Full transcript
0:00Today I have 11 foundational techniques
0:04that you can use to really inform your
0:06prompt engineering, especially when you
0:08are prompting for writing specifically.
0:11Now, if you don't know me, my name is
0:12Jason. I'm the nerdy novelist. And after
0:14writing 14 books the traditional way,
0:16I've since gone on and created this
0:18channel, helping authors realize their
0:20dreams using tools like AI, but in a way
0:22that doesn't compromise their ethics or
0:25their creativity. And in this case,
0:27prompt engineering is an incredibly
0:29important skill. It's a new skill, brand
0:31new skill that hasn't really been a
0:33thing before AI came here, but it is
0:36indeed a skill, and learning to master
0:38it is one of those things that will get
0:40you better results. So, tip number one
0:42out of 11 is probably the most important
0:45one here, and it's deceptively simple,
0:48but it is be specific. I cannot tell you
0:50the number of times I get people asking
0:52me like, I'm not getting this thing out
0:55of the AI. it's not doing what I want
0:56and I look and I can tell that they
0:59didn't ask it to do the thing that they
1:01want. You have to ask it right. This is
1:04one of those things where AI is not like
1:06a ghost writer. A ghost writer could
1:07maybe infer a few things and just
1:10through their own understanding of
1:12storytelling could just figure out what
1:14would make a good scene or good story
1:16based on your outline, right? But an AI
1:19can't do that unless what you want is
1:22specifically talked about in your
1:24prompting. And this is why people
1:26underestimate just how creative it of a
1:28process it is to write with AI because
1:31you have to be specific. You have to be
1:34thinking of all of these things. You
1:35have to know what happens in the scene,
1:37what happens uh in the book overall for
1:40your overall outline. You need to know
1:43the character arc that the character
1:44goes through. The AI doesn't just know
1:47this stuff. It can't just make it all
1:49up. You have to inform it. And so that's
1:51why being specific is my number one tip
1:53because if you're not specific, you're
1:55not going to get anything close to what
1:57you want. And this is why I will go to
1:59my grave saying this that it doesn't
2:01matter how good AI gets. AI plus a
2:04human, especially an AI plus a
2:06well-informed human, a really skilled
2:08human in the art of storytelling, will
2:11always beat AI by itself or AI paired
2:15with a human that doesn't know what
2:16they're doing. I don't care how good the
2:18AI gets. An AI with a skilled human will
2:21always win out. So that brings me to tip
2:24number two, and that is to remind the AI
2:27occasionally about what you want. We
2:29know that every AI model has a context
2:32window. It might be anywhere from like
2:3510,000 tokens to 400,000 tokens. It
2:38could be a wide range. But what people
2:40often don't realize is that even if they
2:43have such a large token window, it can
2:45still not forget but sort of misplace
2:49the information that you are asking it
2:52to do. I tell people, and I'll get to
2:53this in a different tip, but I tell
2:55people like don't water down your
2:56prompts by making them too big because
2:58even if technically an AI can handle
3:00that much information, sometimes it
3:03things get lost. If you imagine that
3:04your instructions are just like 1% of
3:07your overall prompt, you can't
3:09necessarily expect for that 1% to be
3:12remembered more and weighted in a way
3:14that is more important than the rest of
3:16your prompt. So, I'm not just talking
3:17about being in the chatbot and reminding
3:20the AI about what its tasks are as you
3:23go along because it's forgetting the
3:25stuff that you used to talk about.
3:27That's not what I'm talking about here.
3:28I'm talking about making sure that the
3:30AI clearly understands the role that it
3:33is in. So that can be putting things
3:35inside of the system prompt so it always
3:37has those things top of mind. It can be
3:39repeating the task in multiple ways. I
3:42do this with style a lot where I will
3:45ask it to do something and I'll repeat
3:47the same instruction but like three
3:50times and phrase differently each time
3:53because that way I'm really ensuring
3:54that the AI understands what it is I'm
3:56trying to get across. Number three, and
3:58this kind of goes along with the last
3:59one, is that less is usually more. What
4:02I see happen a lot is people put their
4:04entire book into their prompt. And first
4:07of all, you're going to significantly
4:09increase your cost for that. But the
4:12problem with putting an entire book or
4:14entire portions of your book inside of
4:16your prompt is that you're watering down
4:19the prompt. You're like creating a
4:21larger hay stack to find the needle in.
4:23What is often more effective in this
4:25specific use case is to create a summary
4:28of your book. So if you have a whole
4:32book, let's say it's 75,000 words, and
4:35you want the AI to understand what
4:37happened in that book because maybe
4:38you're writing the sequel or something
4:39like that, a better way to do it, more
4:41succinct and easier for the AI to
4:43understand is to create summaries of
4:44each chapter of that book. So you have
4:47essentially an outline of what happened
4:48in that book and then you include that
4:50outline, which will be smaller. It'll
4:52still be a lot of words, but it'll be
4:54significantly smaller than having the
4:57whole book in there. And when you do it
4:59that way, the AI will understand what
5:02you're looking for and what in the
5:04context of what happened in the previous
5:05book much better than if you just
5:07included the whole book. Plus, you'll
5:08save a little bit of money because
5:10putting a whole book inside of a prompt
5:11is expensive. This is one of the reasons
5:13why I'm a fan of Novelcfter because
5:15Novelcfter is a software that
5:17essentially does this for you where it
5:19will automatically filter out the stuff
5:21you don't need for the scene you're
5:23writing and it will bring in the stuff
5:25you do need. So, it's really good at
5:26stuff like that. Tip number four is to
5:28use XML tags in your larger prompts.
5:31Even though everything I said about less
5:32is more in the last tip, I can typically
5:35still go up to maybe 20, 15, 20,000
5:38words in a prompt and still be okay. But
5:41when you're getting into prompts that
5:43big, it really helps to break things
5:45down into little boxes that the AI could
5:47easily understand. And the best way to
5:48do that is with XML tags. This is
5:50actually a tip recommended by most of
5:53the actual developers of AI because most
5:56of these models were trained on a lot of
5:58web data and the web data is coded in
6:01HTML which uses XML tags and that's how
6:03it's trained. So it really understands
6:05this concept really well. What an XML
6:08tag looks like is you get a word with
6:10two brackets and then you have some
6:12content and then beneath that content
6:14you have the same word with another two
6:15brackets but there's like there's a
6:17forward slash and after the first
6:18bracket and this is just a clever little
6:21way to kind of easily divide the
6:24information into little containers that
6:26the AI can understand. So you could have
6:28one for instructions. So you have an XML
6:30tag, an opening tag, and then you put
6:32your instructions there. Then you put a
6:33closing instructions tag after that. You
6:36can have one for style where you have
6:37like, okay, here's your opening style
6:39tag. Then you put all of your style
6:41information and closing style tag and
6:43there you go. So, it's a really neat way
6:44to keep things organized by the AI and
6:47helping it understand your story better.
6:49Tip number five is to experiment with
6:52adjusting parameters. Every model,
6:54especially if you're going through the
6:55API or through a service like Open
6:57Router, you can have access to different
7:00parameters like temperature. There's
7:02also top P. There's max tokens. There's
7:05a whole bunch of different parameters
7:06that you can fiddle with. You can't, by
7:08the way, fiddle with these inside of
7:10chat bots where it's just not allowed.
7:12And a lot of these parameters are
7:15actually really good at helping you find
7:17the results that you want. Sometimes you
7:18might want something that's a little bit
7:20more creative, a little bit more not
7:22sticking to the book. It goes thinks out
7:25outside of the box a little bit. In that
7:26case, you might want to raise the
7:27temperature a little bit. Or maybe you
7:29want something that's that doesn't think
7:31outside of the box and sticks very
7:32closely to exactly what you ask it to
7:34do, in which case you might want to
7:36lower the temperature a little bit.
7:38Things like that can really make a
7:39difference and most people do not play
7:41with them. And then you really should if
7:43you're testing out what works and what
7:45doesn't. Tip number six is to experiment
7:47with different LLMs. So every large
7:49language model is different. Everyone
7:51has strengths and weaknesses. People ask
7:53me all the time, what's your favorite
7:54large language model? And while I do
7:56typically go to the Claude family as my
7:58overall favorites, I use different large
8:01language models all the time because
8:03there are different use cases where you
8:05might need some than others. Right now,
8:07I'm really enjoying GPT5 for
8:09brainstorming and anything that requires
8:11just a hint more of creativity. But I
8:13still enjoy the Cloud models for writing
8:15and then I also like some of the Gemini
8:17models for editing. Everything is a
8:19little bit different. They have
8:20strengths and weaknesses, and you can
8:21learn some of that by following people
8:22like myself, but for the most part, you
8:25may want to spend a little time
8:26experimenting with different LLMs. You
8:28should also think about experimenting
8:31with different prompts for different
8:32LLMs, cuz using the exact same prompt
8:34might work for one, but won't work for
8:36another. And if you tweak the prompt,
8:38then it would suddenly start working
8:39better for that model. And this is just
8:42one of those soft skills in the AI
8:44writing space that I can't necessarily
8:46give you any hard and fast rules on how
8:48this works. It's just something that
8:50comes with experience. You start to get
8:52the vibe of what each model does and
8:55what it's good at and how you might be
8:57able to improve on it by changing your
9:00prompt. Because sometimes I'll get
9:02something out of an LLM and I'll be like
9:03that's not really what I wanted, but
9:05I'll bet I if I changed the prompt this
9:07way, it would make it better. And it
9:09does. And that's the kind of thing that
9:10I can't really teach. I It's just
9:13something you have to get used to. And
9:14playing around with these large language
9:16models, eventually you will get there.
9:17Most of the people inside of my
9:18community, link down below, by the way,
9:20have done this and are getting pretty
9:22good at it. Tip number seven goes along
9:24with what I just said, which is to test,
9:26test, test. Prompt engineering is not an
9:30exact science and it requires lots of
9:33iteration and improvement and testing.
9:36And I often get people coming to me and
9:38saying like, "This doesn't work or that
9:39model didn't work for me." And I can
9:41instantly tell that they have not tested
9:43enough because so often, and this is
9:45especially true of a lot of the AI
9:47haters out there that say they can
9:49recognize AI slop immediately. I
9:51guarantee you that they can't. They're
9:53just looking at the stuff that is not
9:55well prompted. All you have to do is put
9:58a little bit more effort into your
9:59prompting and in testing it and
10:01improving it and you can get something
10:03that doesn't sound anything like the
10:05so-called AI slap that those people are
10:06looking at. So, always be testing and
10:09whether you're testing your prompts,
10:10testing the LLMs, uh testing with
10:12different parameters like this. This is
10:14probably one of the most important tips
10:16that I have for you. Tip number eight is
10:18to break your larger tasks into smaller
10:21steps. I mentioned putting an entire
10:23book inside of your prompt and uh that's
10:26something that's not very efficient with
10:28your prompting because you're trying to
10:29do too much. So, let's say you want to
10:31create a summary of that book. Well, you
10:33can take that entire book, put it into a
10:35large language model, and say,
10:36"Summarize this book chapter by
10:37chapter." But I found that almost every
10:39time that I do this, even with some of
10:41the bigger large language models, it
10:43can't do it. The reason is, I mean,
10:45it'll give me something that roughly
10:48approximates a summary of the book, but
10:49it might miss chapters. It might get the
10:52order wrong. It might number the
10:54chapters incorrectly. Like, it gets a
10:57lot wrong. And that's because you're
10:58asking it to do too much with a enormous
11:01watered down prompt. Instead, a better
11:03way to go is to take one chapter at a
11:05time and say, "Take this chapter,
11:07summarize it, give me character
11:08information, give me world building
11:09information, you know, have it sort of
11:11like list out all of this stuff." And
11:12you do that for every chapter. And once
11:14you have that uh condensed version of
11:17information, it is much easier for the
11:19AI to say, "Okay, now create a summary
11:21of this character and their character
11:23arc in the story. Create a summary of
11:25this world building element. create a
11:27timeline of events in the story because
11:30you have broken it down into smaller
11:32tasks and now it's capable of doing a
11:34lot much more efficiently than if you
11:36were to just go straight to the entire
11:39book and say make me a story bible out
11:41of this. It just doesn't work. Tip
11:42number nine goes back to a framework
11:44that I created and when I initially
11:46created this it was kind of like
11:47innovative and on the front lines of AI.
11:50These days it's a little bit more
11:51simplistic but it is absolutely an
11:53important principle to understand as you
11:56are going about and creating your
11:58prompts and that is the fits formula fit
12:01ts. This stands for framework identity
12:04task and style. The framework is pretty
12:06straightforward. Every AI thrives on
12:09structure. If you can provide some kind
12:11of structure even if it's just make me a
12:13list of titles that have a blank of
12:16blank and blank format, you know,
12:18something simple like that. just giving
12:20it that little bit of structure actually
12:22increases its creativity. And I even
12:24created like I I wrote a whole book
12:25called the plot module which is
12:27essentially a giant structure that I
12:29designed specifically to outline stories
12:32with AI even though I think when I wrote
12:34the book I only have one chapter about
12:35AI in that book. It's mostly a tool for
12:38authors but I designed it in such a way
12:40that it works really well with AI
12:42because it is so rigidly structured. So
12:45that is the kind of thing that you want
12:46out of the framework. The I stands for
12:48identity. It's usually a good rule of
12:50thumb to give the AI an identity like
12:53you are an expert editor, you are a
12:55best-selling author, stuff like that
12:57tends to narrow down its field of focus
12:59because remember these models are
13:00general models. They can do lots of
13:02things besides writing. So we want to
13:04just narrow it down to the specific
13:07field of play where we want it to be.
13:09And uh doing that with the identity is
13:11how you do that. The T stands for tasks
13:13and that's the most straightforward of
13:14the bunch. It's just here's what I want
13:16you to do. And then the S stands for
13:18style. You're not always going to need
13:19style for everything, but often when
13:22you're writing, especially with AI, you
13:24want it to write it in a specific way.
13:26And the instructions on how to write in
13:28a specific way is what your style looks
13:31like. It can be everything from like
13:33avoid show don't tell or rather do show
13:36don't tell to actually giving it like
13:385,000 words of your own writing to act
13:40as a model for it to imitate that sort
13:43of thing. So that's the fits framework.
13:44Another framework I developed is called
13:46the fractal technique, which I use quite
13:48consistently myself and is kind of the
13:50basis for most of the prompting and how
13:52that all works from a global level. The
13:54fractal technique means that you start
13:56out small and get bigger over time. So,
13:58I usually start with some brainstorming
14:00and then I'll take all of those
14:02brainstorming elements and format it
14:04into a synopsis. And then from the
14:06synopsis, I'll build uh into an outline.
14:09And then from the outline, I'll take an
14:10individual scene and create a scene
14:12brief from that. And then from the seam
14:14brief, I'll actually take that and write
14:16the whole thing. And that's basically
14:18how the fractal technique works. You
14:19just start small and you slowly work
14:21your way up bigger and bigger and bigger
14:23until you've got the whole thing. This
14:24kind of goes along with that idea of
14:26breaking things down into smaller tasks.
14:28All right. And last but not least, tip
14:30number 11 is to spend time actually
14:32interacting with the AI. Don't just
14:34expect your prompts to be the beall and
14:36endall of writing with AI because
14:39sometimes you get the best results after
14:40going back and forth a bit with the AI.
14:42This is especially true if you're using
14:44a chatbot where it's easier to kind of
14:46go back and forth in that way where you
14:48might be spending more time in the
14:51brainstorming and outlining stages
14:53inside of the chatbot. I admittedly
14:55don't use this technique as much when
14:57I'm actually writing the story. But when
14:59I'm outlining or brainstorming, I could
15:01be like, "Okay, yeah, I like that idea,
15:03but can we add a little bit more of this
15:04in there? And uh maybe can you rewrite
15:07it with this and this and this?" You can
15:10ask it questions. You can ask for
15:11follow-ups. You can ask it to rewrite
15:13with uh different feedback and you get
15:15this feedback loop that can really by
15:17the time you're done with it create
15:19something that's entirely unique. If
15:21you're just giving it one prompt, what
15:23happens is the AI looks at that one
15:25prompt and it creates a probability
15:27model of what would be the most likely
15:29thing to say given that prompt. And
15:31that's fine, but it can be very generic
15:35and have problems with it. And you might
15:37also get the same results again if you
15:38tried the same thing. But if you're
15:40going back and forth like this, you're
15:42adding quite a bit more of your own
15:44humanness and your own creativity in
15:46there that just makes it unique in a way
15:49that you're not going to really come
15:50across those results again unless you
15:52walked through the exact same process
15:54blowby-blow. And even then, it's
15:56unlikely that you'd get quite the same
15:58results. And so you get a much more
16:00organic and a much more interesting
16:01output when you do it this way. Those
16:04are my 11 basic prompt engineering
16:06techniques. If this was helpful for you,
16:07I'd invite you to check out our one
16:09membership down below. And there's a
16:10silver membership. If you want to check
16:12that out, it's a small onetime fee that
16:14looks good for you. You can uh move up
16:16to the gold membership, which right now
16:18I'm designing to be the single fastest
16:20way to get to 20 bucks for every author
16:22who goes through the program. So, if
16:24that sounds interesting to you, go ahead
16:25and check that out down below, and I'll
16:27see you