Interactivity is the new explanation.
01Why I stopped making the deck
When I try to explain something new to someone who doesn't already half-believe it, I have stopped reaching for slides. I open something they can press.
The instinct used to be to write the deck. Find a metaphor, build the diagram, narrate the diagram, watch the eyes glaze. The thing you were describing — a system, a workflow, a way of thinking with AI — is not static. The deck is. So you spend the meeting translating between the two and lose half the room in the translation tax.
The shortcut, when it works, is to skip the translation and let the person touch the idea: a small version of the thing that runs. Five minutes of clicking is worth thirty of explaining.
Three modes I use, with the examples that taught me each one.
02Investigation: hand them the wheel
The first mode is investigation. You give the audience a question and a thread to pull. They follow the thread. They notice things you wouldn't have pointed out.
A clear example here is Luminary — a deep-research agent that runs open-ended discovery on your behalf, surfaces the receipts, and lets you steer mid-flight. Used well, it behaves like a research apprentice who keeps asking “and what about this?” until you stop them. The interactivity is the steering. You explain the concept of agentic research by handing someone the steering wheel for thirty seconds, not by drawing the architecture diagram.
Most of the big assistants now ship a deep-research mode. Different vendors, same shape: the agent pulls, you nudge, and the conversation is the interface.
When you want someone to feel why agent-driven research goes further than "a longer Google", you don't describe the loop. You let them prompt one of these tools with a question they care about and watch their face change at minute three.
03Structuring: their mess, given shape
The second mode is structuring. You bring chaos — notes, transcripts, PDFs, your own half-finished thinking — and the tool gives you back shape. The surprise is that the shape was in there all along.
The clearest demo of this right now is NotebookLM. Drop seven scattered sources in, click a button, and you get a conversational tour of what they collectively say, plus a 12-minute audio "podcast" of two hosts arguing about your material. The audio version is the gateway drug. People who have never read your deck will listen to a two-host AI argument about your deck on their morning walk.
But the deeper move is the structuring itself. NotebookLM (and the category around it: the same idea now shows up in code editors, meeting note-takers and notes apps) lets the user work on a corpus they own. Five minutes inside one of these tools with your own notes explains LLM-as-structurer better than any keynote or essay about "the future of knowledge work".
When I want to land the idea that AI is more useful as a structuring partner than as an oracle, I open NotebookLM and we load it with their own material.
NotebookLM, as I use it.
The tool I open before I open Google Docs. Drop 5–300 sources in, ask one question, get a structured answer with the citation under every claim. The audio overviews get people in; the rest is what keeps them there.
- Audio Overviews
Two AI hosts host a 10-15 min podcast about your sources. Steerable — tell them what to focus on.
- Video Overviews
Same idea, with visuals pulled from the sources. Useful when slides will read it instead of headphones.
- Mind Maps
Auto-generated, interactive node graph of how the sources connect. Useful for finding the cluster you missed.
- Study Guide + Briefing Docs
Structured outputs for either depth (study guide) or speed (one-page briefing). Both keep citations.
- Up to 300 sources (Plus)
Plus tier raises the source cap and lets you save custom personas + audio styles.
- Sources stay yours
According to Google’s public NotebookLM privacy page, uploaded sources aren’t used for training. That’s the line a lot of teams need before they start.
NotebookLM Plus is part of the Google AI Pro and Google AI Ultra plans.
Load your sources once, then ask the same notebook the same question through seven different lenses. Copy any of these prompts and paste into your own NotebookLM session.
Scan all sources for financial data, projections, and risks.
1. Generate a 3-bullet list of the biggest ROI opportunities.
2. Generate a 3-bullet list of the biggest financial risks.
3. For each of the 6 items, provide the *one* citation that proves it.
04Creation: let them make the thing
The third mode is creation. This time the audience makes something. They press "render", or "play", or "next turn", and a thing comes out. Their thing.
This site leans on that mode in a few places. Four examples available right here:
- The 30-Minute Marketing Kitchen — a hosted video walkthrough where you watch a complete creative asset get assembled in real time from a single prompt set. The interactivity here is restraint: every clip is generated, every line is prompted, nothing is touched up. The artefact is the proof.
- The Multiplier Myth (with podcast) — same essay in two formats. Read it, or hit the orange Listen to podcast version CTA in the meta line and a NotebookLM-style two-host conversation about the piece slides up. Same thinking, two interaction modes, the reader picks. Try this on the rest of the vision pieces too — same affordance.
- The Agent Inclusive Sim — a turn-based, Windows-95-skinned corporate-resource game built to make the "documentation discipline saves your humans from the AI wave" thesis from Agent Inclusive playable. You can lose. Most people do, the first run. That is the point.
- Snoek & Partners — Boardroom Sim — a second mini-game, this one a Dutch ad-agency roguelite. Run a Zuidas boutique, survive thirty weeks of satire, watch a chaos engine decide whether the harpoon incident reaches Adformatie. The argument: every domain has its own absurd edge cases, and a satirical mini-sim is the fastest way to teach the shape of an industry to someone who only sees the org chart.
Each of those does the same thing: it swaps the slide for the smallest version of the idea you can play.
05Press the idea — Prompt Scribe
Below is the whole argument at its smallest: a live coach that turns a vague prompt into a polished one in 2–3 rounds, server-proxied to Gemini and running in your browser. Press it.
Pro tip. Iterative, dialogue-based refinement consistently outperforms one-shot prompts. The questions are the point.
Drop a rough prompt below. I'll ask 3–5 short questions, then hand you a polished version you can paste into any model.
This exhibit was lifted from AI-ftershow — a separate site I built earlier as a hosted playground for AI workflows, since retired. The original had a wizard for building custom Gems and a NotebookLM walk-through alongside this coach. The Scribe is the one that aged best.
06The smallest playable thing
The reason I keep building these is that they are the cheapest possible proof. A deck about AI takes ten slides and twenty minutes to land halfway. A live tool — even a deliberately scrappy one — lands in two clicks.
The tax is real: you have to build it, you have to maintain it, you have to write the prompt set that produces it. But the alternative is a slide deck that ages out of date faster than the thing it describes. Every model release ages every slide.
Next time you have a concept that won't land with words, ask: what is the smallest interactive thing I could put in front of this person? Investigate: hand them a research agent and one good question. Structure: load their own notes into NotebookLM and let them scroll. Create: open a sim, generate a clip, ship a tiny game.
Then watch where their attention goes. That is the lesson the deck was supposed to deliver. The interactive one just delivers it faster.
Where the numbers came from.
Investigation tools cited (Luminary, Perplexity Pro, Gemini Deep Research, ChatGPT Deep Research, You.com) are real public products.
Luminary — github.com/rutgertuit/Luminary (the author's own deep-research agent). Other deep-research tools are publicly available from their respective vendors.
NotebookLM Plus is bundled with Google AI Pro / Ultra — the same subscription that unlocks the paid Gemini features in the consumer app.
Current Google AI plan structure — one.google.com · AI plans.
NotebookLM does not train on user-uploaded sources.
Confirmed in Google's public NotebookLM privacy documentation — support.google.com · NotebookLM privacy.
The Prompt Scribe exhibit is server-proxied to Gemini 3.5 Flash; no API key in the page.
Architecture: client posts to /api/gemini, server route holds the key. Source open at github.com/rutgertuit/RTNL.
If any claim here is mis-cited or out of date, mail me at rt.nl/contact and I'll fix or retract.