Training the AI Dragon

As I progress through this series, I am concluding that learning to prompt is no different from how Hiccup in the movie “How to Train Your Dragon” learned to train his dragon,Toothless.

I am in the process of putting together a NotebookLM for Educators course, which I will post on YouTube later this month. NotebookLM is ideal for educators and students, but what I discovered right off the bat was that to make that application “sing,” you need to know how to write a prompt—NotebookLM calls them “Steering Prompts”—for everything from how the chat responds to you to the design of a slideshow or how the Audio Overview can be customized. For me? A tall order.

Like many newcomers to this madness, I know what I want, but putting it into words is another story. I have lived in a visual world since the start of Desktop Publishing. Over these 40 years, I’ve learned to sketch my ideas or quickly whip up a comp to show what I’m thinking. You learn this the hard way when you tell the client your vision, and then, when you show it, there’s an awkward silence, followed by, “That isn’t what I was expecting.” If you’ve been around a while, you know exactly what I mean. It’s something I’ve been pounding into my students’ heads and at seminars for years: “You can’t describe it. You need to show it.”

This is what I find so frustrating about AI. I am now both the designer and the client. I tell the AI what I envision; the AI goes to work. I look at the result and say, “That isn’t what I expected.” The reason is that AI is a predictive technology. Any result you get is more along the lines of “Is this what you expected?” not “Ta Dah!”

Hardly a day goes by that someone on LinkedIn posts, “These prompts will make you rich!” or  “ Master AI with these 5 prompts.” Search “How to write a prompt” on YouTube, and you are inundated with such videos as  “Master the Perfect ChatGPT Prompt Formula (In just 8 minutes)” or “You Suck at Prompting AI (Here’s the secret.)”. Design tools are just as bad. Figma promises to “Make your ideas real with Figma Make”.  None of these will hand you a “Ta Dah”.

So what does this have to do with training a dragon? Everything.

In the movie, the young Viking Hiccup lives in a world where encountering a dragon requires brute force—swinging an axe and hoping for the best. One day, our hero actually captures a dragon he names Toothless after damaging the dragon’s tail fin. This prompts him to take the time to understand how the dragon thinks and responds. He discovers that offering a fish is better than swinging an axe, that building a new tail fin for the dragon leads to a partnership where the dragon flies and he steers, rather than asserting dominance. His biggest lesson is that you don’t need brute force to train your dragon; you need to communicate.

There is no magic formula for writing prompts. Learning how is very much like training a dragon. I started by barking commands at the AI and becoming increasingly frustrated with the results. My “Ah Ha!” moment arrived when I did what Hiccup did: stop trying to apply brute force to the AI and understand how it works. For example, feeding Nano Bannana a reference image or ChatGPT or Claude a written scenario is no different from Hiccup offering his dragon a fish. Give the AI structure and constraints, and you have built a new tail fin to steer the AI and build a collaborative relationship. Along the way, you will realize that you get extraordinary results not by commanding, but by communicating clearly, providing context, and working with the AI rather than against it.

Here’s one example of how I trained my NotebookLM Dragon.

When you ask NotebookLM a question in the chat pod, it reviews its sources and provides you with a lengthy, detailed response that often tends to ramble. Instead of just chatting, I wanted to create a Custom Learning Guide using a Steering Prompt for the chat. I began with one I got from Robert Pham, who has a series of NotebookLM videos on YouTube.

Break this subject into:        

- Core concepts        

- Supporting concepts        

- Prerequisites        

- Common misconceptions        

Present it as a structured learning roadmap.        

The response was more focused, but I decided to feed my dragon a fish and build it a new tail fin to create a document that could be circulated to students.

The prompt starts with feeding NotebookLM a fish:

Using only the provided sources, propose a logical learning or research order and identify the reading order for the sources.        

This is followed by the tail fin, laying out  the structure and constraints for the Learning Guide:

Using the sources, identify which sources provide the same information, identify the 3 most important sources, and explain why.        

I added a bit more to make the tail fin a bit sturdier: Identify contradictions, disagreements or opposite conclusions between the sources and explain what each side presents and the source of the disagreement.

I gave it a quick test and, as I always do, I asked, “What’s missing?” What was missing was that the student was dropped right into the middle of the response. They needed to be eased into the document. I added to my dragon’s tail fin by entering this right at the top of the custom prompt: In this Notebook:

Differentiate between fact and interpretation        

Label speculation and assumption        

Give evidence-backed claims priority        

When answering a question, first explain it in plain language that a beginner would understand. Then explain again for an advanced audience.        

Which makes me wonder. Maybe prompting is more like feeding a dragon than feeding a prompt into an LLM.