Write workspace context and pick models by cost

Great big hello to each and every one of you. How the heck are y'all doing this fine day? Steve and Dustin here for AI Academy for June the 10th of the year 2026. And how is Dustin Stout this fine day? I'm great, Steve. It's summertime, so there's schedule changes and stuff- Are your kids off school already?

They've been off school, but they're back in school because my kids are odd in that they love summer school. Oh. They actually love school throughout the year, and when they leave school at the end of the year, they are really sad. And, so we've always put them in summer school because they love it. And, so they are, that's where they're at right now. And, I think it's only six weeks long, so won't last all summer, but- So when does your summer vacation start?

Well, let's see. End of June is I think when they're out. Oh, okay. When they're done with summer school, and then we have- When they're done with summer school, but their regular school ends when? I believe end of August. No, it starts. When does the summer start? When does regular school year end? For us, it's the end of June.

Oh, regular school year ended May. Okay. Oh, really? The beginning of June. Okay. Yeah. So it's a full month ahead of us up here in Canada, or at least here in BC. Yeah. In BC, our summer vacation starts July. It's July and August are the two months that our schools are back. Oh, yeah. Our summer vacation started a few weeks ago. Okay. And then summer school literally started, they had a week gap, and then summer school started.

Okay. Well, I'm glad we worked that out. I'm some kind of glad we worked that out. All right. Well, let's talk about artificial intelligence since our- Yeah ... since our biological intelligence seems to be failing us today. So, we've got some plans for everybody today. We're going to continue on kind of allowing the previous questions to guide us in what we're going to talk about today.

Yes. So, we've been in this back to basics series. First week we talked about, I think we just what is AI, right? We step going back to the fundamentals. And then we shifted, next week over and got a little bit more specific and we talked about prompting, right? What does it mean to prompt it? How do we prompt AIs?

And then week three, talked about personas. And so the whole idea is you're not working with a generic assistant, you're working with a specialized version of the assistant. It has expertise and is more effective at a given task, like giving them a job role. And now this week, and this is something we've touched on along the way, because a lot of what it means to be good with AI is giving the AI its proper context, right? So, in your prompting, giving it enough context. In the persona, the persona is simply more

context that laser focuses the AI's responses, right? And we've touched several times on this idea of having a dedicated workspace, and within that workspace, you have a dedicated layer of context around your business. And that's what we're going to talk about today, because oftentimes people are really confused about what do I add in there? How do I do it? How much is too much? How much is not enough?

And so we're going to talk about that today. But before we get into that, I thought it was highly relevant to take a moment to talk about the latest news in AI, because, something happened yesterday that happens regularly in the AI space. A new model is released, and all this fanfare and all these benchmark studies come out and everybody's talking about it and it's buzzing.

And it's from our good, our favorite AI chatbot company, Claude. Claude released a new model called Fable. Fable 5 is the proper name of it. And the interesting thing about Fable, is it's really engineered for agentic work. One of the things that they are saying about it is that it's really great at long-running agentic tasks. So the more and more people start to want to have these

AI agents do things on a regular basis and work for them, quote unquote, 24/7, which is another lie we don't have time for. But the more they want these AI to be running all the time and doing things all the time, you need models that can handle that. And, Claude Fable 5 is kind of built for that kind of work, that long-running agentic type of work.

But here's the kicker. It is twice as expensive as the most expensive previous Claude model, Claude Opus. Which, in Magic we give everyone sort of a baseline. What's the baseline cost of AI chatbots? And we use a multiplier system, if you're not familiar with it. That multiplier system basically says how expensive is one model compared to the baseline? Is it twice as expensive? Is it half as expensive? Is it one tenth?

So we have a whole multiplier system built into Magic to help you evaluate cost. Because the more expensive a model is, the faster you burn through your monthly usage, no matter what app you're using, whether that's you're using the native Claude app, you have usage limits there. You're using the more expensive models, it's going to burn through that usage faster. So Fable comes in at four times more expensive than the average model, and that is not insignificant. And just to demonstrate thatI actually did a little experiment for you, and I wanted to show everybody

exactly what it costs me, or approximately what it costs me to run two prompts. Okay? So let me open up my screen here. You should be able to see Magai. I'll close this out. Let me zoom in just a little bit. So as an experiment, I wanted to have Fable build a landing page for me, right?

Something we're always doing, we're doing for our businesses. I wanted it to go out, do some research about Magai. I didn't want to take the time to... It has context in the workspace that I'm in, which we'll talk about. But I also wanted it to do some research, search the web, and essentially find out what it needs to find out to build me the most effective landing page possible. So, the prompt was simply this: I want you to research our company and product, Magai. Find the best target audience, marketing angles, and build a conversion rate optimized landing page in the canvas that we can use to drive ad traffic to. The design should be absolutely immaculate

and the copy should be irresistible. Use only factual information about our product and offerings. And so it went through, it thought, it did some research, searched the web, thought again, searched the web some more, thought again, and then the research was done, built our target audience, and then gave some details about all those. Actually went to our Trustpilot, found some of our reviews, and then brainstormed some marketing angles here, thought some more, and then it began writing to the canvas, which is our document editor/code writer, where you can build a code

app. So let me- Can I ask a question about this? Go ahead Can I ask a question about that? It was building its own context as it's- Yes ... going through that. So it wasn't interacting with you, it was basically having a conversation with itself as it's building the context. Right. Correct. It's building its own context layer about Magai. And I hid this real quick because I want to show you the first version of what it came up with, and didn't want to reveal the second version.

So here's the first output. And you can see it's very thorough here. It did all this stuff. Well, let me just show you what it built first before I share the cost. And so here's what it's built. The description. Okay. So did it actually build the HTML for you? It built out this whole HTML page that you see here. Now, I didn't give it my logo.

It couldn't scrape that because it's just a text model. It's just a chatbot. It can't pull images and stuff from these web searches that it's doing. So it did the best with what it had. Didn't really match my colors, but the copy, the layout, pretty good. I went through all of it. Yeah.

It's fairly good. It looks all right. So it scraped the data from your existing website and then modified it and updated it. Right. So it basically just figured out all the information about our product, and it's decided, okay, what do I need to build? What does a landing page need to look like? What copy needs to be written to make it conversion rate optimized for these target audiences? Which it built its own audiences, right?

Its target audience, three different ones here. Mm-hmm. And it built a landing page aimed at those people with... I don't know if these are verified Trustpilot reviews. It did search Trustpilot, but I haven't verified these yet. But this is the landing page it made, and it did it in a matter of maybe 60 seconds, is all this took. Has frequently asked questions here.

You can go through. All it could have scraped from our original homepage as well. But yeah, this is what it did. Cost- Now just before you go on, would you consider that the quality of work that it's done met the expectations? Is this an improvement over your existing offering?

I don't know that it's a definite improvement. It's good from a copy standpoint, from a layout standpoint, and from a narrative standpoint. It's really good, but it's also very similar to what we already have. Mm-hmm. Okay. Now, this visualization here of switching models is pretty good as well, albeit it could be improved. Again, I'm a designer, I'm a marketer by trade, so of course I'm going to have improvements, right? The AI is not going to match my level of obsessiveness.

But I'm happy with it. I would be okay with this as a first draft, and I was. But like I said, I maybe had some notes, and so what I did next is I gave it some of those notes. But I could have given it really detailed notes. I could have picked little things here and then been very detailed, but I wanted to be generic here because I think most people aren't like me, like a experienced web designer, experienced marketer, copywriter, et cetera.

So they might have less language and tools in their repertoire to give the AI the right context to make this better. So I'm going to show you one trick that I used that anybody can use. But before I do that, I wanted to tell you about the cost of this, right? So again, if you look at Fable here, it's newly on our list. The usage multiplier is four times.

So when compared to Claude Sonnet, which is a 1.2x, and Claude Haiku, which is a 0.4x, which is cheaper than the average model, and then Opus, which is a 2x, Claude Fable is by far the most expensive. And if we sort these by usage, you'll see here that we've got some really cheap ones, some extremely cheap ones, and we've got some that are much, much more expensive if we're only looking atAction models,

GPT-5.4 Pro is way more expensive, these Pro models. But still, if you're using Fable because it's the newest, bestest one, you probably are overpaying. And so because I own the company and I can go look at the bills, right? I have some tools on my end where I can see exactly how much this request costs. So this one message, this one input, this one output, cost me roughly a dollar and 30 cents to process.

One message. Now, Steve, I don't know if you remember the days where when you were text messaging people and the early cellphone plans, every text message kind of had a cost associated with it. Yeah. It was, I don't know, it was something like five cents, something like that. Yeah. Depending on the plan. Yeah.

But then they came with unlimited text messages, and then for a while it was like you get so many text messages per month, and anything after that you have to pay for, right? Yeah. And most people using these models inside of Claude or ChatGPT or what have you, they're not given this idea of this one costs more. They don't even get any ability to see how much usage they've consumed, right?

It's just like when you hit the limit, they'll let you know. Yeah. But you have no awareness of it, right? If you understand that this model is probably not significantly better than the previous model that you were using, and it's not. Comparatively, I can tell you honestly, this was okay. Like I said, this works.

But it was okay in a rather pedestrian application. Right. It's not designed for that, though. You said that it's- Right ... designed for long-form agentic work. Exactly right. Yeah. And this is the part people have, I think, the most trouble with is they just think, "Well, I'll just use the best one. What's the best one? What's the newest one? I'm just going to use that one on this little research project I have because I want the best one." But that's not- Yeah. It's like going on family vacation using a semi-trailer truck because it can haul more.

Right. But it's not better than the minivan or the station wagon for a family vacation. Right. A station wagon's less capable than a semi-trailer truck, but it's the right tool- Right ... for that particular application. Right. So the analogy I gave, I wrote a blog post about this with my analogy about cars. Oh, okay. Well, I was close then. One of them was you don't jump in your Ford F-150 to go get the mail, right?

Yeah. You walk. Or if you have a golf cart, a Ford F-150, and a Lamborghini, you don't use the Lamborghini to go get groceries. Yep. That would be silly. So understanding that you don't need the most sophisticated model for everything. The most sophisticated models are for very complex and nuanced stuff, things like high-level math, science, chemistry, biology. Really high-level intelligence. That's what they're best at.

And that's what they're optimized for is these really complex tasks. Your copywriting, your summarization, research, web search, all of that, all you need is the simple models, the lower-cost models, because lower cost means consuming less energy and more efficient. So you only need the most efficient one.

So $1.30 roughly for this one output. And again, I had some ideas of how to make it better, so I wanted it to, number one, incorporate my brand colors. Now, number two, the design was okay, but if anybody out there has heard of this concept of Claude skills, there's a lot of this going around, this whole idea of giving AI skills, skills for AI agents.

It's a good way to essentially give the AI a persona. That's what they are. These skills that you're hearing about are just personas. They're just executed in a slightly different way. And so what I did is I went and I searched the internet for a Claude skill that was freely available on GitHub, and usually in these skills that are posted publicly, it's just a markdown file, a text file, if you will.

And you can literally just copy the text, paste it into a prompt, and now the AI can assume that skill. And so that's all I did. I just went and found a Claude web design skill that looked good, copied the text, and then I pasted it here. But I also gave it our brand colors. And all it says, "Use the following front-end design skill to dramatically improve the design. Additionally, only use our brand colors.

Magai branding colors are this, that, and the other thing." And so here's the skill that I pasted. It's very, very thorough about how the AI should be designing things. And then it went to work, and it created the next version of this landing page. And let me go ahead and switch to that version that it created here. And here we go.

I would say it's a dramatic improvement visually. We've got some animations here as it kind of slides into the frame. Obviously, our colors are used, the typography's better. Similar in layout, though, similar in execution, but just a lot of extra visual polish added to this. And it's good. I might use this. I might use this if I needed more landing pages, which I really don't.

But it did the job. It did a good job. Did it do $5 worth of good job? That's the question you have to ask, because the second message in a chat is always more expensive than the first. And so I think grand total, it wasn't $5, but it was probably just over $3 in compute costs, in API costs.For these two messages. Now, if in those two messages, your result was a landing page that generated $1,000 in revenue, okay,

sure. That was totally worth it, right? It was probably way faster and way easier than hiring a professional designer to build that for you, who's probably going to vibe code it himself, right, if he's smart and efficient with his or her time. But we really do have to understand the cost and count the cost for these models and not just assume we need the most expensive one. I could've gotten this exact same result with Claude Sonnet and paid a fourth the cost.

So it's not a matter of using, again, the most powerful model. It's just about using the model that will get you what you need as efficiently as possible. And in real-world terms, you used up probably 15% to 20% of your monthly allocation inside of Magai if you had a basic account. Yeah, if you had just a standard $20 a month account, that's about 20%.

Yeah. 20-ish percent of your monthly allocation. Mm-hmm. Pay $20 a month, get $20 for the usage. This would've been around $3, a little over $3 maybe. Yeah. That's a fair percentage of your monthly allocation. Mm-hmm. So again, maybe it's worth it to have the safety of feeling like you had the best model, but if you could've gotten the same amount of work done for a dollar, you make that decision for you.

I made a note that I think a nice demo for us maybe in a week or two is to talk about finding skills or personas in a place like GitHub and then implementing them and researching and finding other people who have written tools that work for us. Yeah. In a few weeks, we'll be updating Magai to basically use skills. We're making a big upgrade to personas to where personas will be a little bit more sophisticated.

You'll be able to import skills, skill markdown files, point the persona builder to, say, a GitHub repo that has Claude skills, and it'll import them and build them into your persona for you automatically. I imagine a large percentage of our people don't know what GitHub is and how it works. Yeah. And you should. I'm kind of fuzzy on it myself, so that'll be great.

Yeah. One shouldn't need to figure out GitHub to use AI. Nevertheless, we are curious about it. Yeah. So we'll learn. That'll be a big one once we've released the upgrade to personas. It'll be a perfect opportunity to learn about that. But okay. With that said, let's dive into building your workspace because going back to the foundations, one of the most important things that you need to know is filling the AI's context with the right information. And this is maybe a little

opportunity to talk about the other news that hit this week was Apple's new AI. Everybody hear about that? The- I have not yet heard about it. Tell me ... giant upgrade that they're making to Siri. Siri AI is the new thing that they're going to be releasing in September. They just had their big developer conference, and they had all this- Oh, that's right.

Like Apple does, they make a big deal of all the things that they're doing, and it seems like- Just one more thing. Yeah. Well, it seems like they've actually got a much better grasp on what to do with AI inside of an iPhone. Because they really couldn't do much worse. That is absolutely true. There's no way they could do worse. Two years ago, they promised us a lot of things, and I- I thought they were going to be a behemoth ... I said, "This is all going to fail.

This is not going to be good." And I could explain why, but I won't go there. This year, they announced that they're going to have a lot more involved, but they're still harping on this idea of personal AI. And- Privacy. They're all about privacy ... and privacy, right. Private, but it's personal intelligence.

The big thing that they want you to feel like is because your iPhone knows all these things about you and it has all this personal context, it can be more customized to you for the things that you're going to use it for. Now, that's a really bad idea, and the reason it's a really bad idea is because when you give the AI all the context, it can't make the most effective decisions for specific tasks. For example, if you're super

into Pokémon and you like chatting with your friends about Pokémon in your one iMessage thread, but then you have another thread with your boss and you have the AI write a response to your boss and it adds Pokémon-related puns in there, that's not what you wanted. Unless your boss is super into Pokémon, too, I guess.

But that context bleeds over in ways that you cannot control and most times in ways that you do not want. So anybody who is really good and really understands how to be effective with AI, they understand this idea that the right context is what matters. Not all the context, the right context. And one way we've done that inside of Magai is through workspaces. So a workspace, much like

projects in Claude or projects in ChatGPT, allows you to silo off your different areas of what you're working on. So your personal AI work, stuff that you're doing for you personally, doesn't bleed into your work AI work or your side project AI work.You want to separate these things out so that you can laser focus the context. And for each workspace that you create here in Magai,

you give it a name, you can give it a logo if you want to help you recognize it. You can upload documents, files, things that you want the AI to have knowledge of, and you can put in custom context, whereas you can type out the details that you want every conversation to be preloaded with when you start an AI conversation in that workspace.

And so you'll notice in Magai, you're always in a workspace. There's always a workspace selected. When you sign up for a brand new Magai account, your workspace is created for you, and you're in that workspace. There's no, like with ChatGPT and Claude, there's just a general... You're just in there, and you have to go in and create a project and then make sure you select that project if you want to work inside that project.

Magai, everything's a project, and we've done that since day one. So the idea here is fill that workspace with all the context it needs to get the job done. Right? So, our custom context here, if I can extend it... Oh, you know what? Let me see if I can do a better one here so it's a little easier to see.

I'll create a new demo here. So the idea is I want to fill it with all the information it needs about my business so that I don't have to prompt it, and every time I start a new conversation, I don't have to have the AI-- you have to reteach it, right? So here's what we have for Magai.

You work for Magai, Magai.co, which gives you the combined power of every major AI in one place for one price, access to 50 top AI models, including image and video generators, without paying separately for each. Right? So this is just the basic, what is the product? What makes Magai different? And then I go through some of the differentiators of what makes Magai a great product. Core features, I inform it of the product's core features, and then plans. I just inform it of the plans that we have.

And that's really it. I used to have a more sophisticated and more complex and more in-depth custom context for our Magai workspace. But I realized I wasn't using all that context, and it was actually flooding the AI conversations with more info than it actually needs. So, the goal here is to add minimum viable context to every conversation.

And what does that look like for most businesses? That looks like, what is your business name? What is your business's mission, vision, values? Things that inform everything that you do, and any product or service details. Right? What do you offer? Maybe who your target audience is. That's what you want to work into this custom context. So if you had a local service-based business, you would want details about your phone number, maybe your location, your hours of operation.

Anything a new employee would need to know about your business in order to start working for you. So before you get into their job role, their job description, what do they even need to know about your company in order to then go start doing the job? So we're going to use AI because this is maybe something that most people don't know where to begin with.

So we're going to use AI to help us with AI. I'm going to switch over to Claude Haiku because it's fast, efficient, and does a great job. And so for our new demo account here, we have a local coffee shop. Let me see. Uncharted Adventures.

This coffee shop is owned by a friend of mine, and I'm going to go to the website. Wait, nope, that's not it. There we go.

Sorry, I'm not getting the right links here. You guys have never met my cat. This is Minnie Mouse. She's head-butting me. She turned 14 yesterday. Okay, here we go.

Got it. So what I'm going to do is let's pretend that I am this coffee shop, local in town. I'm just going to tell the AI, "I want you to learn about my business." Actually, let me remove the custom context here. Create a new one.

My computer is really slowing down right now, and I don't know why. It's a new one, too, isn't it? It is, yeah. Maybe I have a lot of things going. OkaySo I pasted in the link to the website. Once you learn about my business, we need to build some custom context for our AI workspace. This should be the basic information that any AI model would need to know

in order to work for us. Build a minimum viable context that we can use in our workspace. Okay. Doesn't need to be super sophisticated there because we're giving it the website. It's going to go crawl that website, get all the information from the website, and it's going to actually, I guess, build us a document.

So here we go, writing in the canvas. Okay. So just based on our website, it's got our business overview, name, Uncharted Adventures Cafe. Location, downtown Bakersfield on 19th Street. That is true. What we do, we operate a specialty coffee shop that serves world-inspired coffees, fresh baked croffles?

Croffs? Sure. And provides a community gathering space. We are more than a coffee vendor. We are a cultural experience hub that brings global coffee traditions to downtown Bakersfield. Community rating, hours of operation, brand philosophy, brand promise, founding people, principal, positioning, core offerings. All of this really, really great stuff. Really, really good. And I can verify that a lot of this is true. And of course, if it was your business, you could also verify. You always want to verify, right, when we're working with AI.

Let me see. Did they get the... Oh, here we go. Definitely wanted to make sure this was on there. This is by far my favorite drink at this place, the Caffe D Churro. It's basically a churro-flavored coffee. Best thing I've ever had in my life. Awesome. As long as it has that in there, we're good. So it looks really great. It built all that context instantly for us. Now I'm going to make one simple suggestion here to the AI, and I can obviously read all it says here as well, but I'm going to tell it we need this to be as efficient

as possible. Details are important, but we don't want to overload the context. So let's trim this up to only have what is absolutely necessary.

And one thing I love about Claude Haiku in particular is it is, again, a more efficient model, and when you ask it to be more efficient, it tends to be pretty good at that. So there we go. Here we go. And this is it. There's not even anything to scroll here. It trimmed all that up and I could, again, go through and audit this. I don't think we need the rating and the Yelp score, but all this looks really good.

I'm going to save that, and what I'm going to do here is get that out of the way. I'm going to copy that. I'm going to open up our workspace and hit the edit button for the settings, and then I'm going to just paste in all of that, give it a quick look-see. Got an extra H1 there.

And there we go. We have our minimum viable context for our coffee shop workspace. And that's as easy as it is. And now every time I start a conversation, the AI will already know that information about my business. It will be built into every interaction. And it's easy, after you've used it a few times, to catch the few things that it needs.

For example, I've built a... Well, in this case, it's a prompt, but it's very similar, for creating a synopsis of each one of our AI academies. And it tries to spell Magai the wrong way because it doesn't know. So I just told it in there to correct the spelling, that this is a unique spelling, and never had that problem since.

Hmm. Perfect use case. Yeah. Anytime you have special instructions or special things that you've noticed that the AI will do, little quirks that it has, this is one more place. You could, of course, handle that inside of the persona- That's right ... instructions if you had a persona. That's what I would, yeah. But this is one more place where you can, unlike a persona where you have to invoke or activate the persona for those instructions to be given to the AI, the workspace context, again, travels through every conversation.

Doesn't matter which AI model, it doesn't matter which persona. This information is always present in the AI's custom instructions. And it's not- It's dynamically inserted into the system prompt. And this isn't one and done. You can tweak this as you go. Yeah. And it's super easy, and that was part of why we redesigned Magai the way we did. We wanted this to be easy to edit, easy to access without ever leaving the conversation. So if you find something that it does or says and you're like, "Oh, I need to update this," you can jump right into here, make any edits you need, and

save the changes and go about your business.Now, one thing I will note here is you'll notice that when I copied this rich text, we, as humans, we see these bold bullet points, italics, et cetera, headings. The AI doesn't see rich text. It only reads plain text. And so you'll notice that it kind of transitioned everything to these symbols, hash marks, and asterisks. For those who don't know, this is called Markdown formatting. It is a shorthand way of representing rich text. And so these two

asterisks before and after the word name here, that indicates bold. The hash mark here, single hash mark, represents a H1 heading or a heading one. Two hash marks represents a H2 or heading two. So it's just formatting for language models, essentially. It's not necessary, it's not required.

But just like a human, when we see rich text formatting, it helps us to better categorize and understand things. The same works for an LLM. When you use these Markdown formatting, it is just a little bit easier for the AI to sort of separate things and logically interpret things.

So not necessary, but does help with interpretation and comprehension. Great. All right. We're just about 40 minutes in. Let's get to the Q&A. If you have questions, please drop them in the Q&A, which is the little question mark chat down in the bottom as opposed to the general chat. And our first question's coming from Michael, who says, "Getting solid copywriting from the AI model is important to me. In your opinion, are there big variations of output from the lower cost models compared with Claude Sonnet?" Which is his go-to

model for copywriting. Big variations? No. I would say subtle variations. I tend to go with Sonnet just because I found it to be the best balance of efficiency and quality. Opus does produce slightly better copy in many scenarios, but again, if you're super picky, I guess it's noticeable. For most people, it won't be that noticeable.

So it's just up to you. Manage your costs and how valuable that copy actually is for a given scenario. Carl's asking, "Would you be able to do a session on using the integrations in Magai now that there are so many?" Yep. Yeah. That was actually on the docket for today to see maybe we have time for it, but I was like, "No, we definitely need a dedicated session for that." I actually had an example lined up, too.

Magai can now post to Instagram and LinkedIn for you. And I used it yesterday to post to both my Instagram and LinkedIn in the middle of the night just from our chat interface. So definitely some really cool things you can do, some tools you can be replacing. And as soon as we introduce scheduling tasks in the next week or two, maybe three, there will be a lot more to talk about there.

So I kind of want to wait until we've released the scheduling aspect of that before we dive into the integrations. But we'll see. We'll see where our development's at. But definitely going to dedicate a whole session to that. Okay. Michael has a two-part question. "Should we use similar protection language to what we put in personas about never divulging our instructions?"

Well, I guess that depends. Are you allowing people into your workspaces that aren't working for you? Because Magai is a tool for internal use. Whereas personas, you can share personas publicly with people, and they can use your personas, and they don't have to be part of your organization, they don't have to be on your team, whereas a workspace, that's limited to you and your team.

So I don't see a scenario. Maybe you're letting clients into your Magai workspace and paying for them. That might be a scenario where maybe you do need to protect that workspace knowledge. But again, what I'm putting into all my workspaces are publicly necessary information, things that are mission-critical. I'm not putting any proprietary stuff in there because anything that's proprietary tends to be very use case specific and not applicable to literally everything we do in a given conversation, at least in all the scenarios I've run through.

So we wouldn't need anything proprietary in the workspace context. And that would be in more of a persona. And then the personas, of course, you share only with the people who you want to share them with. And so, yeah. I don't know that there would be a reason to do that unless you're inviting clients or people into your workspaces that you don't want knowing certain things.

Second question, "For my primary business workspace, I filled the context with information about various IT solutions that we represent, which is a lot of info. How much is too much for the workspace knowledge files custom context?" So various IT solutions that we represent, a lot of info. So I would say if your context is longer than 2,000 words, you're way overdoing it.

That's just way too much information. For one, some of the less sophisticated models can't handle that much of a system prompt in such thatThey can't hold all of it in equal value, if that makes sense. So one of the concepts in AI is prompt adherence. How closely does the AI adhere to the prompt you gave it without deviating or without missing things? The less sophisticated models, they

don't have as good a prompt adherence as the more sophisticated ones. So it means the more context you give it, the more it misses, and the more it just can't take into consideration, because it doesn't have that processing power. So you really want to trim it. So would you also consider if it's important for you to have that much context, that much information, would it be a possibility then to look at a different type of solution? Say, going to a NotebookLM, where you load in a large amount of context, a large amount of reference, and you use that as a basis.

Well, even in a NotebookLM, all you're doing in that scenario is you're using RAG. It's not- Yeah ... adding system pro. So it- So explain what we just said. Explain what we just said. Yeah. No, I will. In the workspace context that we're talking about, this is being inserted into what's known as the system prompt. We talked about this with our personas. Yeah. The system prompt dictates everything about the AI's personality and how it responds.

Whereas when you upload documents, you upload knowledge, or you upload things to a NotebookLM, this isn't being inserted at the system level. That data is being converted into what's known as a vector database. It's being put into basically a very sophisticated spreadsheet on a server, and all of the information is being what they call chunked. So it's taking sections of each document, assigning them numeric values, because in the AI, all your words are just math, and it's predicting the relationship between words

to use in its output. So when you upload something to a RAG database, a retrieval-augmented generation database, it's not fully aware of everything all at once. That's the big mistake people have about this. They misunderstand how that works. NotebookLM is probably the biggest reason for this.

They think, "Oh, everything I put in NotebookLM, the AI model in NotebookLM knows all of it." No, it doesn't. It has access to it, and when you make a query or when you send a prompt, what it's going to do is it's going to search that database for the math, the mathematical relationships to your prompt. So it's going to assign a bunch of math to your prompt and say which of these database chunks.

It's going to do a search, essentially, and the top result, it's going to get that chunk, not even a full document, it's giving it a chunk, one piece of that document, and then use that to inform its response. It's not omni-aware of everything that's in there.

It's just a big magic trick, a big pattern-matching magic trick. Now, NotebookLM and some of the more sophisticated ones, they might run multiple queries and use those multiple queries, but again, at the same time, it does not know everything all at once. It's impossible. It just doesn't have that capacity. No AI has that amount of capacity, so... Okay.

Carl's asking, "I use workspaces for different departments in marketing, funding, social media, et cetera. What examples of information would you add to the content?" So I guess he's kind of asking, what is baseline the same information that would go into all of them versus adding specifics for those departments?

Yeah. So ask yourself that question. If I had 10 employees, what information do all 10 employees need at all times? And if you're using it for specific departments, ask yourself, "I have a marketing department. What information does everyone in the marketing department need to know?"

And only you can answer that question, right? I can't tell you about your business. And so you could theoretically have just a workspace for your business and then use a prompt to give them the specifics of, "You are a social media expert," and then give that information as you- Yeah.

If you wanted to streamline things. Yeah, absolutely. Yeah. You can run the same exact query that I ran, right? The same exact prompt that I ran with the addition of, "You are a social media manager. Help me build the context specifically for our social media department," or our funding department, or our marketing department. Run that same prompt.

You might even do the generic one at first. Do the generic one that I showed and then rerun. Instead of saying, "We need this to be as efficient as possible. Details are important, but we don't blah, blah, blah, blah," say, "Now refine all this and create a minimum viable context just for my marketing department. Give me a minimum viable context just for my funding department." Right? So you can kind of refine it, do the big generalized one, and then refine it from there. You can even do that multiple times in the same chat thread, and create all your different contexts in the same chat thread, and it'll become more and more efficient too as you do it because the AI's

learning what's useful and what's not useful for each department. You can stop screen sharing, by the way, now. Oh, okay. Yeah. Brad is asking, "Would you put the website into the workspace as well using your development on the tech?"And as using it in your development of the text. So would you include the base information that you used for the original search?

Yeah. So I put it in the prompt, obviously, because I want it to go scrape the website. But I put it in the workspace context as well, because if I ever need it to write something and it needs to reference the website, it's not going to know my website address by default. Especially for us at Magai, right? We have magai.co, but we don't own magai.com. Most AI is going to assume that we own a .com domain.

So you definitely should add the website into the context, as part of the context. Just like if you had a local business, you would want it to know your address, your street address, right, and your operating hours. That's essential information that it needs to know. And every employee would probably need to know your website address, right? So you definitely want to put that in there. William's asking, "What is the concept of the canvas? Is this an industry standard?"

Different apps will call it different things. This is just a place where the AI creates files, creates and displays files. So I believe in ChatGPT they call it a canvas. In Claude, they call it artifacts. In Gemini, I don't know what they call it, because they have 15 different apps and call things different things in each one.

But this is a window where the AI has created a file, and we're displaying that file and able to interact with that file, whether that's a document, a webpage, a web app, et cetera. Alfred's asking, "Can we format documents in a workspace so letters and other documents are printed with specific fonts, margins, et cetera?"

Can we format documents in a workspace? When you're asking the AI to produce documents for you, it's not as if it has access to Microsoft Word. It doesn't have access to a Word editor like Google Docs or a PDF editor. It's using code.

And if the fonts that you ask it to use are readily available on the internet, like Google Fonts, it can usually incorporate those and write code to use those fonts. But when you're uploading documents into the AI's knowledge, it's not like it's reading the fonts that are in the document files, if that's what you're referring to.

So if I were to upload a Word document that contains all of our business standard operating procedures, for example, or even our brand style guide, unless the text in your documents explicitly says, "Use this font," and that font is readily available, it's not visually seeing the fonts that are used inside the documents.

But if you output your document that you are bringing from the AI into a markdown, you ask it to output a markdown, you bring that into a Google Docs, you can very quickly modify the fonts, the headers, the format, all the formatting. Because once it's in a markdown, it's an entirely formatted document. Every letter on that page is applied to some format that you can change globally.

Yeah. So if you need very specific fonts and very specific styles for PDFs and documents, you still are going to need to manually- Yeah, it's a two-step process. Still going to be way quicker than doing it yourself. Yeah. Copy and paste, much faster. Yeah. Elvava says, "Anthropic released Mythos this week to everyone. What does Dustin think about Mythos?" They did not release Mythos. They released Fable.

Is it Mythos or Meethos? Mythos? I thought it was Mythos. I think Mythos sounds better. Okay. Well, we'll use Mythos. They did not release Mythos, they released Fable, which is sort of, in their words, a tamed down version of Mythos. Mythos, what was it a month ago or so, they talked about this internal version of this new AI model that's so dangerous that they can't release it to the public. That it's scary. All marketing hype, by the way.

We're not ready for it yet. Yeah. So, no. They did not release Mythos. They released a tamed down version of it in Fable 5, and that's what we talked about earlier today. Okay. Michael's saying, "Following up regarding overloading workspaces. Perhaps I trim the workspace, then create personas that are based on the various solutions." Good idea. That sounds good.

Yep. Absolutely right. And let me just add- But again, I will say you don't want to overload. So even in the personas, that's being added to the system prompt, right? If you're uploading documents, it's being added to a RAG database. But so just understand, the more context you give it, the less ability it has to follow everything in that context. It's going to lose something at some point.

Lisa, thank you for asking. I just uploaded into the chat the Magai chat or the Magai link to give you extra, what is it, three months free access. Plus it credits us as the affiliate. So thank you. With that, I think that's the last question. And look at that, five minutes to go.

Perfect timing. Time to get to the park before we do Webinar Wednesday today. Dustin, thank you so much. Good job today. We will see you next week. So you've canceled your trip to Vegas, so we're okay for the rest of the month? Yes. We are in the process of shifting our Vegas trip. Again. So- For the second time ... so heads up for the entire community.

We will be doing our AI Academy for the rest of June. But come July, it'll be alternating weeks as Dustin and I find some time to do a little suntanning. So there, we'll do that. All right. Dustin, thank you so much to each and every one of you. Thank you for your input, thank you for your attention, thank you for your time, and thank you for being you. Till next time, have fun storming the castle.