AI in Context, Ep 04: Turn a call into a document

Machine transcript of the replay, unedited. Times link to the moment on YouTube.

0:00I want to go over like this tagline that I've started saying to everybody, which is agents are amazing at bureaucracy. Like they they are so much better than humans at bureaucracy. And you kind of like, what do you mean? Well, bureaucracy, in essence, if you like, we all deal with government and companies, like, especially large companies, it's like documents and rules. That's it. Bureaucracy is just billions of documents and billions of rules for different scenarios, right at scale. Agents are really good at producing documents. The problem is we're not very good at communicating them the method by which they should go create documents for us.

0:51So in essence my objective today is to give you rules for good AI documents so that you can just set these up and stop thinking about needing to do anything when you want to produce a document. right? Set them up and then stop. Like the same rules that I started with, like I was talking with the owner at the beginning of the year, I had started the rules there. I have not changed my role since about April. So it's like one of those things where I do not now need to go and maintain these rules because the A4 paper size or letter US size is not going anywhere. That's a standard, right? So we're pretty comfortable in creating rules that don't change for that. I do think eventually models will get good enough at doing this by themselves.

1:40The problem is models are awful at art and art is about taste so you'll always have some level of leverage with regards to models of generating something that is more adapted to you. I've got a few examples of documents that I generate here and my goal is I will show you exactly how to generate these kinds of documents. So let me get started. Here are the six rules for good AI documents. First, set the page size. Tell it how big you want. If you want a letter, if you want a four, tell it that. Two, give it an outline. It does not need to be long as an outline. I will show you later. I've got some brief setup here. This is an outline, you know, owner, Priya, two shops, like I'm not talking a big outline here, I'm just talking, give it an outline of what you want.

2:33Three, fewer words. One of the things that AI loves doing is rattling on endlessly. You need to be, you need to put your like editor hat on and say I want you to cut ruthlessly until you have only the essential ideas being transmitted. Otherwise it will go off and talk and talk and talk and and what ends up happening when it talks too much and you have like your repo, which is your company and it's talking a lot, well, it's gonna share all your secrets with everybody cause agents don't care about sharing secrets. They'll say, oh, our infrastructure is not actually ready cause you have not done this specific thing and it will say that in like your documentation and you as a human need to be like, yeah, but I'm working on that. So let's create the document for the final version and not the like stage I'm at right now, right? that's a human thing.

3:26Then show it how you write. The best way I have found to teach an AI how to write and how to do things is I take all of my emails for a year, I take all of my transcripts and I give it to them and I tell it synthesize my voice and remove all of the AI generated words I have sent. So this allows the agent to immediately like create a skill that is your voice based off of data and it's very useful and it's used across everything you do with agents. Five -handed assets don't allow it to generate stuff without you giving it more. You need to give it art. You don't want it to be doing everything by itself. You want to be giving it art and assets so that it can use those art and assets.

4:19And finally, make it check its work. So that's kind of like the six steps to a good document. Let me like go over A, one, set the page size, A4 letter, whatever you want, give it an outline, make it do fewer words, show it how you write, hand it your assets and your drawings and your PNGs and whatever, and make it check its work. That's it. There's nothing else with regards to generating common documents. Now, there are two ways, like this is a common question. How do I create art that is mine and my brand and that I can use them across a lot of things? The best way to do that is to use, I use ChatTPT to create these kinds of like art assets. And I use Claude to create more programmatic style assets. That is the personal workflow that I found works between the two of them and I will be happy to show you how to do that.

5:18So if I go here into chat GPT and I go to here create contract illustrations. This here is like I gave it two references. This is one reference and this you can find online and you can ask it to modify another image you found online initially to like create your own asset bank. And then you give it another asset, and these can be two wildly different styles if you want to make a mashup between them. And you give it a prompt like this, like make one illustration of the exact style of the two attached images, right? And then it creates it. The thing is, once you have this thread within your chat GPT, just save it. That will be your thread for generating new assets when you need it. And the thing is, whenever you have an idea that would take 100 words to explain, like for example, a process to generate a page from an idea to a paper letter, like that is an idea concept that I would need to use a lot of words to explain. I would much rather just use an image asset to generate, to transmit that idea in essence.

6:24And ChatTPT is excellent at creating that in your own style at scale. Like I have generated hundreds if not thousands of images like this now and allows me to create processes. There you go. So you see, this is in my style. The only issue I have is now this is not a transparent PNG. Make it a transparent PNG, et cetera. Really, the process is just that in the essence. If you do have a project that forces it to create transparent PNGs, ChachiPT still messes up sometimes, because it has a hard time understanding the concept that these little square things are transparent. So it ends up creating things like that sometimes, but it's just one thing away. And then of course AI, like if I want to make more isometric things like that, you can ask it to create isometric shapes for your concept. This is all made by Claude in like, these are all lines that are drawn in as an SVG, super easy for Claude to make this. This is a trick. I will send you guys a repository after that has 70 plus of these isometric examples that you can use for your own stuff and that you can just give us reference to Claude and it will be able to then generate them en masse for you.

7:41I've gone through the hard work of creating that initial asset bank, so I'll just share it with you and you guys don't need to worry about that. I'm going to stop there before going on to the exact steps of the document. Are there questions before we go there? So now I've covered the rudiments of what are the rules of making good documents, the importance of imagery and iconography and how to generate that. Any questions relevant to your process? Quick question from my side. Do you know any repository of assets to be inspired from? Like, I don't know, Mobin is a repository of screenshots from different apps. Do you know any place? Because I don't even know the words to describe graphical styles. I would like to just browse a thousand of them and pick that one is the one Yeah, no, good, good, good, good question.

8:35What you want is image, asset, bank, illustrations. There's a few, there's a few massive ones. Most of them are paid to some extent. There is a question of copyright infringement as well. If you just like straight copy, that's why I'm recommending pulling multiple of them. You have a lot of free asset banking, like, Like free assets like this that you can go ahead and download from hundreds if not thousands of websites. In essence, what you're looking for are free packs of assets. The, you can also have a conversation with Claude with regards to explaining to you the major styles. So there's vector there's flat art, etc. But in terms of references, I often go to Dribble for inspiration.

9:26You just need to log in with an account and you can go ahead here and, you know, look, illustration. And this will give you a bunch of, like, really nice illustrations from some really good artists as inspiration. But as I said, again, do not steal other people's works directly. always try to mash up to create something that is your own and like modify it but does that does that give you do you I can also send you like a list of ten websites after that I go to with your as a reference yeah to be great thanks that's a good that's a good direction cool wonderful so in order to actually generate these documents so it's always outline sources page size draft short check safe. This sounds complicated, I've made it a bit more simple here because what I have done is I've created a bunch of context files that explains exactly how to go about this, what are the outlines, what is the brand, the charts, etc. So these are like you create a project for creation of documents. I try not to create documents in every single stream of like if I have a chat with with my agent and I get to a point where I say, oh, I need to create a document for this. I would much rather tell that chat to summarize for me. And so like create a prompt to generate a document for this and then go to the project that I create documents in instead of trying to do it in line because the agent has polluted context of like all the steps to get to where you got to. And so it's gonna make it bad at generating documents.

11:06You would rather have a fresh session for document generation so that the agent is solely focused on the document generation component and not be thinking through all the logic to get to where it needs to create a document. Does that make sense? Cool. Once you do do that, very much it is one -page proposal for Bloom, Twig, and Flores. Like this is at the point where you have all of this context, it's really just about giving it the context and allowing and allowing it to generate the image itself. I'm gonna go back here and I'm gonna go to the other. So one page proposal for Bloom, Twig, and Florus. I'm doing the same thing on Chad and Claude to show you. Like here you have all of the different sample documents that I've given to it so that it has my assets, it has my pictures, it has my writing style, it has the outlines, et cetera. And so that it knows what to generate when I ask it to generate something.

12:03Generally speaking, this takes about a minute. nothing more. So that's all you need to do with regards to that. What's really cool about document generation is once you've figured it out once, you no longer need to re -figure it out. Like it's it's just the same project with the same script file so that you just create a new thread every time and you can generate documents like that. I'm gonna wait for these to be finished. I believe chat GPT is usually a bit faster but we will see generally speaking I My personal take is I prefer using Clod for this. I have used both ChatGPT and Clod for this. I do find that Clod has generally slightly better taste when it comes to design, but honestly, it's something I revisit every two months, every time there's a new model released to see if they've gotten better at front -end and generating documents. So, here you go.

12:58Document generation, proposal for Bloom and Twig, prepared for Priya, where you are. It included the asset. that here's what we'll do, here's what the proposal is, here's the price and time, and here are the next steps, and then you enter the information. Like, as simple as that, same thing on white floors, as soon as it's done, we'll be able to see that one. What a, one thing here is, like, this is just my experience with this. I try, as soon as I've generated a document, a PDF, I want to save it elsewhere than Claude, elsewhere than chat GPT documents get lost so easily and if another like quick tip here is if you're doing a lot of documents and you're doing a lot of versions of those documents start versioning faster than you think like so anything that's related to a company you're building and you're thinking security documentation you're thinking onboarding documentation you're thinking, anything to do with that kind of dock, start versioning from the get -go. Because you're going to need to change that document, and you don't want to start from scratch every time. You want to start from the previous version and improve it iteratively to get to the best version. So you start with version, like I personally always start with version 0 .1, because I don't consider the first version of whatever I do to be something that I would give to a client. Generally speaking, my version 0 .4 it's the one I'm going to give to a client. I've already gone with the model and made three passes of modifications to get it where I want it to be.

14:40There you go. So let's see if there you go. It has it done here. As you can see, as I was saying before, I much rather play with Claude than play with chat TPT on document creation. Generally speaking, I find that it makes better assets. Chat TPT does not use the assets. I could tell it, add an asset. And I could have this kind of conversation with it to improve it. But my default is to go with Claude, because it likes to think with pictures. It has a bit more personality there. Um, so yeah, keep it in your, oh, Matt has that question. Yes, please. A quick question about it. So I might've missed it. So do you have, is it just the assets you store in a project and those and Claude co -worker and chat GPT, or do you have skills for them? Do you have the same skills and chat is just using, choosing to ignore those skills? Like, is there any difference between what they get into context or is it just like their behavior? your gets not yeah this is this is a situation where is a one one exactly the same context chat GPT tends to not go for image and visual visualizations first as a first reaction Claude does the ways around that as you've already said is yes skills can help with that especially if you've got your cloud FD or your agents at these or like that mother document indicating use assets as a default, that's a great way to do it.

16:20I'm not sure if you asked this question, but I'll say it anyways. It's a good idea to put all of your assets into a Drive folder or whatever Cloud storage you have, and have it pull from there and also save news assets to that folder just to stay organized. It's very, very easy to get lost in terms of asset generation and not remember, I remember creating this one picture and you can only find it in a document and it's really annoying to extract it from that document. So it's a good habit to have that whenever you generate a new asset to have the agent save it into your drive or whatever so that you can access it another time. Do you have them describe those assets in some way when they save them? Like for example, verbose file names or do you just rely on simple names that they just analyze what's in the picture to know which one to look for? I do the latter personally. I think both could work. It depends on how technical the images are. I've got some of my images are a bit more technical in nature and I have like infrastructure documents. Those I would much more, I want to be more verbose to describe them, so.

17:33The, this is like, this is what we're talking about here. Keep it in your drive, not in chat. So your brand should have its own folder in your cloud. So logos, colors, your writing style, I consider that part of brand. It's just an MD file that explains your voice. If your skill, like this is where context sprawl is a problem, where you're trying to keep up to date your skill with regards to your voice, but at the same time you have this version in the drive. I have not found any magical solution to that problem yet. I do think it's a question of time before we get there, but like you have to be aware that it's better to have a version of your voice that these baseline in your brand folder so that when you do generate even documents that they are consistently in your style of writing. When, when I say brand voice, I'm not talking about no em dashes, like you can add that, but I'm more talking about real examples of your writing that the so multi shot prompting is the like technical term, but like real examples of your writing that the model can reference to understand your writing styles, Because the more you give it, the better it is.

18:42So if you have 200 examples of your writing, that is better for the model than having three. So the more you can give it, the better it is in terms of quality of writing. Same thing. Library pictures you approve, assets you approve, templates, outlines that work. So you should start building a repository, a saved bank of proposals, of onboarding documents, of like letters of understanding, things like that. So a letter of understanding is part of a sales process when you have a meeting with someone and you communicate with them and you want to send them something after that meeting. That's not a proposal, but it's like something for them to, for you and him to say, for you and them to say, I understood what you told me, I'm just saying it back to you so that we're on the same page. Like those are the kinds of things you would like to have as a template in your document folder. and then finally the finished PDFs themselves. So I keep everything in one folder and then you can dispatch it accordingly and it's really easy to just provide that reference and then your models always have access to everything.

19:53Making your first document is exactly what I showed you. So it's here, I've got a project here. My project has the instructions. So you make documents that look good in our right. read the project files before every document. So every document, every time, page, letter size, outline, few words, my voice, my look, check. So this is like a basic system prompt that allows me to have very consistent process with regards to that. And I tell it personally, if something you need is missing, make the document anyway with a clear blank because the agents are a bit messy sometimes that they refuse to produce it because they don't have a specific information so now I'm just like no produce the document first please and then we can have a conversation.

20:44Context add this as you go you can also add connectors here etc. Scheduling is also useful as in terms of a managed agent so if you want to create like a monthly report or a weekly report this is a great place to do it if you also have connectors connected to it. If you are doing client work for people this system will allow you to make weekly reporting a breeze. because give it the connectors that you need, give it the process to make documents, and suddenly everyone's happy because they're getting their PDF report every Monday that tells them exactly what you did as their subcontractor, as their vendor, so much easier, and that's a big lever in terms of using agents. There you go. Then because I've done all of this previous work, it makes it very easy for me to just go ahead and do simple stuff. So let's give you another example of that simple stuff. So here I've got a CSV in my context that has some data, right? Where is it? Sample data. So I've got Jan, I've got January to October, bookkeeping, payroll, client questions, tax prep, admin. And all I want to do here is make multiple styles of graphs to showcase what you can do on one page.

22:05So, like, one of the most useful tools of agents is making pretty visualizations of data. It's like one of those tidbits that you might not think of. I do not spend any more time in Excel for making graphs and visualizations. I don't use Power BI anymore or anything like that. It's so much easier for me to just on demand have an agent generate dashboards visualizations for individuals. And unless I need like a data pipeline or something, this is a really useful way to show your late person, your late visualization of your work. So strongly recommend using cloud for data. I tend to use slash design for that nowadays.

22:56Design has been so much better and creating artifacts to give me options, multiple options of visualizations, and then you can tweak them and combine them. And I really like that process. That is like, that's an amazing tip as well. Like if you are doing this as a more constructed format of like, oh, I'm just making visualizations, cloud design is very good. Note here, Claude Design as a separate app is being deprecated in December and it's being merged into the Claude itself. If you go into Claude and into your settings over here, you now have design systems. So everything is actually integrated into Claude and you can use Claude to generate new design systems, etc. So anyone who's gotten familiar to the Claude design interface, it's going away, unfortunately.

23:55we can't get used to anything in the AI era before they take it away. You need to do that now into your cloud stream, but I agree, you're very right with regards to that. Now you can see it's already making some data visualizations. What I like to do personally for Data Vis is have it produce multiple different Data Vis, and then I get to pick and choose what's my favorite among those that he's generated. So that's always a fun one. I'm looking forward to this one being done. So you can see converting Inter to TTF, building the PDF and previewing it. So it's going to be generating a lot of like these interim steps and you can always ask it to go back and generate the HTML version or like a shareable version or a webpage version of these same things and make them interactive, so. Simon, your audio is a bit, it's going up and down or maybe my problem and don't know whether anybody else hears it, sometimes you're not really loud.

25:01Apparently I was, I was moving away too much from my mic. So here you go. These are just different examples of graphs that it's been able to generate from that same data set. Do I find these good? I find them a good baseline and then I can start integrating. Oftentimes what I'm going to do is I'm going to create my graph separately and then use these and download them or add them to my drive and then go use them elsewhere. Because I have found personally that trying to do graphs at the same time as doing the rest of the documentation can end up with weird artifacts because the agents tend to want to prioritize one or the other. It feels in my experience that in the weights, the place for the numbers and the graphs and showing data information that way and the place for art are not close by and so the model needs to kind of jagged itself into both of them and so it ends up creating something that's closer to the mean in terms of like average which is boring. So from a design perspective I've tended now to do graphics on one side and then assets on another and then assembling them as a third step.

26:17There you go and then pretty charts this is the what I'm doing right now this is allowing me to do these charts like this, here you go. This is another version with regards to chat GPT allowing you to do the same thing. So I've given it more of a structured prompt to create a PDF of that same data. And this is what it ended up coming up with. So this was again, the same process of generating the graphs first, giving that data and then allowing it to generate a document, giving it the flexibility to be able to modify like the graphs. I'm not being like a hard ass. I'm saying, do not touch the graphs. I am telling it, here is your reference. Here is your data set. Here's what I already created. Now generate a PDF document for this. So you can clearly imagine how you'd be able to generate documents around this.

27:08There you go. This is the end of my talk this time. See, I kept it more concise because I wanted to be able to have more conversations. So give ages the rules. You keep the judgment. This allows you to create documents that scale very easily with regards to process. So questions, conversations, give me a challenge, have me create things on the fly. So I can have a question if no one wants to go. One thing I noticed is that for documents and for slides, and this might have changed in the past like three months, like everything did, but agents are kind of bad at like keeping things on the page because they just, due to how the HTML works, they just keep adding content and eventually it falls off and they're not very good at detecting that and checking that, do you have something for that so that the agent can verify that everything is laid out properly.

28:04I have found, I went through multiple iterations of this. What ended up working for me is simple is better. So, what I mean by that is within my instructions, let me go to Claude because I prefer this, I gave it this instructions all the time. So, I give it the workflow steps to follow in terms of one, start with the page size. Then at the last step is always check. Every fact comes from message, never endpoints, etc. etc. So this process in terms of a very simple, very concise system prompt has made the document generation process much easier. I do not put anything else into the system prompt.

28:55So your Cloud MDE file that's for documents should have a very small, very succinct system prompt that is constantly reminding the model that it needs to do these steps every time it produces a document. So that has worked. Another question. So yeah, I just I just wanted to ask. I've been playing around with like the document generation quite a bit, because in like research, my prof always asks, like generate a manuscript, whatnot, whatnot. Yeah, I wanted to ask if you've tested how far these things can go. Like, I know there's things like the anatomy scale, and like, they can now also do a really good job in blender, like for 3d stuff.

29:53I don't know how like, relevant this might be just, I don't know your work regularly, because I don't know that many people that need 3d modeling in general, but yeah, I was wondering if you if you know if they're like, capable enough to like do a good job at this sort of stuff or? So I can show you this kind of stuff, which is like white papers. You can generate these kinds of documents fully within the model. So from creating the isometric views, embedding them, doing embedded views of images with tables, with logos from companies, step -by -step processes, the different pro, the different steps of the, the engine, and things like that. These are absolutely within the realm of what document generation is possible.

30:48Any of these could be replaced by a 3D model, for example. That's not a problem. So I have not found any... limiting factors with regards to generating documents, more or less I can go as complicated or as simple as I want here. Let me let me find you other ones that I've generated with for my own company. Here's the Sherpa playbook that I'm working on. So again, same isometric view style, different steps, different team members, different processes, arrows pointing to the process, Gantt charts, different items, like, you can, by using the same building blocks, you can actually assemble extremely complex documents using entirely the process. So does this, like, is this kind of what you were imagining in terms of more complex documentation?

31:47Yeah, yeah. Like this isometric view specifically, is not. And And so this is, I'm guessing you still had to somewhat like iterate to get this or it was just able to like one -shot it. Once you have the artifacts working in terms of like creating the library of, and like the basic rules of how to design an isometric document, it becomes relatively easy to have it iterate and do thousands of variants on that. Um, the, the bottleneck at this point becomes you actually going through and reading this, um, like this, this version, I believe here, you see, this is V 0 .3. So I've already done three different steps of going through this and making sure that it's a bit more, it's a bit more, uh, cause six cent in essence. So like you can do that. Um, you can even go a bit crazier on this and let me go get you something a bit crazier, which is, where is it?

32:55I believe I have that on my desktop somewhere. They modified. I'm trying to find motion animation because you can use the same thing to generate actual videos for yourself. Here it is. Product screens, art scenes, films, there it is. Hero V3, Ask V4, MP4. There you go. Pretty sure it's this one. There you go. So this is the document generation process brought to another level. So the same principles that allow you to generate documents allow you also to generate motion design and anything to do with animated views. The thing that brings it also when you do start going from text to video, it's you have to be aware that you can also add stuff like sound design.

33:56So you can go crazy in essence with regards to document generation. Fair enough, yeah. That video was very impressive. Thank you. If you guys want me to do a session on just making motion design and videos I'm more than happy to. It's a fun process, but the fundamentals to make a video are the same fundamentals to make. It's very much design your assets, have the process in place, and then ask the models to generate, and then iterate on it. So that was the fourth version of that video, by the way. I can show you V1. No, that's okay. I was just wondering, Sorry, maybe I'm asking too many questions.

34:50Are you using a skill or you just kind of iterate through it taking your time? Or what do you prefer usually for this? I don't like using skills as much as other people because skills are... Like if you think about a skill, a skill is a process and the knowledge of that is like experiencing that skill is the same model. Like the weights don't change, skills just give it a direction. So I would much rather use a skill as a reference. Like when I go on the internet and find skills, I actually almost never use skills that I just download from the internet. There are very few exceptions. Grilled me is one of them. I love the grilled me skill.

35:42There are like two others that I use, but most skills that I use in my own workflows, I have taken inspiration from something online and then I have absolutely generated the skill for myself. And oftentimes what I do is I condense the skill and I make it smaller. So my recommendation is use skills as inspiration but make them your own because your context is your context and the value of your judgment and your taste in your context is much more valuable than the SOP, the process that is provided by that skill from some random stranger on the internet as well as, you know, the safety concerns of downloading random skills off the internet. No, very cool. Thank you. Yeah, Fritz. I thought I could connect to the question I was having. Since you're jumping between the two tools, GPT and Cloud, how do you keep your context and instruction similar and both?

36:45So essentially, since you're working on the same thing manually, or just wondering? My take is always, I go ahead and use the connector to go get my GitHub repo. So when I am working across multiple tools, even if you're non -technical, it is a very good idea to generate a GitHub account and have a repo for your business activities for everything to do with your general purpose model usage and put everything in GitHub, that makes it easy to switch because then the only requirement is that you push whatever you've been working on to main or to the GitHub repo so you commit it so that when you're on the other platform it allows you to have access to whatever the other model works for.

37:35A little aside on that if you are working more locally through the terminal on your computer or even through the apps personally my agents .md at that github repo and my clod .md are system linked so they are the same file. I do not, I never make a modification on one without it showing up on the other. Personally, I've used AGENTS.md as a convention because I find CLAUDE.md very, like, yeah, let's just use AGENTS.md. Let's not put the name of the company, but so every time I make changes on AGENTS.md, it automatically gets copied, if you will, to my CLAUDE.md. And if no one else had a question, I was curious about that load distraction and notification that came up on your screen.

38:26I use a tool called Rise. It's another startup that allows me to have a visibility with regards to, so far today, I've worked six hours, 16 minutes, focus, meetings, breaks, et cetera. This allows me to have tracking with regards to all of my, how I'm spending my time on average. I tend to work 10 to 12 hours a day and I need to have a good idea of where my time is going. So it's, in essence, a screen logger. I think it's like a hundred dollars a year or something like that. I'm not sure, but yeah, it's rise R -I -Z -E. And it's really useful for tracking and knowing where I'm spending my time on my computer.

39:16And unfortunately... Do you have the CPU eyes? No, he's pretty good at getting it performance monitor. I can show you that right now in terms of start monitoring. Yeah, you see it's using about 110 MB of memory, wake up per sec, 77. So yeah, you can actually, because it's a developer tool, more or less, you have access to every single thing that goes on and you can have like good telemetry on exactly what it's looking at. You can choose to do it cloud -based, et cetera. So, but yeah, it's, it's pretty, it's pretty light in terms of CPU and memory usage. There you go. Stop monitoring that. I don't care too much about the monitoring. So.

40:04Wonderful. I have a hard time with these types of tools because I'd rather lie to myself than know the truth. I have a bit of a meta question. The document you showed, the Sherpa document. I struggle from nowadays, I struggle a bit with the the the the the allergy that people tend to have developed for AI slope. And if you and rotten thing is that sometimes certain things that I like are seen as AI slope, like a sand color background. I like that. It's something that I find stylish.

40:56If you go through this document, there are situations where you have cards Yeah, where it shows cards. Well, the way it shows cards here and pills, for instance, they give it a feeling like this is really AI generated. Somehow, I'm struggling with the fact that my clients seem to have, it's like you didn't do the work or something. Um, if you like, let's say if you, if you give them a lousy old fashioned word document without any illustration in it, yeah, it is, it has the handmade feel, uh, where, where is this, uh, um, it's unclear what your, uh, what, what the, what the amount of personal stuff that's, that's in there, how do you look at that? I know that this is probably a bit beside the point. No, it isn't. It isn't.

42:03It's actually a very, so my answer to it has been I need these for myself oftentimes, and I will make short versions of documents for people. So here you see this is a short version of the document. And people have no problem with AI generated documents if it's only a page or two. Because one of the big push backs they have is when they receive a mountain of text from someone that clearly no human has ever reviewed. Well, if you choose to use that same process to make like a one or two page document. So this is like my security audit for what I do. And it's just like it summarizes everything. So I have read this, I have gone through this and really validated. it, okay, this makes sense with regards to doing it. So I tend to send the short documents to the individuals that are wanting to do pages. And have you never contemplated adding everything gone through my hands, Simon, at the end of the document to just to show that it really represents you and not the agent.

43:23I would never do that for documents because at that point it would be alive for me. Everything that I do at this point has been touched by AI at some point. And so I would feel disingenuous to say that I would do that for emails. And oftentimes you can feel when an email comes from me specifically, because A, I make it clear if it's Simon's agent or Simon, generally speaking. I'm getting, I have messed up in the past about that, but whenever I send something that is like from my own hands, it's going to be exceedingly six -cent and very short, and it's going to be like 10 words or 20 words. I personally think that I'd rather be judged on the quality of the documents I produce and the information in it. And the only way I can do that is if I have someone read it. And so what I, I used to send like 20 page documents to people when I was like among the first doing these like AI generated stuff. Now I'm actually more on the side of it doesn't really matter if AI generated or not, as long as it's extremely short and to the point and someone can look at it and just get the information right away. Um, it's, it's forces me to be very critical of what's actually the core message I'm trying to send. Uh, and that thought is what people feel and what makes them trust it more. So as long as it's extremely succinct and to the point, it's, it's, it flies very good with people and it lands good.

44:50Yeah. Okay. Yeah. Like this, for example, the playbook is not something I'd ever send to a client. It's meant for agents actually, it's meant for agents more because I know that humans are going to go through this and look at the pretty pictures, look at this, and then they're going to ignore this, which is completely fine. I made this for agents to, to read, and that's actually one of the things is make your, make your documents for agents to read. I have a lot of my clients right now that I send them something. They don't even open it. They just take that document, put it in their AI agent and tell, and tell the agent to summarize it for them and that's okay. That's completely fine. They trust their agent more than they do me and that's okay. I can use that to my advantage as well. Yeah, that's true.

45:39Okay, thanks. By the way, for anyone watching this, be careful when you do just put in a document into your AI. It's a very simple way to get prompt injected. FYI, very easy to do prompt injection through documents by putting in a bunch of texts that's white or a bunch of texts that will be machine readable but not human readable. So be cognizant of it that this is something, there are a lot of ways to prompt inject that are coming up, this is one of them. It's always better to just do a short document that you can just read. And I'm not sure whether it's a true story but I heard that about resumes. That is really good to include some live text in your resume. So when an agent looks at your resume, you do a prompt injection to put you on top.

46:33Yep. There is the same war that's happening between drivers and speeding cameras. It's happening between HR departments and resume writers right now. So we're in that world. Any other questions with regards to document generation, with regards to AI, with regards to anything? Since you said anything, have you ever played or worked with soft learning agents? In the sense of agents that are learning on the job or as it goes? Yeah, like using a lower model, a cheaper model to do a work, but then having the work evaluated by a more expensive model, and then also based on the result regenerating the instruction or code for that cheaper model.

47:30I have done some of that, but it's not a focus of mine right now. My focus This right now is on model ablation as much as possible in terms of, so to answer your question directly, anything to do with a smaller model and a larger model, you're going to end up, it's a lot of oversight in terms of you need to spin up multiple, multiple of the smaller models to get quality results when you're doing that. Um, it's, it's much better when, when you, I tend to do that once I have already figured out a process with a frontier model, a big model, like really well, and then I will just get it done by a smaller model and just review one in 10 by a larger model. Um, so it's, I tend to not see the benefit of having smaller models do the work when a bigger model is actually more token efficient, even though more expensive to just do the work. and then once it's really fine -tuned that work, then I offload it to smaller models.

48:36But I would be happy to hear what you're working on and see where, like, if I can provide advice or my opinion on that, I'd be happy to. I find it very fascinating to see what people are doing with models right now. Just to like finalize my little comment on ablated models right now, there's a lot of hacking happening and so I want my systems to be secure and I'm finding Claude and Codex just do not want to play nice when it comes to cybersecurity so I'm having to use open source and ablated so no guardrail models to to keep my infrastructure secure and make sure that they're doing the right that they're doing the pen testing more or less so that I know what to fix. Cool well then I will wrap it up there. We are at the end of an hour already. I am really looking forward to the next one. This was this one. And the next one will be watching your competitors and putting up little watcher agents that are going to go check out what all of your competitors are doing.

49:43I have a list of subjects that I'm planning on sending to the people who have been joining these so far to see what people are interested in in me covering but I'm having fun I think I'm getting better at them I hope by you know you know ten weeks I'll get even better at doing these webinars and we'll get to a point where it's useful and and really fast and good information I want to be tactical as much as possible so I'll let my agent evaluate your session yes Yes. Tell me what it says. Perfect. Well, thank you, Simon. Apparently I have a meeting coming up. Perfect. Thank you. I always appreciate it. See everyone. Bye.