Introduction

I am going to talk a bit about writing with AI. I'm gonna step down the technical level a bit and talk about something that doesn't really need any tools at all other than like maybe ChatGPT or Gemini or on topic or whatever you want. And I will do that in such a way that by the end we'll cover three things.

Why generic prompting falls short

The first one is I will walk you through the framework that I currently use to write stuff with AI, which is a little different from what I've seen online in that it's not this typical of like, here's this magic prompt, plug it in, it spits out 100 LinkedIn posts in two seconds, off you go. If you've tried them, you know they don't work, and I've tried a lot of them because I really want this to work, and so far it hasn't worked yet. So this approach is slightly different.

It requires you to put in more effort, which is actually a good thing because it means that you add more context to the AI to give it stuff to work with and do useful things, which is one of the things that's missing if you just have a prompt and say, well, write me a LinkedIn post, which I can show you in a second.

The second thing is we will do some live writing if time permits. I will try to write a LinkedIn post with exactly that method that I'm going to show you so that you hopefully see how does it actually work when you do it in practice.

The third thing is you go home and hopefully know how to write stuff with AI. if you pay attention. If you're not, results may vary, right? That's how this goes.

Agenda for the session

All right.

Demonstrating the limits of one-shot prompts

Before I jump in, I want to show you why just prompting doesn't really work.

Is this big enough? Probably not. I can zoom in.

Cheap to generate, costly in impact

So if I just do the stupid thing of asking, in this case Gemini, write a LinkedIn post about me attending the MindStone AI meetup today, it will spit out this, which is, if you want to read it, it will sound very generic because it's just, ah, it's left with so much to think about. But if you've seen this kind of stuff, it gets very repetitive very quickly. And you'll see a lot of it because it's very cheap to generate.

So from a cost perspective, this is amazing. But from an actual impact perspective, unlike me posting this on LinkedIn and actually getting engagement out of it, it will not perform very well.

It was very practical and insightful. I made a lot of cool new connections. I can't wait to do it again. It's super amazing.

Context memories can mislead

And you see the same if you try it with, say, ChatGPT over here. it spits out a slightly different answer.

It magically knows that the MindStone AI Meetup is in the Microsoft Nerd Center. The reason for that is what you've heard about earlier, this memory component that ChatGPT now has. And in there, I've mentioned that the Microsoft Nerd Center is where the event happens, and it remembered that.

I've run through the talk that I want to give before with ChatGPT to see if it actually spits out something useful. So now when I ask it again, it can put it in there, right?

So that's a... first sort of hint that context is one of the pieces that's missing in good writing, right?

Evidence AI writing can work

At the same time, I know that it must work that you can write with AI which is because of this blog post. This is from a guy called Mark Lawrence. Does anyone know Mike Lawrence?

He's an author. No, so he is a fiction writer and he won a bunch of prizes. He published a bunch of books.

So in the area of like writing and sort of fiction, yeah, in the area of fiction writing, he's a relatively big name. And he did this fun thing when I came out, he was like, okay, I know a bunch of other very famous established professional writers.

Human vs AI short-story experiment

I'll give them a prompt, let them write a piece, and then I let AI write a piece. And we take four authors, we take four articles from AI, and then we just show that to people. And that's what this blog post does.

It goes through and says, here is a short story of a couple words. Is this story any good? Was this story written by AI? Please vote. And then he just showed this to his audience just to see what happens, right?

And to make a long story short, this was the result.

The green circles are stories written by the human professional writers. The red circles are stories written by AI. And the outer circle indicates what a large group of people thinks about the post.

So this one, green on green, means it's written by a human. Audience thinks it's written by a human.

Red with a green circle around means written by AI. People think it's a human.

Red on red, AI, AI, and green red means written by a human, but people think it's an AI.

Key takeaways from the study

So there's two takeaways from this. The first one is, this is pretty close. We're not that far off, so it must work. There must be some way that we can use AI to actually write useful stuff that people like.

The other one is, which is in the post itself, it's not so much in this particular chart, If you do this on aggregate, you get this result.

If you ask an individual human on each individual item, they're not much better than random. They're basically guessing. On average, our guess is better than random, so we get it right collectively, but individually, we're really poor at this.

So I know it must work, and also this blog post was from 2003. So we know this works for years, but how to make it work, I didn't know.

So I've decided, even though my initial attempts, which I showed you, basically of me just throwing it out there, LLMs are amazing, let's see if they are amazing, didn't work, I know it works, so I just decided to keep bashing my head against the wall until I figure out how it works, because it must work, right? And if it does, that would be great.

How to make AI writing work

So, continuing with this, how to make it work.

Two missing pieces: process and context

The trick is, there's two pieces that are missing from just plain prompting the LLM to get a reasonable output.

Lesson from programming workflows

The first one or the first insight comes actually from programming. Does anyone here like program or vibe code or something like that? About a third, a quarter, yeah. of those of you that try it, you probably share this experience, right?

Like you go to Lovable or something and say, write me a tic-tac-toe game. And it goes, it writes a bunch of stuff, and it doesn't work, right? If you try it slightly differently and you look into like, how do people vibe code?

It's like, oh yeah, first you need to write a PRD, like a product requirement document that says, this is what I want to do. These are the features I want to have. This is how the user experience looks like, yada, yada, yada. Then you go and maybe make some pictures or some wire drawings of like, this is how my app should look like.

And then you take all of that, feed it into the LLM, and then it suddenly starts to do more reasonable stuff. The thing that we're doing here is we're going through a process of first write this document, then go to AI with it, then it spits out something.

Okay, do this thing next, it spits out something. So there's some kind of structure to the process of how you work with AI. It's not just one-shot prompting of we write a prompt, magic happens, it's the result.

Why context matters in writing

So that's the first piece. The second piece turns out is context.

And that one I got from, not from Mike Lawrence, but from looking at how do novelists actually deal with the fact that AI is a thing now. Does anyone actually try to write books with AI? Because if I can't write the LinkedIn post, maybe if people figure out how to write really long form content, It should work for LinkedIn posts and all the other content as well, right?

And the thing there that that community realized is that the reason why AI writing sounds hollow is because it doesn't have enough context to work with. So the best example I can, or the best explanation for that that I can give is if you ask AI to write something, it will A, follow your instructions, and B, try to not hallucinate. That's what we train AI to do, which means it's really trying to not be too creative.

But if you only give it one or two bullets and then ask it, please write me a five-page document, the only way it can do that is by stretching the content as much as humanly possible because it tries to satisfy two constraints. Write a lot of stuff. Don't invent anything. So that's how you end up with this very empty-sounding stuff because it doesn't talk about anything because you didn't give it anything to talk about.

And if it does... if it finds itself really unable to make those two ends meet, it will invent stuff, but the stuff that it invents will be the most average thing that it can possibly say because that's how AI predicts, right? It predicts the average most likely next word, so you get the most likely thing to talk about.

So that's the two pieces that I'm missing. A, some framework to help us interact with AI constructively, and B, a way to provide a lot of context so that the AI has more than enough things to work with to actually write something good.

Enter the snowflake method

And that turns out what works really well here is something called the snowflake method, which again comes from creative writing. And if we take that snowflake method and combine it with the LLM assistance, the pieces where AI is really good at, we actually get something that works surprisingly fast and surprisingly good. Like all the people that I've shown AI-generated stuff to were like, oh, this looks really good.

Did you write this? I was like, no, I wrote this.

Like the abstract from my talk, which I think only you saw, not everyone else. That one was written by AI, for example. Right.

The snowflake method explained

So what is the snowflake method?

From sentence to story, step by step

The idea is that writing a full chapter in the novel is really complicated if you just do it from scratch. There are some people that can do this, discovery writers. They just have an idea of like, wouldn't it be fun to talk about this thing? I have this character. Let's see where this goes. And they can do this because, I don't know, they're very skilled, but others find it really hard and they just stare at a blank page and like, where do I even start? And that's me.

So one technique you can use there, which works nicely, is you start just with a single sentence of like, OK, this is something I want to write about. And then rather than trying to write the full text in one shot, you expand it a little bit. You write a paragraph. Then you write a one-page summary, like a short story. Then you take that short story, take pieces of it, and say, okay, I can make it even longer.

So we start small. We iteratively build outwards. That's, in a nutshell, the snowflake method.

What LLMs are good at in this process

And the other part is what are LLMs good at that might be useful when we try to run through this process? Well, first of all, they're really good at bootstrapping. If I give an LLM a couple of bullet points, it will be able to generate a couple paragraphs out of it or one paragraph out of it. So rather than me trying to find the initial words, I can start off with something that's written and then start editing.

AI is also really good at brainstorming if you use it right. The wrong way to use it is write me 10 ideas for LinkedIn. That's not very useful because what you get is the average next word, which means you get average posts that everybody else gets. And everybody else that asks the same prompt will get the same thing, so you all talk about the same stuff, which doesn't really differentiate you.

A better way to use it is to Let it give you ideas, and then you judge if that idea is any good. Or maybe you read something that the AI says, and you're like, oh, well, that actually reminds me of something else. That something else is a really good idea.

So we can then use the AI to cycle through variations of stuff. For example, I want a title for my post, or I want the way to write a sentence. I ask it for three, four variations of that sentence. I read through those, pick the pieces that I like, put them into my normal text editor, write something that I like, take the thing I like, put it back into . It's not perfect.

I don't like this part. Give me five variations. And then I go back and forth and ping pong until it looks nice.

And the third one is AI is quite good at critiquing your stuff in that it will find obvious holes that you are forgetting. So if you want sort of a silly little thing, if you have an idea or if you need to prepare a presentation or if you just need to, I don't know, write a document for work or something, you just ask it, pretend to be the person that I'm going to give this document to. What am I missing? and it will spit out generic things, but those generic things are likely stuff that are not in the document. And you probably should have that because that's baseline, right?

Turning method into a practical workflow

So that's the idea, how these fit together. And that's the two methods. How do we stick them together into a process?

Step 1: Premise and ending

And we start doing basically this snowflake idea.

We start with a premise, which is we choose the idea we want to write about, and we describe it in a sentence or two, just so that we could go to someone else and say, I want to write about this. And the other person goes, yeah, sounds about right. Sounds interesting. That will give the AI context to understand what you're writing about.

The other thing is, we choose the ending first. And that's for two reasons.

The first one is, especially if you're writing nonfiction, usually you write because you want people to do something. Like you want them to share your post, you want them to click on a button and sign up for a course, you want them to, I don't know, just learn something in the process. And if you're clear about what it is you want people to do when they read your thing, then AI will benefit from you telling it, because then you both work towards the same goal.

Because ultimately, AI are what I like to think of as interpolation engines. You give it two points, it can draw a line through that. But once you go outside of the data that the AI knows, it can draw the line in arbitrary many directions and will just choose one at random. And that might not be the thing you actually want.

So those are the two things we get clear on in advance.

Step 2: Choose platform, format, and headings

Then we pick an outline for the thing we're writing. Three steps there.

First one, we pick the platform we want to publish, a blog post, Twitter, a book, a white paper, a research paper, whatever. And then from that platform, there's different formats that different platforms expect. Those of you who are active on social media know that some posts work better than others.

People like me who are in research know that certain journals have a specific format they want you to stick to, and if you don't stick to that, you're doomed. So we choose the format based on where we want to publish the thing and also based on what fits best. And from that, that format is basically like a generic blueprint that says like first you write an introduction, then you write a section about related work and related work in your field, then you lay out your methodology. Very high level stuff like that.

From that, we can then take the format plus the general idea that we have and start choosing headings for each section. And the reason we do that is we're basically now doing the next level of the snowflake method. We're writing like a very small abstract of what it is we're going to talk about.

So we have the idea. Now, with the headings that we choose for the piece we're writing, there should be enough content there to see that this is sort of the abstract of what I'm talking about. So if a reader then ends up on your, say, blog post, and they are not really sure if it's worth spending half an hour to read through everything, they will just skim read and look at the headlines, right?

The goal there is, A, you want to make sure that you prove to the reader that they came to the right place and you actually deliver on whatever hook you had when they clicked on it. But also, you want them to learn something useful. And that you do by just looking at your headlines together and saying, is this in itself coherent? Does it break down the thing I want to talk about nicely? And does it deliver value?

So that's what we do here. We can bootstrap those headings from the context I'll show you. We can do the brainstorming and critiquing that I talked about earlier.

Step 3: Flesh out main points

Once we have the headings, we go to the main points. And this is the part where we sort of combat the slop. We have the points we want to talk about, or we have the idea we want to talk about, maybe broken down into individual pieces that we now need to flesh out more. Here we come up with the actual content and the value that we want to deliver.

And this is something that is very much on your side in that if you ask AI to generate this, you will end up with generic stuff. But your ideas will make the content worthwhile. You don't need to formulate it out in detail and say, oh, this is the exact way I wish to write it. It's more like bullet points under each heading.

We should probably cover, I don't know, the venue that we're at at this meetup. We should cover maybe some people that we've met, maybe some lessons that we've learned.

And then it's not just cover the lessons, but I learned this, I learned this, I learned this. Doesn't have to be fleshed out yet, but we just need to write down that this is something we want to talk about.

And this is what A, makes the piece valuable, because this is the meat of your content. And B, it gives the AI context to write meaningful things rather than make up stuff or try to stretch it. And the more bullets you write, the better.

And then again, we can do the thing where we use AI to critique what we have. If we say, oh, this piece of thing that I could talk about seems really useful. Is it actually, or is it fluff? Is it too much information? Is it too detailed? We can ask AI. AI is not good at generating stuff for us, but it's very good at telling us if the things that we have are actually any good.

Step 4: Draft, iterate, and edit

And then finally, that's when we actually start writing the prose. We now have the bullets we want to talk about. the structure we want to write it in, and the headings for each structure.

We can take all of that over and say, AI, write me a first draft. That first draft will, again, not be perfect, but it will be much, much better than me going, write me a LinkedIn post about the event that I just attended, because it just has much more information, because we just put in all that effort.

And then we start iterating. So we go through the text that we actually have, paragraph by paragraph.

We read it. We don't like something. We edit it. We like it. We move on.

And then we have a piece, and then we're done. Sounds easy, right?

Live demo: Building a LinkedIn event post

All right, I have a demo now.

Did I lose anyone? Did I run through this too quickly? Any questions so far?

No, then I will show you the demo. All right, we will do exactly that thing.

Premise and ending for the demo post

Let's write a post about me attending this meetup.

So the premise would be something like Let me see. What would I talk about? Okay.

If I get too far off, then I have that entire thing already prepared elsewhere. Something like this.

Ending. Let me see. What would be a good ending?

It was a really fun event.

We would like people to actually join the community because it's fun and we should make more of it. We should have more people here, right? If you haven't yet, go check out the community.

Sourcing outline blueprints

The next piece that we need is an outline blueprint. Where do you get them? Well, there's two ways.

Either you go to YouTube or Google and ask what are typical formats that work well on platform for type of thing I want to write about, and you will find lots of creators, marketing experts, agencies online that would be like, here's some free content that shows you how to write this, because the current theme is we give away the frameworks, but we sell the framework, but not the implementation. No, we give away the framework, we sell the implementation. That's the way you turn it, right?

So a lot of people tell you how to do this. But if you want them to do it for you, you'll need to pay. So this is something you can find online.

Or you bootstrap it with AI. You ask it, like, I'm writing a post about this thing. What's a typical outline? And that's actually a good question to ask AI about, because I'm asking it, what is the common usual thing? And if I ask that, I'll get the common usual answer, which is exactly what I'm expecting here. So it's a good start.

It's not perfect. but it's a good start.

The other thing is, as you publish more and more on a specific topic in a specific format, you accumulate data points on what actually works, and then you can adjust this, right? So if you have previously published a post of this style, you can just reuse the outline blueprint from last time. So getting this is fairly straightforward, right?

Generating section headings with AI

Right, so now we want the outline.

So let's go to Clob and ask it, I am writing a Actually, I will do one thing. I have this here, and I had a nicer way of phrasing this prompt.

So let's use that because... So let's change this to a LinkedIn post about me speaking at the MindStone Boston September AI Meetup at the Microsoft Nerd Center about AI co-writing to speed up the writing process while keeping the human in charge. Slightly more context here, so hopefully we'll get better output.

I am writing a LinkedIn post about me attending a in-person meetup.

These premise ending and And then what we do is so I really the instructions are very straightforward It's like here are the premise the ending and the outline blueprint that we have earlier Please generate headings and now we need to paste those things in which is this piece Then we have the ending And then we have the outline blueprint and That's this one. There's my tab.

Is this big enough, by the way? Make it bigger. I can try to zoom in a bit.

Also, tell us all about how frontier models love Markdown and why you use it. That is fairly straightforward. You want some amount of structure in your prompt for the AI to find which section you're talking about. There are different strategies for that.

Claude usually recommends using something called XML, which is how websites work. It essentially would look like this. Where I then tell the model this section here, is the premise. It starts here, it ends here, and everything in between is the premise. If I talk about the premise up here, the AI model can figure out, OK, this is the thing I'm talking about.

Personally, I don't like using XML for that because it's actually really expensive. So Markdown works as well. Markdown is, if you don't know what it is, it's a language to, or it's a way of embedding styling into the writing using simple characters rather than sort of a what you see is what you get editor like Google Docs or something like that.

So for example, these two stars here make text. Oh, you need to zoom in on this. I'm sorry. There we go. These two stars here make text bold.

So the bold starts here, ends here. And if I have a thing that can interpret Markdown, like, for example, this cloud editor, it will be bold. Hooray.

And these two hashtags that you have there, those are different levels of headings. So one hashtag is H1. Two hashtags is a H2 headline. So it's essentially the same as me doing this. This is like heading one, heading two, heading three. This would be like one hashtag, two hashtags, three hashtags as a prefix. That's Markdown in a nutshell.

What I'm telling it here is we have the premise, the ending, the outline. Please use these three things to spit out titles. Here's the premise, here's the ending, here's the outline. If I now hit enter, then hopefully good stuff happens.

I need to format this a bit. I will take these, go back to my editor. and just paste them here. Let's try this.

This is a bit large because I will now very briefly format it like this because then we can actually read what it says.

How am I doing with time? How much do I need to skip over the details of doing this? OK, so I need to wrap it up now.

I have 15 left. Oh, wow. I thought I can talk 15 minutes. Online it said I have a 15 minute presentation.

So now I'm like, ch-ch-ch. All right. Then we actually have time to go over this.

So there's a hook missing here. So the hook that it spat out was opening statement about complements. OK, this is not good. This is the curse of demos.

There we go. Can you prefix it with the titles of each? Let's try this instead. I will reset this, and I will do a new chat. There we go.

Now we have bullets, which is great because bullets mean we get a structured input. One, two, three, four, there it is. Let's try this again. All right. This is not working as I wanted it to work, but that is okay.

All right. Never mind. I thought we'll try this. It didn't work, but that's okay. Let's go back to this.

Just dropped some AI co-writing knowledge at Microsoft. This is not a hook that I would want to use, but However it formats it like this. Let's start with this. Speaking at the Mindstorm Boston meetup. Now these are not the actual titles that we will use in the end, but they are a good enough start. And then I can show you how this works to iterate.

Amazing to connect with name from company. Okay, which one did I miss? I think I missed this one. There we go. Now it makes sense again. Don't miss the next Mindstorm Meetup. Right.

Refining hooks and headlines

So for the hook, I already kind of know what type of hook I want to use here, so there's no real use of asking. I can just say I want to say something like, guess where I was last night. And the reason why that hook seems more interesting to me, sorry, I'm a bit zoomed in here. I'm not used to it being so zoomed in, so it looks a bit creative on my machine. The reason why I want this type of hook is because, if you remember in the beginning, I took this group picture. If I post this picture and say, guess where I was last night, I think that might hopefully trigger some curiosity and get people to click on the article. I know this, so I don't need AI to generate it.

Three game-changing insights on AI co-writing. I don't really like this one. So I go back to AI and say, the headline for the, what is the piece called?

one second I need to get rid of all of this here and all of this doesn't have to be heading it can just be normal text so that we can actually see things again wonderful there's the venue missing here and I think that was the piece that this one There we go. Right. Speaking at the AI MindStone meetup. This is good enough because all we do here is we want to mention where we are and what the meetup event is about. This one is not so good. So the headline for the key takeaways are a bit generic. Can you give me five other variations? Here is the current draft. Take away.

I'm sorry. I'm not typing very fast. All right. Same story here. Here's the draft. Give me five options.

Now it's starting to work again. Now we have other graphs. It's like what I learned about keeping humans in control of writing AI, the surprising truth about AI co-writing, speed versus quality. Now, these are again a bit generic, but something here might look good, right?

Like I can maybe look at things like what I learned sounds already much better than key takeaways to me right now, spontaneously. So let's keep that and keep that around for a bit. What else is here that is good? Why most people get AI writing wrong. That sounds nicer.

I should add that I'm not very good at playing the social media game because I started quite late. So if you are a marketer and see something that I don't see or you just know how to do social media, feel free to point it out to me. Anyway, so let's say we like the first piece. I like the what I learned bit. This sounds a bit funny.

Can you give me 10 more variations starting with that? And then this way we can just cycle through different variations and now it says, okay, what I learned about AI co-writing, making AI writing work about something. And then this way I can cycle through this again and again and again until I find something that I like. words, pieces, phrases, I piece together the entire title. And once I have that, I can move on. I will do a cheat code because I've done this before up here. And then if you go through that entire thing, you can get something like this. I'm mainly doing this for the sake of time.

So then you can get something like, guess where I was last night, speaking at the MindStone event in Boston, three key insights from my talk, networking gold, your turn. The reason why I have, for example, networking gold here is because on LinkedIn, you don't want a full long title because you're not writing an actual article. It's like bullet points and short sentences. So I want those titles to be short. Yes, so let's actually steal these guys so that we can speed up that process a tiny little bit.

But does that make sense on how to get sort of titles or general topics to talk about within each section of the piece that you're writing? Right? Good. That's good.

Adding main points and specific context

Now come the main points. And the way that works is we don't really need to use AI that much. We are essentially going to put bullets here. Under the hook, we're not really going to write anything.

But under the venue, it's... The location was the Microsoft Nerds Center in Boston. It was an in-person event.

All the things that come to mind that I wish to talk about, I can just list them out here. Three key insights from the talk. What is a good insight for the talk that I'm giving right now? What would I like you to know?

I would like you to remember that context matters. If you don't give the AI enough to write about, it will stretch it will use a lot of words to say nothing, right? Stuff like that.

And then I can go through this, write stuff that I want, and if I'm not sure, if I don't like this, it's like I can do something like this. I can go back to my AI and say, okay, I'm trying to come up with good things to say in each section how valuable would it be for a i don't know audience that wants to use more effectively for pro doc t vt all right let me see how valuable would it be for an an audience that wants to learn how to use AI more effectively for productivity to know about this bullet.

Wonderful. And then it will say, well, that would be highly valuable, da-da-da-da. It isn't the final judge of things, like AI isn't the authority over you. You read it and then you think, is this something I agree with? And if you do, great, move on. If you don't, you can choose to ignore it, right?

And then you can do that bullet by bullet and you can figure out how to add pieces of context to your post or to your blog post or to your book, whatever that makes sense, right? I will use the cheat code again to try and speed things up a bit. I have some main points here. And once you go through this exercise, we can look it over. There's quite a bit of stuff here.

So I gave a talk at the Boston AI Meetup. It was an in-person event at the Microsoft Nerd Center. I don't know how often these events happen. Are they monthly? Yes? Given our attendance here, I think monthly. Okay. The event happens monthly.

It helps non-technical people get more value out of AI. It's a great event to discover what's the rub with the current AI adoption. Meaning that what I hope, this is me obviously writing this before the event, so I'm partly making stuff up as I go along because I need to actually do this here to figure out if it's really true or not.

But my assumption is that you guys try to use AI and are probably stuck somewhere. And learning where you are stuck will help me do more useful things because in my free time I build AI tools and I program and I blog about how to use AI for productivity. And if I know where the problems are, I know what I can talk about to help you. That's the gist. So that's why I'm here.

Key takeaways. When generating text with AI, soulless writing comes back. Soulless writing comes from lack of context in the prompt.

When you work together with AI on bigger pieces, you need to incrementally go the piece. Don't just one-shot it. So that goes back to the snowflake method of I don't just write a prompt and expect it to get perfect the first time. I slowly, piece by piece, build things out and make it bigger and bigger. Those are the three key lessons.

Attendee call-out. I, of course, picked Michael because he was the one person I know will for sure be here. Then I can say, I met Michael, he's the CEO of Full Crowd Dynamics, a company that does useful things with AI and collecting data. He's really curious and fun to talk to, and just something about the person that was there. Because this format of the post is, I've been here, this is where it is, here's what I learned, and here's some cool fun people that I met along the way.

If you like this, great, come check it out because it's fun. That's that's basically it. Right. So now let's see if I get some good results with this.

Asking AI for a first draft

If I go back to my AI, and then say, Okay, new chat. Now we put everything in and see what it does.

I am writing a an event attendance post on LinkedIn. Write the first draft using the premise ending outline and main points below.

And now we copy paste all of this in premise. How am I with time? Do I need to Ah, might be good to at least get the first draft.

Wonderful. Ending. The outline is this. And the main points are here.

And now I hit Enter, and then we hopefully see some magic, or we will fail horribly. Now it will write a lot more because it has more context, but we can trim it down.

Style passes and tightening prose

We can say, well, linked in prefers short sentences and active writing. Please refactor your draft. And then it should hopefully shorten it. And that we can then look at.

Away it goes. And now I have a first draft that I can work with, so we can read through this.

It's like, guess where I was last night? That's good, because I wanted it to say that. I spoke at the MindStone Boston AI Meetup at the Microsoft Nerd Center. This monthly event helps non-technical people get more value out of AI.

And it turns this into something useful. So three key insights from my talk.

This is not something you would just take and publish immediately. You probably want to do one or two passes of editing through this.

But it is much better than this generic, oh, this was super exciting. I met a lot of fun people. It was really amazing. Because it's more specific. It has more meat on the bone, if you want to think of it that way.

Iterating titles, sentences, paragraphs

And if you then go through this process, and to edit this, you do the same thing that I just did with the titles. Give me five suggestions for it. Look at the suggestions, pick the words you like, and say, OK, I like these words. Can you give me five titles with these words, five more variations with these words?

And then you do it. It works for words, works for sentences, works for paragraphs. You iterate.

It's like, this thing, context kills soulless writing. I don't like this.

And then if you do this entire process, at the end you can get something like this, which I will find out if it works well or not. I will have to do some adjustments based on the networking tonight, but that's essentially the gist of it.

Is this big enough? No, I don't want nodes. I want this to be larger. All right.

I guess I was last night at the Microsoft Node Center speaking during the AI Boston Meetup. Here are the three key takeaways. Context occurs, AI slop, you're the editor-in-chief, scaffolding meets, networking gold, your turn.

Results, time savings, and reflections

Much nicer. Yes, it's still back and forth.

Yes, it's you writing, but because you can use AI, at least for me, I get stuff done much faster because usually writing something like this will take me like six hours because I'm super slow and I really nitpick on stuff. But with AI, this took me maybe like an hour, which is still not perfect for someone that actually wants to do this professionally. But if you want to do this professionally, likely you type much faster than I do, and you're likely much better at this. So for me, as someone that is not good, this is amazing.

All right.

Q&A

I'm probably over time, but I can take some questions, or do we just straight go to the next one?

All right. Any questions? Go ahead.

Reusable style guides and tone cues

I thought in your demo, the style guidance you gave at the end had a big improvement to quality. You linked it to shorter sentences. It suddenly made bolder statements and was just better.

Have you found, are there style cues that repeatedly Yes, and you can actually reuse that.

Like something that I'm doing for a fun project of mine where I try to write some fanfiction for myself to read. I need to go to this one though because this is nothing you can use unfortunately because it's my private chat GPT because I wrote this. For example, when it comes to this one, this is like something that writes a Warhammer 40k fanfiction with me and my brother, because we just love reading this kind of stuff and it's fun to just play with technology, right?

And in there, there is this like tone guidance. right so this is something that i always add to the prompt for this one it's like right in a grim dark gothic military science fiction style be sure to show not tell and use the point of view or and then you can play with that right or for the pros i think this is a better one no this is one i showed right so then you can play with it but and once you get the right keywords you can just reuse it every time so then what you would do is if you're in this step here

You would, where am I? Yes. Where I give it everything and then ask it to write, I would just add another context here, like style guide, and use the stick to the style guide, and then bottom left, style guide, bullets, bullets, bullets, bullets.

And then it will immediately get it nicer. But this is something you can go back and forth with. Yes.

On AI and design/taste limitations

Sir, a quick question. Yeah.

Recently, I found that the or some other AI tool, they are not good at the formatting thing. For example, it's like I create a poster for the conference I am going to present, but the thing is I want to ask them to do some kind of alignment, change some form, and also give me some recommendation how to change color, more attractive, but it didn't do that.

Right. Which makes sense because AI isn't a very good judge of taste. AI doesn't have any sort of intrinsic taste for things. You need to provide that.

So if you ask it to like, oh, what's the right color scheme to use here, it will give you a random answer because it doesn't know. So it's the wrong question to ask the AI, unfortunately.

Thank you. Yes. All right.

Conclusion

Thank you. Thank you so much.