This is MVP Unplugged. In this episode,

we're going to learn how to write novels

faster and more easily. We're joined by

Michael. Michael, welcome.

>> Thank you. Thank you for having me.

>> Yeah, so you have built an amazing app

called AI Story Builders that uses AI to

write novels and kind of hold together

the elements of story and narrative

verse and length and everything like

that, all the references. I'm really

excited for you to show your app for us

today. What inspired you to kind of

build this?

>> The thing that inspired me to build it

was when ChatGPT first came out. I guess

it's been 3 years now.

>> Uh-huh.

>> I wanted to come up with an application

to use that technology

that was something that I really cared

about and needed.

>> Mhm.

>> And I

really wanted to get back into writing

fictional stories. But whenever I sat

down to work on a story

I would have all these notes and, you

know, the here and papers here and I'm

using, you know, software like

Scrivener and um

I realized I'm building a database. So

initially I wanted to make a software

program that would just have a database

of my characters and that sort of thing

and then to use AI to kind of you know,

help, you know, write the story, you

know, to assist and that sort of thing.

So that's what inspired me to start

working on it. And then that led me down

a whole path.

And

you know, we can kind of get into that,

but but that's what really started me on

this

this path.

>> Mhm. Yeah. And you're a Microsoft MVP as

well. What inspired you become a

Microsoft MVP as well?

>> Um, why

>> became a Microsoft MVP because I was

involved um, in an open source uh,

uh, project called DotNetNuke.

>> Mhm.

>> And it was through working on that

um, and really

that's what made me fall in love with

open source and that's why AI Story

Builders it's fully open source, okay?

So, you have the full source code,

there's no, you know, it's MIT license,

so it it's the most permissible license

ever. And I saw that the power of open

source is when you give something back

to the community what you get back is

you know, it's measurable because the

feedback you get um,

just the support of people um, because

they realize they're all into this

together, we're trying to make something

good together.

Um, so that's what got me started in

open source and

uh, that work led to the MVP, you know,

it's a recognition of basically, you

know,

uh, service to the community, you know,

it's a recognition award, right? Um, so

that's what started that and then I've

uh,

you know,

off and on I've been an MVP for uh, I

guess almost 20 years now.

>> Mhm.

>> And um,

you know, I just love the whole you're

part of a community um, and just really

that, being part of a a dev community,

you know, you're not just coding on your

own.

So, that's why I love to stay in the

program and every year I you know, um,

you know, I hope to be uh, to receive

the award.

>> So, I'm really excited Michael to kind

of get into this and for you to share

your project. So, uh, let's let's start

there. Why don't you share AI Story

Builders?

>> Okay. So, uh, this page here, if you see

my screen um, it's just showing if if

you go to AIStoryBuilders.com

uh, you'll see this page here which kind

of, you know, it's kind of a little

landing page here. There's actually a

little video that

uh

you know, gives you an overview.

But, it basically then talks about how

to um install it. And um the way you

install it is you can actually get it at

the uh

the the Microsoft Store. Again, it's

totally free. Um but, you can just

download from the Microsoft Store.

That's probably the easiest thing to do.

And there you go. However, of course,

there's a GitHub repo. So, this is

showing the GitHub repo. And here's

where you can see all the, you know, all

the code here. Of course, still

explanations, um you know, some more

documentation about it. Um so,

you can get to it uh either way.

And uh so, when you run the application,

um it pretty much looks like this. Uh

here I already have some stories loaded.

But, when you first come in, what

happens is is that you're going to first

want to go to settings, and you're going

to want to

uh select your AI provider, either

OpenAI, Anthropic, Google. Um I will say

that the Anthropic model, especially the

latest uh Opus models, that's the one

you want to use. Um yeah, it costs a

little bit more, but the performance is

just uh immeasurable.

>> It's amazing to see the innovation in

models. It seems like every day there's

a new model coming out and the new one

on the leaderboard uh as well. Pretty

exciting.

>> Oh, of course. Yeah. And remember, I

started this uh

uh project 3 years ago. So, I was using

models that were less much less capable.

So, I had to make this tool work with

those models that were

you know, looking back at them now are

quite limited. Um yet

uh

with the newer models that are capable

of so much more,

what this application can do today, um I

feel is really night and day. Um so,

once you put that in there, you just

really want to start with the new story.

And uh let me just grab uh my little

sample here that I normally use three

little pigs.

And um

And then you can select how many

chapters you want to create. It's only

going to create the first paragraph of

each chapter, but it just kind of gives

an outline. And then of course your your

AI model and again you see I'm using

Claude Opus 4 and then you just hit

create story. So as this is going,

let me just quickly

show you the challenges I had with

creating this. So initially one would

think, "Oh, I can just send to the

prompt this is the story I'm working on.

Please either fix something or write the

next paragraph or you know, or actually

some people thought write the next

chapter." And I found that that doesn't

work. And I did I had a number of proof

of concepts I created over a period of 6

months. And I remember I didn't have the

AI coding wasn't what we have today. So

I could have made those those

proof of concepts much faster today, but

back then you got to you know, put

yourself 3 years ago.

Um

So I made a proof of concept after proof

of concept and and thing everything I

tried wasn't working.

Okay? So this is a story about failure,

all right? And and what do you do when

you just something's not working? And

what I do is I just I I try to be I try

to have a little bit of humility, right?

And just see things the way they truly

are not the way I want them to be. And

the way things truly are was

I could not get AI to write

even a page, never less a chapter. I

could not get it to write something that

was coherent.

>> [laughter]

>> You know, and again the models were less

capable of that it then. Um, but it

would write something that was well it

was coherent, but it wasn't good.

>> Yeah, yeah.

>> So,

I realized that the best I could get the

AI to do is to write one paragraph.

And for it to even write one paragraph,

I had to provide so much information to

the prompt

>> Yeah.

>> that you know, the program just needs to

keep track of that.

>> Yeah, this was Yeah, I was going to say

this was my experience as well in the

MVP program as I would have to write

huge prompts including do not do this,

do not do this, do not do this, right?

And I would like through process of

elimination, it would get better, but

enter much better models nowadays,

right?

>> Okay, enter much better models. However,

um, and hopefully [clears throat] we're

going to touch upon this uh

is

um

I found that it's still about how much

you give the prompt, you know? As we

know, AI is just an algorithm. It

doesn't really think. It's an algorithm.

It's a calculation. And therefore, what

you put in is what you get out. The base

program was was based upon this

structure that I'm showing here, which

is I realize I have to break story down

into core elements. And

it turns out that it's truly in this

order. So, this image here is just

saying this is story, right? And obvious

things, a story consists of chapters and

and characters and look, you know, and

locations and that sort of thing. So, we

know that, you know, people know a story

has characters and that you need to keep

track of the characters. Okay, that

part's obvious.

But that's actually not the first part

of story.

>> Yeah, yeah.

>> That's what I found in my experience.

The first part of story

is timeline.

When So, even in a story you have

flashbacks, right? Um, so a flashback is

is controlled by time. So, if the

character is 5 years old in the

flashback, that is controlling, okay?

Everything has to hang off of of that

timeline. So, I realized that you start

off with timeline first. So, in my

database structure, I start off with

timeline. After timeline, even before

you get to characters, is location. So,

the when and the where

>> Yeah. Yeah.

>> With with fictional stories,

um

is or actually even a true story, but

even story versus say you and I are

talking right now, story is controlled

by timeline and location.

Even if the Even if the location is

unspecified, it is still a location. You

know what I'm saying? Um for example,

it's all it's unknown to the to the to

the reader, but it is a location that is

distinct from say another location. So,

these two things are something I

discovered. Um so, timeline, location,

and then we get to character. And then

with character, you have attributes,

things like the hair color, um

their the background history, and all

that sort of thing. But, it goes in this

order. So, for example, a character's

attributes, whether they have red hair

or blue hair, is on always on a

timeline.

>> I see what I love about this, Michael,

is that you're you are defining a

combination of, you know, relational

data that is explicit, that is you're

you're telling the model this is what

you want in the story or you're helping

generate it. But, then you're allowing

the model to then bring these different

pieces together to form, you know,

paragraphs or you know, to to write

>> Right. Yeah, I feed those Right, I feed

those to the model. Um and then of

course I I

Uh you have chapters which uh compose of

paragraphs. Um another decision I did

make with this application is

um I

didn't want to have a special database

um

So, I actually use text files. So,

everything is using uh the text files to

store all these elements. Um I course

use embeddings so that I can use cosine

similarity

to surface you know things that are

related to each other. And you know,

that's pretty much the

the base of the application and then now

you can see with the three little pigs

here that we started off with. It

creates a little synopsis

and this course is fed to

the AI in the prompt. You can also have

system messages and these are things

like you know, don't have foreshadowing

maybe the the

perspective is third person limited.

Again, this sort of thing you can change

and you can of course even have world

facts things like you know, magic is

real and and that sort of thing. But

then you see it breaks things down. So

even with that little paragraph that I

fed to it, it said hey,

you know, there's different timelines.

Farewell to mother

you know, the three little pigs leave

their mother. They're constructing their

houses. The wolf arrives. We know this,

right?

Mother is a flea to safety. The wolf

fell to tax because this course has a

happy ending and the wolf's demise.

You have the locations. The pigs straw

house, the mother's forest home that

they live. The pigs stick house, the

brick house, right? These are the

locations that you would expect in the

story. Then you have the characters and

the three little pigs Joe, Tom and the

wolf. Um

and

and then you have the mother course is

off screen I guess in this first

And the course you have the chapters.

Chapters have a synopsis. Apologize if

this is hard to read but just it's there

and notice we only have the first

paragraph of each of the chapters. And

the reason for this again is is that I

found and this even with the new

frontier models

you only want

the AI to help you write one chapter at

a time. So what happens is is that you

can come in here and you can kind of

start and work on the next chapter. Um

or you can just click this AI button and

describe what you want.

Or you can just click this button and

it'll just continue from the previous

chapter.

>> I like that.

>> Right. And what's happening here is is

that I'm feeding to the AI

all the context of okay, you're working

on a chapter. Um and of course let's go

ahead and save here. Um working on a

chapter and this previous chapter

these are the characters that are in

this chapter. Um

this is the location that's this

chapter. So therefore everything that's

you know, tied to that location. Um of

course this is the timeline to that

chapter and everything is tied to that.

Um so it's all fed to the AI and then

the AI of course writes this prose.

Which again, the frontier models this

prose would be

it's better than say a uh

uh you know, a less capable model. But

if I tried to tell it to write too much

it'll write something good, but it'll

kind of get off and it's just not a

human, you know.

Um and one more quick thing. Um

updating this the database of of say the

characters and and uh you know, all

their uh

you know, elements. While that can be

time consuming, what you can do is you

can simply click um auto detect

attributes. And what happens is is that

the AI will look at this this and say,

"Hey look, there's something new and I

can just you know, add that to the

database." So that's a way for the you

to add things to the database uh

quickly and easily.

>> Yeah, I suspect

Yeah, I suspect that as you write more

chapters, the AI gets better and better

at writing chapters aligned with what

you know, what your vision is.

>> cuz you're giving it more information.

You are, yes.

>> Yeah, yeah. So like so say you write

chapter one, but then you make human

refinements to it. Say, I don't know,

say you edit it by 50% compared to what

AI did. But then the next chapter you're

going to edit it by 25% and the next

chapter you're going to edit it by 10%,

right? Because uh theoretically or or,

you know, as we think about this like,

you know, um it it understands the

context for how to make the story more

focused, especially when you give all

these attributes like timelines and

locations and characters and such, you

know.

>> And a couple of other quick features

about it. Um you can import uh an

existing story, whether it's uh you

know, Word, a PDF, or text. So, that's

why I um imported, for example, The

Great Gatsby. It's in it's in uh

you know,

uh public use. Um

>> Yeah, one of my favorite books. It's one

of my Yeah, one of my favorites.

>> Right. See, so it's all uh here and of

course it broke down, you know, these

are the characters.

Um

You You'll see there was actually a

little bit of uh

uh issues here where, you know, Daisy

and Daisy Buchanan and Daisy or it

thinks it's two different people. So, um

this is something I'm still working on

here. Um

but yeah, and these are all the

different locations and that sort of

thing. So, you can import an existing uh

uh story and of course you can um

uh export um your story to um

you know, as a Word document.

>> You know what I like about, you know,

The Great Gatsby in this case and some

of the other stories that you've

imported is uh it's one thing to write

The Three Little Pigs, which is a fairly

simple story with, you know, a few

characters, a few settings. Uh but when

you get to The Great Gatsby or, you

know, J.R.R. Tolkien or other more

complex things, it it articulates the

scale potential of AI. An AI-based

solution like this can help us keep our

own story together and uh avoid

inconsistencies in the story, which I

think is is especially as it gets more

complex.

>> Right. So, like here,

uh this right here is the last paragraph

in The Great Gatsby, and you can

basically say, "You know what?

Continue the story."

>> Yeah.

>> And and this will actually uh you know,

literally

continue the story of The Great Gatsby.

So, it's kind of wild when you kind of

think about it.

>> fan fiction.

>> Yeah, exactly. Yeah, fan fiction,

exactly.

>> [laughter]

>> So.

>> Well, you mentioned Michael a little

earlier about, you know, others, you

know, for our viewers and our listeners

out there that that you all can build

this. Um tell me a little bit more about

like how would you how did you build

this?

>> So, if you look in your hard drive, your

documents folder, you'll see a folder

called AI Story Builders, and there

here's where you'll see all the

uh folder for each story, and then

you'll see uh folders for each part of

the story. So, for example, I go into

chapters, and then you'll see folders

for the chapters, and then when I go in

there, you'll see um a text file for

each paragraph, and if you open up the

paragraph, you see something looks like

this, which is these are the characters

that are in that paragraph. This is the

prose of that paragraph, and here are

the embeddings uh

that represent that paragraph, and

the program literally

you know, opens these files up, reads

these embeddings, uses rag, uh which is

retrieval augmented generation, which

then uses a calculation called cosine

similarity to then determine what uh

paragraphs are related to each other,

and therefore it then creates the prompt

that is then sent to the AI.

Which leads us into the why I'm now

using knowledge this I'm still using

this, but I'm also now with rag using

knowledge graphs to

uh what I what in my opinion really

makes the this program 10 times better.

>> All right, so Michael, uh, tell me a

little bit more about this new graph

feature, uh, that you are building. I'm

pretty excited to see, um, you know,

what you've learned and how you're

growing and how you're improving this,

uh, this overall AI story builders.

>> The first thing I did was I,

uh, went into, you know, kind of I have

coded an app and I said, "Hey, look at

my existing AI story builder files and

create a, uh, knowledge graph

>> Okay.

>> and it created uh, something that looks

like this.

>> Yeah, so [clears throat] this is a JSON

file, um, that then is going to

aggregate all this, all this data into a

visualization.

>> Right. So, this is like the raw thing,

okay?

>> Yeah. Yeah.

>> However, I then vibe coded, uh, an app

that where I basically said, "Hey, can

you, um,

show me,

um,

let me just start here.

Uh, can you just visualize this graph?"

So, here, uh,

using the same Sherlock Holmes,

I

basically come up with something. I know

this is impossible to see, so don't

worry, I'm going to zoom in here, but

this is Sherlock Holmes, you know, that

knowledge graph I just showed you. This

is it visualized.

>> All right, so Michael, now you're, uh,

visualizing the adventures of Sherlock

Holmes. Uh, tell us a little bit more

about how, you know, an author would

utilize this to kind of keep, uh, the

story straight and understand where to

go next.

>> Right. So, this right here is just a

visualization of the graph just to get

an idea. Um, a graph, of course,

consists of entities, like these are the

characters, and then you have the edges,

which contain information that this

character interacts with this character,

um, and so on. And of course, um, we

also have everything broken down, um, by

paragraphs so that you can see what

characters interact [clears throat]

with, uh, which paragraph, uh, with with

that. So, what happens is is that this

graph is then exposed to the AI

as tools.

>> Mhm.

>> So, therefore, the AI is able to use

those tools to make um as many calls as

it needs to interrogate this graph.

So, what I've added is a chat window

that allows you to do things like

indicate the interactions between Watson

and Holmes in chapter one.

>> Yeah.

>> is it's able to come up with this um

very, very detailed uh response that I

would not what I normally would not be

able to surface using simple retrieval

augmented generation because with rag,

I'm limited by how much I can can

surface, you know, to the prompt.

Whereas here, the frontier models using

their advanced tool uh calling

capability, it can make as many calls as

it needs to that graph and therefore

surface everything that it needs. Um and

what I've also added is the way to go um

the other way, whereas you can say,

"Okay, you know, you've identified

issues. Please update um

you know, the story or update the

chapter to, you know, fix those issues."

>> Mhm.

>> So, therefore, I've exposed these tools

that allow it to update the files. So,

this

you know, so this capability is what's

being added. Um I should be releasing

that this weekend.

>> I love this cuz you're you're asking

questions to keep your story straight,

but then you're using this as context to

then write more story and have that be

more accurate as well.

>> Right.

>> Wow.

>> So, I look forward to the hopefully the

community uh will uh will like this

feature and then also can see how they

can adapt this technology to their own

applications by identifying uh things in

their applications which uh

um they can identify, you know, entities

and edges to create knowledge graphs and

then expose that knowledge graph to the

AI through tools that then allow their

applications to um you know, just go

deeper than just mere rag can can

currently provide.

>> Well, Michael, thank you so much for

sharing uh this application and sharing

your passion, being a Microsoft MVP. Uh

this is a fascinating uh project. And

so, I really appreciate you sharing it

today and thank you for being here on

MVP Unplugged.

>> Thank you for having me.