committing, NotebookLM is faster than anything else I've used.
I stopped designing slides. Started writing visual briefs. The pitch deck got 10x better.
That sentence is the whole workflow in 14 words.
Want me to share how I used the same process on educational decks in 2 languages? Same workflow, different problem.
Top comments (0)