Full transcript
0:00Here are four prompting techniques that
0:02will make it so that your books will go
0:04from being obvious AI slop to being
0:08something that nobody could guess is
0:11actually written by AI. It's that good.
0:13I use these techniques heavily in my
0:15automations. I'm doubling down on all of
0:17them as I'm going through and creating
0:20my own tool for writing with AI. And
0:24this is, I believe, these these four
0:26techniques, if you can figure this out,
0:29this is going to be extremely helpful
0:31for you in the writing process. Now,
0:34most people are not using these
0:36techniques because, frankly, if you're
0:37using them inside of a chatbot or
0:39something like that, it might get a
0:41little tedious, and most people aren't
0:43into that. And that's why I use
0:46automations for a lot of these because
0:47it can run through several of these
0:49steps and techniques
0:50without me really having to supervise
0:52it. It just does it automatically. But
0:54you can absolutely use all of these in
0:57the process. So, let's dive into these
0:59four techniques and figure out how can
1:01you level up your writing using all of
1:04these. The first is by utilizing checks
1:07or critiques of your work. Now, this
1:10might sound simple, but a lot of people
1:11get this wrong. Let's say you've written
1:14a chapter using AI based on what you
1:17told it to do. There are going to be a
1:19lot of issues with that chapter right
1:21out of the gate. It's going to have a
1:23lot of AI-isms in it. It's going to
1:24probably have some poor writing examples
1:27in there, like maybe a few many adverbs
1:29or something like that. And you,
1:32ideally, would love to filter out as
1:34many of those things as possible so that
1:37you don't have to change it manually.
1:38And when you go in to fix it up
1:41yourself, you're only focusing on those
1:43aspects of the the world, the
1:46chronology, and just making it match
1:48your own vision as much as possible, and
1:51maybe just adding a little bit extra
1:53human
1:54flare to it
1:56uh that the AI wouldn't necessarily get,
1:58but we can at least minimize a whole lot
2:01of the issues the AI has, so you're not
2:02just playing whack-a-mole with all of
2:05these issues, that you're actually
2:07starting with a fairly clean draft from
2:09for you to work on. The way this would
2:11work is you take your chapter and you
2:14would run a prompt on it that says
2:16something to the effect of I want you to
2:18search for all instances of superfluous
2:21adverbs and come up with an improvement
2:24plan to remove these adverbs. Now, this
2:27is a key point of a check like this.
2:29We're not going to ask it to rewrite
2:31anything. We're just going to ask it to
2:33create a plan on what it would change.
2:37This allows it to focus on a very narrow
2:39set of parameters, and this is key to
2:42the uh to the checks approach is that
2:45ideally you want to focus on just a few
2:47things with each check. And thankfully,
2:51if you narrow it down enough, you can
2:53use a really inexpensive model, like say
2:55Gemini 3 Flash or or one of those to do
2:59this check.
3:00Uh because you've narrowed it down, like
3:02let's say we're only focusing on
3:04adverbs. You could have a check that
3:05includes a giant list of a whole bunch
3:07of things that you wanted to watch out
3:09for, but that runs the risk of it
3:12getting uh too spread out in what it
3:14does. If we focus what it what it looks
3:17at, then we can do this, and then we can
3:19run multiple checks
3:21uh to fix this thing up. So, you ask it
3:23for an improvement plan, and then in a
3:26second step to this, you ask it to
3:28incorporate the feedback from the
3:30improvement plan, but to change nothing
3:32else other than what was on the
3:34improvement plan. This is another key
3:36aspect to it because if you don't do
3:38that, it's likely to rewrite the entire
3:41thing and do it probably worse than the
3:43original uh draft was, especially if
3:46you're using a less powerful model in
3:50that redraft process. But, if you just
3:52asked it to change what was in the
3:54improvement plan and leave everything
3:55else the same, you can use a lesser
3:57model to essentially improve your your
4:02draft. So, this is a technique that not
4:05enough people are using, partly because
4:07it takes quite a lot to go through all
4:09of these checks if you're just using a
4:11simple chatbot, but you can do it. This
4:13is why I do it inside my automations
4:15where I will just have a series of these
4:17steps that it goes through and it just
4:18does them automatically. And you can
4:20have all kinds of checks. You can have
4:22checks for adverbs, check for dialogue
4:24tags, checks for the logic and
4:26plausibility of the scene, checks for
4:29various types of AI-isms. Like, you know
4:31that thing the AI does a lot where it
4:33says, "It's not just this, but it is
4:35also that."
4:37That's something the AI does quite a
4:38bit. You could create a check that
4:40specifically looks for that particular
4:43type of phrasing and gives you an
4:46alternative to say instead, and then you
4:49just run that check and it filters out
4:51all of those things. It's a great way to
4:53to utilize humanization in this whole
4:55thing. So, that's the checks or
4:57critiques process and creating the
5:00improvement plan. It's one of the most
5:02powerful ways to improve your writing as
5:04long as you have
5:05a solid series of checks and prompts for
5:08each one to go through and filter these
5:10things out. All right, the second key
5:12prompting technique that I think is much
5:14more important than people give it
5:16credit for is the value of context
5:19engineering. So, context engineering is
5:21this term that's come up recently, this
5:22idea of just like it's not really about
5:24the prompt so much as the quality of the
5:27context that you give the AI. So, from a
5:31fiction writer's perspective, this is
5:33pretty obvious. This means like you give
5:35it all of the information about your
5:37world, about your plot, about your
5:38characters that you want it to know. If
5:40you don't give that information, it
5:41won't know you know, it'll come up with
5:44something more generic that
5:45doesn't match your vision.
5:48Uh it actually goes a little bit beyond
5:49this. For instance, if you're writing a
5:51chapter and you're providing it with a
5:54whole bunch of information, it might
5:56pick the wrong bits of information to
5:58write that chapter. Instead, we kind of
6:00want to curate what gets fed into the
6:03AI's input when it's writing this
6:06chapter. So, maybe only the characters
6:08that are in that scene or just the key
6:11details that it needs to know for that
6:13scene. If you're using a tool like
6:15NovelCrafter, NovelCrafter actually is
6:17excellent at managing your context and
6:19making sure that only the little bits of
6:20information that the AI needs to know
6:23for that scene that you're writing is
6:25the stuff that the AI knows.
6:27Um NovelCrafter can be a little bit of a
6:29complex platform, but it is excellent at
6:31context engineering. If you're running,
6:33say, an automation, which is how I do
6:35like 90% of my AI writing these days,
6:38I would put some steps in there that
6:40kind of look at what needs to be written
6:42and then looks at the total character
6:45sheet and all of that and just selects
6:47the pieces that need to be used in that
6:50particular scene, and then those pieces
6:52are the ones that are passed on when
6:54writing the chapter. Another key aspect
6:57of context engineering that you need to
6:58keep in mind is not giving it too much
7:01context. This is not just because you
7:03don't want to give it information it
7:04doesn't need, but also because this can
7:06actually water down the prompt. I've
7:08seen too many of examples of people
7:10putting their entire first book into the
7:13context while they're writing their
7:15second book. Or even if they're writing
7:18just the first book, they're they might
7:20have the entirety of everything that's
7:22been written up to that point available
7:25in the context as they're writing the
7:26next chapter. And I strongly advise you
7:28don't do this because what you're doing
7:30is you're watering down the prompt.
7:32You're making it so that if it would
7:33needs to access certain details, it's
7:35like a needle in a haystack. So, part of
7:38context engineering is also providing
7:41the AI with succinct and important
7:43summaries of the text that it needs to
7:45know.
7:46This is why I actually have an
7:47automation that I use called creating a
7:50summary plus document. One that scans
7:53through your entire book and summarizes
7:54each chapter one by one with just the
7:57key details from that scene like
8:00like it gives a one paragraph summary of
8:02the scene. It gives the list of
8:04characters that are in the scene and any
8:06descriptors of that character, what
8:08happens to them. And a few other details
8:11like that and it condenses the amount of
8:14text that is used
8:16or or
8:17condenses the entire book into something
8:20that's much more manageable for the AI
8:22to understand. It also signposts
8:23everything so the AI can quickly find
8:25what it's looking for.
8:26>> [gasps]
8:27>> Which is much better than just trying to
8:29dump the whole book in there and just
8:30hoping that it knows everything that's
8:32in the book. Which even though it's
8:34supposedly in its context window, that's
8:37not really how it works. The bigger the
8:39context, the more it has to sift
8:41through, the worse it gets at actually
8:43identifying relevant details that it
8:45needs for the scene. So this is why I
8:48tell people that you should try to
8:49minimize your context. Try to keep it
8:51under 50,000 tokens if you can because
8:55that's just going to result in a higher
8:57quality output because of that. Quick
8:59pause. If you're an author who wants to
9:01start writing more books, I put together
9:03my complete AI prompt pack that helped
9:05me and many others in my community write
9:07a book in a month. It's completely free.
9:09Hit the link in the description. Join
9:10the free waitlist for Story Hacker Gold.
9:12Then you just answer a couple quick
9:14questions and I'll send you the full
9:16prompt list plus a few extra free
9:18goodies. I'll also email you when I open
9:20the paid Story Hacker Gold group so you
9:22can join if it feels right. There's no
9:24commitment to join though and you'll
9:25still be able to get all of the freebies
9:27you signed up for either way. All right,
9:29back to the video. All right, the third
9:31major prompting technique is one I'm
9:32kind of excited about and kind of a new
9:34one, too. Uh this was first Future
9:37Fiction Academy came up with this idea
9:39that they called narrative physics. And
9:42I've expanded on it a little bit to
9:44where I think of it as sliders. Or you
9:47can think of it as game stats for
9:50different aspects of your world. Now,
9:52this is still just a prompting
9:53technique. There's really nothing crazy
9:55about it. But I'll give you an example
9:57of how this looks and in the future I'll
9:59probably make a whole video get it going
10:01deep on one or more of these techniques.
10:04But uh, here's the idea. If you have a
10:07character, let's say you have a
10:08character bio for a character that's in
10:10your book, and you include that
10:11character bio in every scene that that
10:13character shows up in. Well, what if
10:16that character changes across the course
10:19of the story? And what if their emotions
10:21are in the same in a different place so
10:23that they react differently? Like you
10:26might have a character who is normally a
10:29very nice person, but under
10:31massive amount of stress might break and
10:34like kill somebody. You know, like under
10:36the right circumstances. Feel like
10:37there's a joke there from like The Dark
10:39Knight and The Joker that if you just
10:41give people a little push, they'll do
10:43anything, right? That's this idea that
10:45people behave differently in different
10:47situations. And so what I came up with,
10:49and this
10:50the while the concept of these narrative
10:53physics is something that Feature
10:55Fiction Academy came up with,
10:57I have not seen any of their prompts. I
10:59have no idea how they do what they do. I
11:01came up with my own based on their
11:03concept. And the way I came up with it
11:05was this idea of sliders.
11:08Um, [snorts] so for every character I
11:10will now give them sliders for things
11:12like stress to uh, harmony.
11:16Uh, or or peace, right?
11:19And if they're like
11:21a number 10 on the harmony side, they
11:24might actually be like too harmonious
11:27and a little bit too out of it. Like
11:30maybe they're on drugs or something.
11:32Uh, but if they go all the way to the
11:33other side of the slider, like a -10 on
11:36the stress side, they're like so
11:37stressed they're having a panic attack.
11:39They're they're literally on the verge
11:41of dying.
11:42Um
11:43whereas we can pick somewhere in the
11:45middle is where their baseline is. And
11:48then for each scene, we can get the AI
11:51to identify, okay,
11:52where are they still at their baseline
11:54or have they moved up or down from that
11:56baseline? And what this does is is it
11:58allows the characters to shift and react
12:01more realistically to the scene.
12:05And so this is an excellent technique to
12:06get a little bit more variation about
12:08your character so they don't always
12:09sound uh monotone and the same all the
12:11time.
12:12It's also an excellent technique if you
12:14use something similar for, say you have
12:16a magic system,
12:18right? How do What are the rules of that
12:20magic system? How does it work? What are
12:21the costs? All of that. Making sure you
12:23you establish all of those things in
12:25almost a computer-like or game-like way
12:29with actual numbered stats and things
12:32like that that can then change and
12:34evolve over the course of the book. You
12:37can even do this with certain emotions
12:39that you want to get across inside of
12:42the of the story. Like the amount of
12:44tension that's in a scene. You know,
12:46what's that look like on a slider,
12:48right?
12:49Uh or the amount of uh
12:52spiciness in a scene. What's that look
12:54like on a slider, right? Sometimes
12:56you're going to have zero spiciness cuz
12:59the characters hating each other or
13:00something like that. Some sometimes
13:02you're going to have like in a romance
13:03you'll have it go up and down up and
13:05down and up and down but slowly getting
13:06bigger and bigger and then suddenly like
13:08bam, right? Uh you can take it all the
13:10way to a number 10, which would be like
13:13pornographic erotica, right? Or um so
13:17that's the kind of thing we're talking
13:19about here is different sliders,
13:21different ways of telling the AI this is
13:24the sort of tension that we're we're
13:25for. This is where the character's at in
13:27terms of their own sliders and so on.
13:29And so this is a really cool technique
13:31that I'm excited to be playing with more
13:33in the future. All right, the last
13:35technique that I really want to talk to
13:37about is, for lack of a better term, I'm
13:39just coloring calling it layering. All
13:42right? And this is the idea that we can
13:44actually utilize multiple strengths of
13:46multiple different LLMs in one batch.
13:51And rather than
13:53just choosing one LLM to give us
13:57everything that we want it to give us,
13:59we can take it in steps and use
14:02different LLMs for each step. This not
14:05only allows us to play to the strengths
14:07of the LLM, but it also allows us to
14:09kind of avoid some of the sameness
14:12that we get from any one LLM. For
14:15instance, I love the Claude models. I
14:17love Claude 4.6 Opus, which is the most
14:19recent one as I'm recording this.
14:22Uh but when I work with it too long, it
14:24starts to all sound uh a very similar
14:27kind of tone all the way through.
14:29This layering technique allows us to mix
14:32it up a little bit and make it a little
14:33bit better. So, what would this look
14:34like in practice? So, for instance, the
14:37Gemini and the ChatGPT models tend to be
14:39really good at uh logical thinking and
14:42structuring things. And so, maybe we
14:45could start out by having them look at
14:47the information we provided about the
14:49scene and say, "Okay, let's expand this
14:51and create a scene brief." And so, it
14:53creates the scene brief along with all
14:54of the little things that we needed to
14:56have.
14:57And then uh we know Grok is really good
15:00at dialogue. So, then maybe we could
15:01pass the scene brief over to Grok and be
15:03like, "Can you create a dialogue-only
15:05draft of this and just put in little
15:08placeholders with a little bit of
15:09blocking for what happens in between the
15:11dialogue?" And then from there, we pull
15:14it over to, let's say, Gemini. And uh
15:18Gemini is pretty good at prose overall,
15:20but uh not as good at other things. So,
15:23we could say
15:24write all the pros, you know, fill in
15:25all the pros for this, identify the
15:28chill scenes, and write all the pros for
15:31this stuff. And then we pass it over to
15:35Claude Opus 4.6 and be like, you're
15:37really good at writing really serious
15:40dramatic climactic scenes. Go through
15:43anything that's left in this scene and
15:46write in those little bits and
15:48and bringing all bring it all together
15:50so that now you have a chapter that's
15:52been gone over by like four or five
15:55different LLMs
15:56all providing what makes them unique to
15:58them and you end up with a chapter that
16:01is absolutely mind-blowing compared to
16:04what you would get if you just went to
16:06Claude or Gemini and asked it to write
16:08the whole chapter in one go. This is a
16:10really powerful technique that I think a
16:12lot of people aren't utilizing. It does
16:14admittedly increase your costs of being
16:17able to do this
16:19especially if you're using an API. It's
16:21going to take more especially if you're
16:23doing this in like an automation type of
16:25format. It's going to take a lot more
16:27money to do that. But if that isn't an
16:29issue for you, this could be an
16:30incredible way to get really really good
16:32results out of your LLM especially if
16:34you combine it with the context
16:36engineering and the checks and critiques
16:39and the narrative physics and the
16:41sliders. You combine all of these
16:43together, you're going to get some
16:44pretty fantastic results. So, these are
16:46some four pretty advanced tools that I
16:49think are really helpful for you. If you
16:51want to get access to a lot of my stuff
16:53including all of my automations that get
16:55into this sort of thing that I'm
16:57currently building as we speak.
16:59You can access the Story Hacker Gold
17:02group down below and if it's on the
17:04waiting list when you join, you'll
17:05actually get a whole bunch of my prompts
17:07all for free that you'll that'll just be
17:09sent to your email when you join the
17:10wait list and then you can you know,
17:12join the Story Hacker Gold group when it
17:14opens more publicly. So, check that out
17:17and I will see you in the next video.