Full transcript
0:00The single biggest bottleneck for most
0:02authors is time. We have the ideas, we
0:04have the outlines, but turning that into
0:06a polished book is a marathon that can
0:08take years. What if I told you you could
0:10turn that marathon into a 4-day sprint?
0:12Hi guys, my name is Jason. I'm the nerdy
0:15novelist. I used to work for Kindle
0:16Prneer, one of the largest author
0:17training websites, which taught me a
0:19whole lot of things about the author
0:20process. And uh now I teach authors how
0:23to write with AI in order to get their
0:26ideas onto the page using it as a
0:28productivity tool without compromising
0:30your creativity or your ethics. And um I
0:34ran an experiment on my own novel. And
0:36in this video I'm going to show you how
0:38I used AI to create a genuinely clean,
0:41well-written first draft in just a few
0:42days. And why I believe this could allow
0:44me to publish a new book based on my own
0:46ideas every 2 weeks or so. Now, when I
0:48say I wrote a novel in four days, I need
0:50to be completely honest. I did cheat a
0:52little bit. Uh the foundation of the
0:54book, meaning the detailed outline, was
0:57something I had written a long time ago.
0:59So, I've had that in my back pocket for
1:00a while. Um this wasn't just a throwaway
1:03story. This was book five in one of my
1:05established fantasy series. It's
1:07actually this fantasy series right here.
1:10Uh this fantasy series is really
1:12important to me because this is the
1:14fantasy series that I burned out on. I
1:16wrote the first two books. Uh I was
1:18halfway through the third when I
1:20released the first two books and it
1:22absolutely tanked. Um and the reason was
1:25it was I had not targeted the genre very
1:27well and uh the cover that you see and
1:29the title uh that you see here is
1:31actually different than what I initially
1:33published. Um but I was expecting a lot
1:36from it. I was expecting to make a lot
1:37of money from it and it ended up doing
1:39worse than the previous series that I
1:41had published which was my first and so
1:44writing book five it's actually the
1:46first one I've completed in years and I
1:48used AI to do it and because it was book
1:51five in a series it was very important
1:52that it feel like my story and that it
1:55sounded like me. So that was a priority
1:59for me as I was going into this project.
2:01Now the engine behind this entire
2:03operation is an automation that I built
2:05in a tool called N8N. Think of it like a
2:07digital assembly line. You create a
2:09series of steps that take information
2:11from one place, process it and send it
2:13somewhere else all automatically. And
2:16for this project, its job was to feed
2:18the AI into a variety of different
2:20models u in different steps getting
2:23everything it needed to write one
2:25chapter at a time exactly to my
2:26specifications. And for every single
2:28chapter it wrote, it was given the
2:30detailed scene by scene outline, full
2:32character sheets for everyone involved,
2:33a worldbuing bible, and even samples of
2:35my writing from my previous four books
2:37to use as a stylistic guide. It was a
2:41massive amount of information, but it
2:43was necessary to ensure that Claude and
2:45the other models didn't just write a
2:47generic fantasy scene, but a scene that
2:48belonged in my universe and sounded like
2:51it came from me. I want to quickly show
2:53you what this automation looks like.
2:54It's the kind of thing that looks
2:55complicated at first glance, but the
2:57logic is pretty simple once you break it
2:59down. So, this is the automation right
3:02here. There are a number of steps here.
3:04The first few, you can see these are all
3:06just grabbing documents from Google
3:08Docs. Uh, I use Google Docs for all my
3:10automations. I find it's the simplest
3:12and easiest and, uh, most people have
3:14Google Docs, so they're able to do this.
3:16Um, but basically what happens is I just
3:19click a button uh, this one right here.
3:21It says execute workflow and it will run
3:23through this process and uh I
3:26specifically give it my outline, my
3:28characters, my world building, sample
3:31pros, a stylesheet based on that sample
3:33pros, a list of forbidden words that I
3:37don't want it to be using and so words
3:39that AI uses quite a bit. Uh as well as
3:43I I bring in the document where I want
3:46the finished text to go. So, if you're
3:50just starting on this project, this
3:52document will be one that you create and
3:54is blank when you start. But you'll need
3:56that. And the re the way I'm able to
3:58call all of these things is if you click
4:00on one of these, it has the document ID,
4:03which you can just get in the URL of
4:05whatever document you're doing. U you do
4:07also have to connect NADN to your Google
4:10Docs account. And this this overview is
4:13I'm not going to tell you how to set up
4:14NADN. We do have resources for that in
4:17my story hacker gold group. And I'll
4:19make more videos about uh n and
4:22everything. That's not really the
4:23purpose of this video. I'm just sort of
4:24showing you the workflow. So, we pull in
4:26all of these documents. And then I use
4:29this node. Uh it has a prompt here that
4:32says, "You are a parser. Given the above
4:34book outline text, return a JSON array
4:36named chapters containing each chapter
4:38title exactly as it appears, one element
4:40per title, including those without
4:42numbers like prologue or epilogue. Make
4:44sure the chapter numbers are also
4:46present where listed. do not include any
4:48additional keys only output valid JSON
4:50and sometimes I'll say actually I
4:52specifically want you to do these
4:54specific chapters not all of the
4:56chapters what that does is it then makes
4:58a list of chapters and from there they
5:01get parsed in this node right here and
5:04we start a loop so this this thing right
5:08here uh goes through each of those
5:10chapters that were listed in the
5:12previous step that were parsed out and
5:14it creates a loop that then goes goes
5:16chapter by chapter through each of them
5:18and writes the whole thing. So, in
5:21theory, you could give it an outline and
5:23let this go and it would go through your
5:25entire book and write the whole thing
5:26all at once in uh and yeah, I found that
5:29that took me about 2 hours for it to go
5:32through a 40 chapter outline and
5:34actually write all 40 chapters. That was
5:36expensive. I don't actually recommend
5:37doing it all at once like that, but it
5:39was uh you know, it was for science. Um,
5:43so then, um, if you're writing the first
5:45chapter, these two steps won't matter,
5:47but, uh, I like it to be able to read
5:49the previous 2,000 words or so. Uh, you
5:52could also change this to be three,
5:54four, 5,000 words, however many you
5:56want. Um, but I just like it to know
5:58basically what happened in the last
6:00chapter, so it can continue off of the
6:03end of that last chapter in a way that
6:05is uh that makes sense, that is
6:07seamless. Uh so we grab the last 2,000
6:09words of the chapter and then we run
6:11through these four steps uh four AI
6:16uh models different models for each of
6:18these except these last two are the
6:20same. Uh and once it goes through those
6:23four steps it then takes that chapter
6:25that it has written and it plops it into
6:28your document that you pulled in
6:30earlier. So we this note is just called
6:32add chapter to document. uh and then it
6:36goes to this and replays the loop. But
6:38on the next loop, it if you started with
6:40chapter one, it will go on to chapter
6:42two and so on. Uh so it's a very handy
6:46loop. Uh you can kind of just press
6:49execute and don't worry about it until
6:51it's finished the number of chapters you
6:52want. Uh but these four steps, there's
6:55four of them. There's um the first one
6:57is creating a scene brief. So I'll show
7:00you a little bit of what that prompt
7:01looks like. So, we pull in the outline,
7:03the characters, the world building, the
7:04previous chapter text, and then I have
7:06it create a scene brief to sort of spell
7:08out what should happen in this specific
7:10scene, and uh who are the characters
7:13involved, what are they doing, how are
7:15they, what's their mood in this
7:16particular chapter, what are they
7:18wearing in this particular chapter,
7:19because their physical appearance might
7:21be slightly different from one chapter
7:22to another. And this was actually useful
7:24because I had like a character that got
7:27injured in one chapter. And because I
7:30was doing the scene brief and it was
7:31able to look at the last 2,000 words, it
7:34knew that the character was injured um
7:37and it was also looking at the outline.
7:38So, it was able to derive from that that
7:40the character was injured and it was
7:41able to pass that information on into
7:43the next chapter it was writing. Had it
7:46not done that, um it would have just
7:48like he would have been injured in the
7:50previous chapter and in the next chapter
7:52he would have just been fine. Um, so,
7:55uh, it's important to be able to let it
7:58have that level of awareness of the
8:02chapter. So, it creates the scene brief,
8:04uh, to give it all of the details that
8:07it needs for that specific scene. And
8:10then we move on to the most important
8:12step, which is the first draft. Uh, by
8:14the way, the I'm using Sonnet 4.5 for
8:17the scene brief right now, but for the
8:19first draft, I'm using Opus 4.1. Um, it
8:23opus is an expensive model and I
8:25literally throw the entire kitchen sink
8:27at it in this particular prompt. So, it
8:29is this is not an this is an expensive
8:32step. It would cost me about 75 cents
8:34per chapter for just this step.
8:36Everything altogether was costing me
8:38about a dollar per chapter. U, but it's
8:40worth it to me because I get good pros
8:42out of Opus, better than than Sonet 4.5
8:45and a lot of any other model that I've
8:47tested. Um, so first I give it the style
8:50sheet and the style pros examples from
8:53my text. I also give it the prohibited
8:56words that I don't want it to use. I
8:57give it full context including the
8:59entire outline, the characters, and the
9:00world building. I give it the previous
9:022,000 words of text so it knows where to
9:04start off from. And I give it the scene
9:06brief that was generated in the previous
9:08step. I then go through a couple of
9:10instructions here on what to do and let
9:12it go. And then from there
9:16it writes the first chapter. Now, as we
9:19all know, if we've worked with AI a lot,
9:21there's all kinds of little AISMs and
9:23little quirks and cliches and uh things
9:26that are a little bit dull about uh AI
9:29writing. And while Opus 4.1 is very
9:32good, it's got better like scene
9:35cohesion and everything that I've seen
9:36from any other model, it still has some
9:39of those things in it. And so these next
9:40two steps are meant to polish that out a
9:43little bit just to make it a little bit
9:46better for when I then go through it.
9:50Um, so it would go through this
9:51improvement plan and then it would go
9:53through and using that improvement plan,
9:56it does a rewrite. The improvement plan
9:57looks something like this. It's uh I've
10:00got a list of things I wanted to look
10:01for. So like metaphors, identify any
10:04metaphors and instruct them to be
10:05removed. Adverbs. Identify any over
10:08reliance on adverbs. Dialogue tags.
10:10Identify dialogue tags that are not said
10:12or asked. Make suggestions for which
10:14ones should be changed to said or asked.
10:17Um, flag instances of passive voice. So,
10:19stuff like that. I'm just having it go
10:21through and polishing this up just a
10:23little bit so that I don't have to do it
10:26later. And so, it goes through and this
10:28step crucially does not do any rewrite
10:31rewrites. All it does is it highlights
10:34the problems and then in the next step
10:37I'm also not having it do a rewrite. I
10:40ask it to simply implement the suggested
10:43changes but do not in change anything
10:45else about the original chapter. Uh re
10:48reproduce the entire chapter with the
10:50suggested changes. I found that if you
10:52use the word rewrite it will rewrite the
10:54whole thing and it's often worse uh
10:57depending on which model you're using
10:59than the original draft. So instead of
11:01asking for a rewrite, I ask it to just
11:03implement the the changes and not change
11:05anything else. Um, so that's all this
11:08prompt does. And it so it goes through
11:10those and it reproduces the chapter at
11:12the end with all of those changes made
11:14and then adds that chapter to the
11:15document. And then it loops back around
11:17and starts on the next chapter and it
11:19continues to do that. So you could do
11:20this for the entire book all at once.
11:22Um, the reason I don't recommend that is
11:25because I found that uh if it got a
11:27little off on one chapter, and there
11:30were actually instances where it
11:31actually wrote the wrong chapter, I've
11:32since corrected that, but um that was an
11:35issue as I was going through this
11:37project. Um, if it does get a little
11:40off, like even if it's off by one degree
11:42in the next it might be off by two
11:44degrees and then three degrees and then
11:45it just gets worse and worse as it goes
11:48along. Uh, and so I find it's good to
11:51write um, you could go one chapter at a
11:54time and uh, and do that. Uh, that would
11:58take a while though and then you'd have
11:59to wait for the generation to happen in
12:01between each each chapter. So, I found
12:03the sweet spot for me was two to three
12:05chapters where I would have it generate
12:07all three uh while I'm doing something
12:09else and then I'll go through and revise
12:11those chapters and then after I've re
12:13revised those chapters, I have it do the
12:15next two or three chapters and I'll go
12:17to do something else and then I'll come
12:19back and revise those chapters and that
12:21way I was able to get through a bunch of
12:22different chapters all at once. Of
12:24course, every one of these chapters,
12:25especially the ones that I had to
12:26regenerate, came with a literal cost.
12:29The Claude Opus model is not cheap and
12:31it's the best large language model for
12:33creative writing on the market right now
12:34in my opinion, but you pay for that
12:37quality and each chapter cost me on
12:39average about a dollar to generate as I
12:41mentioned. And so for a 40 chapter book,
12:43that's 40 bucks. Factoring in the
12:45handful of chapters I had to regenerate
12:47due to mistakes, the total cost for the
12:49raw first draft was probably closer to
12:5050 or 60. Now people might b at that,
12:53but let's put it in perspective. Would
12:55you pay 60 bucks to have a complete
12:5780,000word first draft of your novel
13:00appear on your computer tomorrow? Like I
13:02have worked with ghost writers and paid
13:04them four $5,000 for a good draft. Now
13:07granted, it is a little bit cleaner and
13:09I can expect more from a ghost rider and
13:11have to guide them a little less than I
13:13do with AI, but at the same time,
13:16[snorts] $40 versus $4,000 is a big
13:19thing for most people. Most people are
13:21not going to be hiring ghostriters,
13:23right? um you're either going to be
13:24writing it yourself or you're going to
13:26be using AI to do it. And so for me, $40
13:29is nothing. $40 is absolutely worth a
13:32clean first draft. And you're going to
13:33get a lot cleaner first draft using Opus
13:36than another model. I've tried this with
13:39uh using Sonet 4.5 for the actual pros
13:42generation, and it just isn't the same.
13:44It's not as good. It's not bad. Uh, but
13:47you're going to be paying for it later
13:49with more time to clean up those drafts
13:51than if you used Opus. At least that's
13:54been my experience. By the way, if you'd
13:55like to learn more about how to use AI
13:57as an author-led productivity tool to
14:00write uh at least 12 books per year, if
14:02not more, and without compromising your
14:04creativity or ethics, go check out the
14:05Story Hacker Gold link down below. You
14:07can also get this exact model that I'm
14:10using uh to use on your own and uh
14:14implement into your own books. So, with
14:16the raw manuscript complete, it was time
14:18for the part of the process that most
14:20authors dread, and that's the first big
14:22edit. But this is also where the title
14:24of this video comes from, because the
14:25actual writing, the generation of the
14:27text only took a few hours of total
14:28processing time spread out over a day or
14:31so. And the human part of the work, the
14:33editing, only took me 4 days. Uh, I had
14:36a simple process. This was a 40 chapter
14:38book. And so, I set a goal of editing
14:40roughly 10 chapters a day. That didn't
14:42take me that long. Um, and it usually
14:45took me between two and two and a half,
14:47maybe three hours for some of the more
14:50complicated parts where the AI really
14:52messed up. But in that amount of time, I
14:54was able to get through 10 chapters and
14:57a lot of it was really good. And there
14:58are actually quite a few moments that I
15:01thought were actually excellent inside
15:03of the book. And and uh I can attribute
15:05this to Opus because Opus is the only
15:06one that does this sort of thing. For
15:08one, I had a moment when two of the
15:11characters were having this quiet
15:12philosophical debate between the merits
15:15of justice versus mercy. And I looked at
15:19the text that it gave me and it was
15:20actually a genuinely thoughtful
15:22philosophical debate. A lot of which was
15:25not provided by me, at least not beat
15:28forbeat, right? Uh I just said they they
15:31debate justice and mercy and this person
15:33has this view and this person has that
15:35view. But then when they came together
15:37um it it was actually a genuinely good
15:39conversation. There was another moment
15:41also when my main character uh she has
15:43this magical ability to uh mirror the
15:47abilities of others and she comes in uh
15:50into contact with this kind of
15:52Lovecraftian entity that's really
15:54horrifying
15:55and she tries to mirror its ability and
15:59it it just uh it just uh it it does not
16:03go well for her and she starts to
16:05transform into this mass of ugly things
16:07and it did this layered
16:10uh really strong showing instead of
16:13telling scene where it just described
16:17her transformation. And it also layered
16:19in like the uh the reactions of other
16:23characters around her because it was
16:25from one of their perspectives watching
16:27her make this transformation. And um and
16:31there are other characters who don't
16:32realize what's going on and they're just
16:34kind of like laughing until they figure
16:36out what's going on and then then they
16:38like start freaking out. And all of the
16:40reactions from each of the characters
16:41cuz there were villains there and allies
16:43there. They were all unique and separate
16:46to each other. Um and not just like the
16:48same reaction from each one. And so I
16:52got moments like that that I was just
16:53like flying through it. I was just like
16:55this is amazing. And then there were
16:57others that where it was not so good and
16:59I had to really uh tighten it up and
17:01spend more time on it. Uh and as I
17:04mentioned there were a couple of times
17:05one of the one of the problems that I
17:07noticed is that sometimes it would write
17:11the wrong chapter. Uh and that's because
17:13in the scene brief steps sometimes it
17:15would get the wrong one and so it would
17:17say okay we're writing this scene and in
17:19fact I wanted a scene that was like
17:21three chapters earlier. That seems to
17:23have been solved with a tweak of the
17:24prompt. But I haven't had a problem with
17:25it since. Uh, but that was something
17:27that happened that caused me to have to
17:29regenerate huge sections of the novel in
17:33order to get it right. And it was
17:35through that process that I figured out
17:36that going two or three chapters at a
17:38time might actually be better than
17:39trying to do uh huge sections at a time.
17:42So, the only real downside to that is
17:44that you you can't like edit 10 chapters
17:48at once because you have to edit a few
17:50then generate more then edit a few then
17:52generate more. And so you're going back
17:54and forth between
17:56um editing a text and then you have to
17:58go do something else for a while while
18:00it's generating the next few because it
18:02does take a little while. It's not like
18:04something where you're going to be um
18:08sitting around for just a few minutes
18:10and then you get the next chapter. Like
18:11it takes for for like three chapters it
18:14would probably take uh 15 20 minutes.
18:19Um, I'd say probably for each chapter it
18:21took about five minutes to go through
18:22all of those steps. [snorts] Um, that's
18:25my that's my rough estimate. Maybe maybe
18:27more. Uh, but yeah. So, I don't like to
18:31be constantly bouncing between things uh
18:33to try and multitask because humans
18:36actually can't multitask properly. So,
18:39uh, I, you know, if I'm going to be
18:41editing text, I'd like to just sit down
18:43and edit all 10 chapters, but that
18:45wasn't really ideal with the way this
18:48worked. But regardless, I was still able
18:50to get it done in four days. 10 chapters
18:52a day. Didn't take me that long overall.
18:55The only reason I couldn't do more was
18:56because I also like I film YouTube
18:58videos and I'm doing other marketing
19:00things and I'm uh administrating in my
19:03group and like all this stuff that I do
19:04in addition to this. Um, but this, you
19:08know, it it was totally doable. And I
19:09think anybody, even if you have a
19:11full-time job or something like that,
19:12you could probably, if you could spare
19:14an hour a day, you could get through
19:17this in a week or so. And if I were to
19:19factor in the cost, uh, the time cost of
19:22creating an outline,
19:24uh, as well as doing like secondary
19:26edits and, uh, formatting and
19:28publishing, I think this would probably
19:30take me about two weeks. Now, I have
19:33sent this manuscript over to a line
19:34editor just to do a second pass on it.
19:37Uh, if you don't have the budget for a
19:39line editor, then uh you would want to
19:42do at least two more passes, I think, on
19:44on it. So, that would take extra time.
19:47Um, but I already knew just from having
19:49read through it uh that I would probably
19:52glaze over a lot of the problems uh that
19:55were that were left in the manuscript.
19:57So, I wanted to get a uh objective third
20:00party to come in and and do a pass on
20:03it. Um, and that's just more for my own
20:06peace of mind that I know that the the
20:08text is going to be decent. So, if this
20:10process sounds exciting to you and you
20:12want to learn my entire framework for
20:13going from a blank page to a finished
20:15book in hand, including how to market
20:17it, you should check out my school
20:18group, Story Hacker Gold. And in there,
20:20we dive deep into how to use AI and
20:21other tools to enhance your process
20:23without ever compromising your
20:24creativity as an author. The group also
20:27includes weekly live calls with me, all
20:29of my ebooks and audiobooks for authors,
20:30and a community of writers who are all
20:32pushing the boundaries of what's
20:33possible. So, if that's interesting to
20:35you, the link is in the description and
20:37I will see you in the next