GLM 5.3 AI Full COURSE 1 HOUR (Build & Automate Anything)

GLM 5.3 just dropped and it might be the most powerful open AI model ever released. But only if you set up the right way. This is the full GLM 5.3 course and by the end of it you'll be able to build and automate almost anything for your business. Any tool, website, or app you've ever wanted.

Any boring job you're tired of doing. You'll discover how to hand it to the AI and get it back finished. And I'm going to give you the exact prompts to type so you can just copy what I do. And this one prompting trick I'll show you that makes everything you build dramatically better.

10 little words most people never use them. Then at the end I'm going to show you how to make all of this run in the background so the work gets done whilst you're at the gym, whilst you're asleep, whilst you're living your life. If you stick with me to the end I promise you'll walk away being able to build things you thought you needed a whole team for. And you can start using GLM 5.3 today on a free trial.

Let's get into it. So let's start with what GLM 5.3 actually is because there's one detail here that changes how you think about every other AI model from now on. GLM 5.3 comes from a lab called ZAI. It launched on August 14th 2026 and here's the part that surprised people.

It's the exact same base model as GLM 5.2. Same size, same architecture, same pre-training. Nothing about the core model changed at all. Every single improvement came from post-training.

Let me explain what that means because this is the most useful idea in the whole course. Pre-training is where a model learns things. It reads huge amounts of text and builds up raw knowledge and raw problem solving power. Post-training is where a model learns how to work, how to plan, how to use tools, how to test its own answer, how to recover when something breaks, how to keep going on a job that takes hours instead of seconds.

So for years the assumption was a better AI meant better and bigger AI. But the parameters have not changed in this. Whereas most people used to think, oh more parameters, more pre-training. GLM 5.3 breaks that assumption completely.

Now if you want to use it, one of the easiest ways is to use it inside Zedcode. You can get a free trial and then just make sure you are using GLM 5.3. You can see that we've already built stuff here and it kind of works like cloud code. We'll get into that later.

So more about the model itself. ZedAI actually took the model they already had, spent about an extra month on post-training and got results that used to require a whole new generation of models. So if you think about it like this, GLM 5.2 was a smart student who only ever did homework. GLM 5.3 is that same student after a month-long internship doing real jobs.

It's the same brain but a completely different worker. And what they trained it on matters just as much as how long ZedAI deliberately moved away from coding puzzles and towards tasks that look like real units of professional work. Long tasks. Messy tasks.

Tasks with many steps where you have to remember what you did 10 steps ago. And some of their training tasks represent several days of work for an experienced engineer. In one example they describe the model gets dropped into the same working setup a real engineer would have with access to files, documents, past results and everything else. And a real problem to diagnose.

It has to figure out what's slow, fix it, run the experiment and prove the fix worked. That's the training and that made this model good at the thing you actually care about which is finishing a whole job on its own instead of handing it back to you in one finished piece. Now there's a second piece to how they did it and explains something you'll feel every time you prompt this model. So ZedAI couldn't hand build enough of these realistic training tasks so they actually built a factory that makes them.

So research agents took patterns from real work and turned them into readable environments with many steps and hidden prompts. Then a judge agent attempted each task to check it was even solvable. Then they built verifiers which are the graders without letting those graders see the answer key and that last detail is a clever bar. So a grader that can see the answer learns to grade against the answer instead of against the goal.

So by hiding the answer the grader had to check whether the goal was truly met which means that GLM 5.3 spent a month being judged on real outcomes not on whether its work looked right. And you feel that when you use it. So this model actually checks itself before it says it's done and when you write prompts that give it something to check against the quality jumps as well. I'll show you exactly how to do that in the first build.

Now let me run you through the numbers because the jump here was huge. There's a test called Terminal Bencher 3.0. It drops the AI into a computer gives it a real task and sees whether it can finish the whole thing alone. It gets up to 600 turns in a 10 hour window.

That's a long horizon task. Now GLM 5.2 scored 4.6 on it GLM 5.3 scored 28.3. That's around six times better in eight weeks on the same base model. There's another test called Deep SWE which measures how well the model fixes real software problems.

GLM 5.2 got 46.2. GLM 5.3 got 66.9. There's SWE Marathon as well which as the name suggests measures endurance on long problems. That went from 19.4 to 42.5 more than double.

And there's Automation Bench which measures how well the model can automate multi-step work. That went from 26.2 to 48.2 also nearly double. And there's Agents Last Exam a hard agent test which went from 23.8 to 28.5. So if you notice a pattern in those numbers the more and longer multi-step the test the bigger the jump.

So short test moved a little long test moves anonymously and that tells you exactly where this model is heading. It's really good for long horizon tasks. However ZAI actually ran their own tests on these benchmarks and published their own table so every lab does this and healthy skepticism is fair because labs grain their own homework all the time. So there's one row in the table they actually didn't run it's called GDP VAL and it was scored by an independent outfit called Artificial Analysis.

It measures real professional work the kind of tasks that people get hired to do. So GLM 5.3 scored 1,769. GLM 5.2 scored 1,508. Claude Fable 5 one of the strongest closed models on the planet scored 1,743.

GPT 5.6 Sol scored 1,730. So on the one test in that entire launch that ZAI did not touch the open model came out on top. That single row is worth more than the rest of the table combined and it's the strongest thing in the whole release. If you only remember one number from this course remember that one.

Now let me give you the other side because of course it only shows you the wins it's setting you up for disappointment. GLM 5.3 does not beat the top closed models at everything. On ZAI's own hardest internal test Claude Fable 5 still leads 39.5% against 34.5% at maximum effort. So ZAI printed that in their own blog post which honestly earned some sort of credit.

And there's another open model called Kimi K3 from a lab called Moonshop and that beats GLM 5.3 on several rows in ZAI's own table too. So on DeepSWE, Marathon, on a couple others Kimi K3's weights have been downloadable since late July and it's beating GLM 5.3 in many ways. So that's a quick background on what it is. How do you actually access it?

So you can go to zcode.ai you can download it here and this basically works like Claude Code if you're familiar with that. So you can see an example what it looks like but basically you can start a free trial. Now bear in mind that you don't get a crazy number of tokens so when you do use it for free it's great but just bear in mind like you may run out of tokens quickly. Just want to be 100% transparent with you if you're using the free plan.

Now once you've done that it's going to look something like this. So I've just zoomed in here just to make it easy for you to understand. So you want to make sure that if you are coding with this and you don't want to take any risks for example then you can just leave the permissions access to full access and you want to make sure you select GLM 5.3 in the drop down. So if we look at zcode for example and we pull this up you can see that we can select GLM 5.3 there and if you don't want to have to approve everything that it builds which honestly gets kind of annoying all the time then I would select full access.

If you do want it to plan first or to for example ask for permissions before it makes any changes if you're working for example on a system or website or an app or a tool where you don't want to take any risks and you don't want to make any edits then you can just ask before changes and then you approve everything first. Once you've done that you want to make sure that you select GLM 5.3. So there are older versions of GLM so for example you've got GLM 5 turbo you have GLM 5.2 those models are nowhere near as good you want to make sure that you select GLM 5.3 and then also when you're using this I tend to use effort levels max right so basically the more you the more intricate the project you're building and the better you want the outputs to be the higher the reasoning level you want to use now bear in mind the higher the effort and the higher the reasoning level the more tokens it's going to use up as well so it depends on you but basically I would recommend that you go with max for building and that's going to use up more tokens but you're going to get better outputs and things are going to be right first time around as well. So this is what the whole thing looks like now if you want to start coding with a new task you create a new task over here what you do need to do is make sure that you select a project now you can see here that you get three different options when you're coding with this so you can open up a new folder you can select a remote connection or you can work outside the project as well so if we open up a new folder like so we can select a folder locally that we're going to build with if we wanted a remote connection basically this can connect to docker or it can connect to ssh and basically this can access a remote workspace so if you want to build inside a remote workspace or for example if you want to connect to a vps or something like that you can go with that system i usually just code locally because it's simpler and it's easier and faster to use so we can open up a new folder like so and then you just give it the name of the folder now this can basically automate anything right so anything you want to build you can start with this if you've got a name of a website an app a tool if you for example have an idea of something that you want to build today even if you have for example an idea of a workflow that you are just really needing to get off your hands you want to automate you can just start here as well right it all leads back to this so basically you just describe in plain text what you want to build and you can go from there and so for example Well, if you have a look at this example here, we have said, build me, and I told you, I would tell you the exact prompt.

So I'm just going to read the exact prompt right here. We said basically, build me a one-page website for a dog grooming business. The goal of this page is to get visitors to book a call with me, include a strong headline, three short reasons to choose us, a section that answers the most common worry a new customer has, and a contact form that collects their details. Make it clean, modern, and fast on a phone screen.

Before you finish, test that the form actually saves submission. So you see how easy it is to start building a website. Now, if we look at version one, you can see that it's worked for seven minutes, 54. So it actually tells us how long it was working for.

Now, that means that you can basically type in the prompt and then you come back, right? So you go and make a coffee or whatever, if you want, or you can go and do some work in your browser or on a, you know, inside cloud or whatever. Then you come back in eight minutes and it's done. And you have the site right here.

So it explains where it's saved. So for example, it saved it inside this folder and then it explains what it did. So step-by-step, we've got the output here and then how we've tested it. Now, the reason that I'm not going to like build this right live is because obviously, if I was just waiting for eight minutes for this to code out, that would be very, very boring for you to watch as a course.

So I'm not just going to sit here and do that for you, but what I will do is show you how it looks. So we can actually open it inside our browser, or we can, for example, go back to said code here and preview on the right-hand side. So if you've ever used Codex, or if you've ever used, for example, Cloud Code, looks and feels very, very similar. The one thing that I will say is when you're coding out, for example, with zed code, the outputs you get are pretty nice, right?

And they don't look like AI-slop websites. So typically, for example, if you're building out a website with AI, it's going to look and it's going to feel sometimes like it's built with AI. But if you have a look at this, it's got a nice logo. It doesn't feel or look so much like AI.

And also, you can always go back and forth and change the outputs right here. So for example, if we open up this setting here, you can see that we've got the full website that we've created, and then it links to a book or call page, and it looks nicely designed and clean and everything else. Now, you might be saying, okay, which model is the best? Is it Fable 5, or is it, for example, zed code?

For me personally, I still prefer using Fable 5, but this is cheaper, and you can actually try it for free as well. So, you know, if you just want to get started and you're not using it for heavy stuff, then you can just go with this. So that is one example. That is how you can build out a website with this.

Here's another example, which we came back with here. So we actually gave it a goal, right? So it's quite interesting. You can use something called forward slash goal, and this is a new mode.

So I've shown you basically how to prompt very simply to build out a website. You just give it the prompt, you tell it what you want. If you don't like it, you tell it exactly what you want to change, and you can iterate and go back forth from there. What you can also do is you can start building out trackers or that sort of thing.

So we said goal, build me a weekly planning page, open every Monday morning. Shows seven days of the week. I can add tasks to any day, mark them done, and drag leftover tasks. So this is like creating an app, basically.

And it shows me how many tasks I have finished this week. So we're creating like a dashboard. It works for 39 seconds, and it built it out. And then what it did after that is we gave it another task and just improved it slightly on the prompt, and it worked for 18 minutes autonomously.

If we want to see like a deeper breakdown of what it did in those 18 minutes, we can just click on this, right? So we click on the dropdown, and then it shows us everything right here. And then as we scroll down, we can see everything that we've built. Now, what you're going to see is that we can actually open and see what was created, and it will give you a local file address right here inside the browser.

So if you want to open this up inside your main browser, you can actually just click and open up like so, right? So that is the local file address, and you can view that inside your main browser here. Now, bear in mind, this is not stored on an actual website yet, but if you did want to host the website or the app or the tool that you built with a actual website, then what you can do is get a Netlify website, right? So you can go to netlify.com, and from here, you can log in and get a personal access token and you give that personal access token to Zedcode directly.

So I've shown you how to build websites, and I've shown you how to use a goal. Goal mode, if you're wondering, right? Basically, you set up a goal and you say forward slash goal, do this, and it will just loop around and check the work is done until it finally completes. So it's a way of like getting your agents to just keep going round and round and round, and you have one agent that builds for you and one agent that judges a work.

So the agent that judges a work actually checks if the work is done or not, and if it's up to standard. If it's not up to standard, it keeps looping round with the goal until the work is judged as complete. Now, here's another example of a website. We can actually build, we built this website in like a different language.

So you said build a website for Casper Dash in Chinese and actually build out a website in full Chinese, as you can see right here. And then we've got this language option in the top right where we can click between English and Chinese, right? So you can see that we can switch between them right here. So just to recap, we've covered how to build projects, how to use the goal mode, how to build websites in different languages.

Also, what we can do here is we have this automation section. So we can create a scheduled task, and the great thing about with a scheduled task is that we can basically set up an automation to run daily without us. So let's say, for example, you wanted to automate searching for the latest AI news. Well, we can just have this running and you have to make sure that, number one, you have the app open, and number two, that it's running on a computer.

So when you are running this automation day-to-day, if you set up a scheduled task, it's kind of like OpenClore or Hermes Agent, if you've ever heard that before. If you don't, it's like having an assistant just sign in on a daily basis, and then on a daily basis, it runs his task. So we can set up an automation, and we can say, for example, research the latest daily AI automation news, and then from here, we can select the schedule. So we can say, okay, every week, we're gonna do this on Monday at 9 a.m., and then from here, we can say, every weekday on Monday, make sure that you check and research the latest AI automation news, and then give me a bunch of content ideas that I can create content around based on what's just dropped this week.

I want you to focus specifically on content that has dropped this week and new ideas this week, not old stuff, so make sure you actually check that the content is new, and then also what I want you to do is make sure you put that into a beautiful webpage so that I can actually view the content and see what it's like. So from here, we have the instructions. Now, by the way, when I'm creating instructions for anything complex like this, I like to just create a voice note. If you're wondering what I use to create the voice notes, I will use something like WhisperFlow, right?

So WhisperFlow, this is what it looks like, and that way I can give detailed prompts without having to just sit there typing away, which is super energy and time consuming. So from here, we've given it the instructions, we've got the schedule, we've got the task title, and then we can select the project. So where do we want it to work on? Now, I'm just gonna have that work outside a project, and also, bear in mind, this is a scheduled task.

So for me, if it's running without me, I need it to be fully autonomous. If I want it to be fully autonomous, I'm gonna set the permissions as full access. Then from here, we can choose the model that it runs on. Again, you wanna select GLM 5.3, and then in terms of the effort level, I would go with max.

From here, we can create the scheduled task, and boom, shakalaka, we now have a task that will run on a schedule every single Monday at 9 a.m. If we ever wanna edit that, we can just click on the three buttons, and we can edit the scheduled task, and we can also do a test run now. So if we click on run now, we can actually see that that is running inside our app with GLM 5.3, and so we can see inside our tasks all the previous tasks and the history of what we've done. We can pin anything if we wanna come back to it later.

We can unpin it like that if we don't want it to be pinned. So we can see all the runs of our previous tasks, including scheduled tasks, inside this section right here, and then you can see it's thinking, so it's looking at the latest AI automation news. We can see that it's planning out how it's gonna do it, et cetera, and this is basically a scheduled task right here. For me personally, what I like to do with scheduled tasks is I will keep the app running, and then I will have the task scheduled, and also I just have my computer running, right?

So I have a Mac Studio, it's pretty much always on, and that's just running in the background so that any scheduled tasks actually run. If you close the app, won't work. If you shut down your computer, it's not gonna work. If you don't have access to the internet when you schedule a task, not gonna work, right?

So you have to make sure that it has all the connections that it needs to be able to access this whenever it needs it. So those are automations, and those are tasks. What we also have over here is the mobile remote control section. So we can actually scan this QR code here, and what it would do is allow us to connect this to our phone.

We can also copy this link and share it with people, and we can basically allow our phone with the app to control our code over here. So this means that you don't just code from your computer, but as long as you have this open, this whole ecosystem open, like I was talking about before, then you can actually scan the QR code on your phone and connect and control your computer phone using Z code as well, which is pretty nice. And also, you can use a different channel here. So you can connect this to Telegram.

How'd you do that? You just click on that, right? You can connect this to Lark. How'd you do that?

You just do this, right? How'd you connect it to FaceShoot? You just go. How'd you connect it to WeChat?

You just go over here. A lot of these apps are Chinese because ZAI is a Chinese-based company, but that's how you can connect it. For most people watching this, probably Telegram is gonna be the most useful, and also this QR code right here. And then you can actually stop that if you want to as well.

You can also refresh the QR code as well, right? So when we're using this, for example, if I've just shown this on a setup like this, I can refresh it like so, and that way we've refreshed it, and no one can get access to the old one. Now, we've also got a bunch of settings. So what you're going to see over here is that we have a bunch of skills, so we have specific skills that we can do cool stuff with.

Now, skills are basically workflows where we train the agent on workflows that we're going to use. So for me, for example, I actually have a skill for SEO. It's inside our AI platform boardroom. And what this allows me to do is basically create nice content for AI SEO.

So, for example, I have a bunch of SEO projects I work on, and I use AI automation to help me rank my websites and get more traffic. Here's an example of a website that's grown with AI automation. I've never logged into this website, but I can get AI to control it. How do I get AI to control it and post content there?

I use this skill. Now, how do we give a skill to the AI agent? We go inside the skill section here, and we click on the plus, and then you can just paste in the skill or train it on the skill that you want to use. And you would just say, hey, do this.

So we can go inside here, we can create a skill, and we're going to go to a new task, I'm going to paste that in. And I'm going to say, add this as a skill so that you can create SEO content to my websites. And basically, this is a saved workflow. So if we have a look at the default skills that I've got inside my setups here, so if I click on settings over here, I've got skills here.

These are skills where we can basically view custom workflows that were saved for later. So if we have a look at this skill, it's got the name, 2D Games. It's got the description. It's got the allow tools.

Then it has all the details of how that skill works. And these are very detailed instructions that basically teach any AI agent. It could be, for example, KimiK3, the skill of how to do something so that we can recall it later. Why do we want to do that?

Because it's basically a way of saving a workflow for later. So for example, I can say, use your guide skill to create a guide. Use your mobile design skill to create a mobile app. And these are skills we can save.

We can switch them off, we can delete them, we can edit them, we can add new ones, we can download our existing skills, we can refresh them. And also we can import skills. So you can see here that if you have, for example, saved workflows inside Cloud Code, you can import the skills from Cloud Code into said code. You can import the skills into Cloud CLI, into Codex, into OpenCode, into OpenCore, whatever you want.

This is pretty nice. Then we also have sub-agents over here. So these are different sub-agents that are built into the system. We also have plugins.

Plugins are super good because basically what this allows us to do is connect to external tools. So for example, we've got document skills here. We have different skills that we can add and create. We can also add new skills here.

So what this allows us to do is connect to external apps so that we can control apps with our AI agents. So for example, if we have a look at this here, we can actually create or we can add new apps. We can pull in an existing GitHub repo for an application that we want to add. Or we can connect directly to these existing ones.

So if we have a look here, for example, we have like Context 7. We can install the app for that. We could install the plugin for cloud-based skills for GitHub, for iOS simulator, browser use. What browser use means is this can actually control our computer so it can control our browser.

It can basically automate and build whatever we want with it. And also it can control everything inside our browser so that we can basically use AI to control our browser. Honestly, for me, in a practical sense, do I like browser use? Not particularly.

I don't think it's particularly good. And that's why I don't want to go into it too much in this course, simply because I'm not going to teach you stuff that you don't find useful, right? Like I could teach you how to use browser use, but I just think with any AI model, whether that's Claude, whether that's Zcode, whatever it is, it's pretty bad. Like it's not that useful.

As impressive and fun as it sounds. Then we also have memory over here. So we can save our memories for later. So if you toggle this on, this allows you to save memories for later.

So essentially this means that if you don't want to re-explain yourself and you want to give a lot of context and personalization to an AI agent, it can save memories inside this section. So you just toggle it on. For me personally, I actually prefer to use something called Obsidian. Obsidian is a really good external memory that you can plug into all of your AI agents.

Why do I want to use something like that? Because basically this gives context to all of our AI agents, not just Zcode. So you can use the memory that's inside here. For me personally, I prefer to use Obsidian.

If you're wondering what Obsidian is, it's a free app that stores all your information. So for example, if you click on this button, you can see it has information about our AI Profit Boarding Community, how it works, what it is, how to use it, etc. And that connects to all of our agents. So for example, if we have a look at our agentic operating system over here, all of our agents pull from our memory, and all of our agents update our memory system here as well.

Why is that useful? It's useful because basically our AI agents have a lot of context on who we are and what we do. How do you get your agents to use it? You can just go into Zcode, make sure you have Obsidian set up, and you can say, based on my Obsidian memory locally, can you make sure that you use that for your memories and also context?

So let's do a little test right now. Can you just tell me what happened recently using my Obsidian memory and database? It's stored locally. So from here, it can start to work with our Obsidian memory, which is just a folder and a vault.

So it says, I'll look for your Obsidian vault on your system, and it's going to ask for permission to access it. It's found it straight away, which is fantastic. And now we instantly have a memory system that we can plug into Zcode. And then you can see here, it says, your vault is very well organized.

So thank you for the praise there. And it's got daily notes, it's got agent memories, it's got all of our digests and all of our notes. And then basically what it's doing here is it's grabbing the information from our memory system and using that for context. Why is that useful?

Well, for anything that you want to automate, when you're using a generic, blank, out-of-the-box agent, it doesn't know anything about it. So how do you give it context? You can use Obsidian. What is Obsidian?

Obsidian is just a free app with notes and files about you. How do you plug it into Obsidian? You just literally say, here's what to do. And then anytime you have a workflow that requires personal information about you, you can say, hey, using my Obsidian vault, can you just use that for information so that you can give me a personalized answer?

So for example, I can say, using my Obsidian vault, can you give me some SEO keyword ideas? And you can apply this to whatever you want. I'm an SEO, so for me personally, I would like to use this for SEO. So I'm going to say, using my Obsidian vault, can you give me some SEO keyword ideas?

For you, it could be like, using my Obsidian vault, can you tell me what we did yesterday? Using my Obsidian vault, what should I automate with GLM 5.3? And so it's a powerful second brain that doesn't just plug into GLM 5.3, but plugs into all your agents. It gives GLM 5.3 context from day one.

It's a system that no matter what model comes out or whatever system comes out next day or next month or next week, you have the Obsidian memory vault to plug into it. And also it's just a very detailed way of personalizing it. Bear in mind as well, when we look at our Obsidian vault, it looks complex, but you can actually tell your agents to update it for you. So when I'm using any AI agent, Claude, Hermes, OpenClaude, whatever it is, GLM 5.3, I will tell it to update all of my information inside the Obsidian vault.

And that means that all of these notes are automatically updated and linked together beautifully in a very organized way without me having to do it myself, which saves hours of admin. And then also for basically recalling that context, the agent can use it and any other agent can use it as well. And then if we have a look here, you can see that it's actually pulled in the latest information. So it's looked at our Obsidian vault, it's looked at the impressions we're getting for those keywords, and it's pulled in keywords.

So for example, Hermes workspace, that is something that I use that will be a great keyword for my website. And so it's found a bunch of new, relevant keyword ideas that we could basically plug into our SEO system. Look at that. These are really, really personalized ideas.

And I didn't have to start from scratch. So now I have a personalized GLM 5.3 system that has skills and workflow saved into it, that has a memory plugged into it as well, that connects to my plugins and also has automations running as well. And also I've shown you how to build a website and an app with it already, which is fantastic. So we're making a lot of good progress here.

So let's talk more about GLM 5.3 and how to build with it even better. So build number one, a lead capture page for your business. So this is a high value first project because a lead capture page is the front door for new customers, right? It's the thing that turns a visitor into a name in your inbox.

You can set your effort levels to max, and you can see an example of how this would work in practice right here. So you can set your effort level to max, open a fresh folder for a project, and you can type in this prompt like you can see right here. So we've already talked about this, but this is the dog groom website example. And then when we pressed enter, and this is actually it filmed live, so you can see how it works.

So it doesn't start typing code straight away. It plans first, you see it break your requests into a list of steps, and then it works through them one at a time. And then at the end, it goes back and checks its own work before telling you it's done. And that last step is the trained behavior I described earlier.

So it learned inside environments where its work got graded against a real goal. So when your prompt includes a line like Tessa, the form actually saves, you're speaking its native language. And that one sentence measurably improves what you get back. So when it finishes, you can open the project folder, you can double click the main file, you browse, your page opens in the browser, and it goes from there.

Now, a couple of prompting lessons. that I would recommend for you. Lesson number one is you want to describe outcomes, not steps, when you're prompting this. So if you look at that prompt from before, it said what the page is for, right?

Getting visitors to book a call. It didn't say, for example, create a file called index, then add a header tag, blah, blah, blah, right? The model was trained to plan. So planning is a thing that got, like, way better with this model.

So when you micromanage every single step with it, you're using a long horizon model, like it's a simple autocomplete tool, and you throw away the entire upgrade. So like a bad prompt when you're using this system would be make a header that says docrooming in blue, then underneath make three boxes. A good prompt would be like make a page that convinces a nervous first-time customer to book and looks trustworthy. That second level of detail gets you a better page because the model can make a hundred small decisions in service of a goal it understands.

Now, lesson number two is always give away for it to check itself. So you can say, for example, test that every link works before you finish. Make sure it reads well on a phone. Check that the numbers add up before showing me.

Don't finish until saving actually works, right? And when you have these check-in lines that tell it exactly what to do, basically these are just 10 extra words, but it's a big single quality upgrade because it makes it check its own work and improve it, right? And there's a reason this works so well in this specific model, rather than being generic advice. So Zed AI's whole post-training approach was built on verifiers on graders to check whether the goal was met without seeing the answer.

So when you hand the model a criteria to verify against, you're switching it into the mode it was trained for. Quick pause here, by the way, if you want more training on this sort of stuff, feel free to check out the AI Profitable Boardroom. Link in the comments description or go to the AIProfitableBoard.com. Every single day, I answer your questions.

I also create video tutorials for the members there to help you as much as I can. And then also inside the classroom, you get access to all of our best training. So you can get our full agentic operating system that I showed you before that has all of your agents plugged in and we update it all the time. Also, you can get access to all of my new trainings based on what's actually working.

Inside the classroom, you can also get access to all of my other trainings. And then additionally in the calendar, you can jump on four weekly coaching calls about AI automation to get help and support in real time, share your screen, meet other cool members. And then in the map, you can meet people in your local area who are building out with AI agents like you. So feel free to get this, link in the comments description or go to the AIprofitable.com.

So build number two, and this can be like a client tracker. And this build basically teaches you goal mode, which is a really important feature in the entire app. So let me explain the difference first, because this is where most people are leaving the biggest gains on the table, right? Normal chat mode works like a conversation.

So you ask it for something, it does it, you look at it, you spot a problem, you ask again, back and forth, and you're the one holding the plan in your head and you're the one catching the mistake. Now goal mode actually works differently because you hand it a finish line instead of an instruction, and it plans to work, builds it, tests it, finds some mistakes, fixes them, tests again, and keeps looping until the goal is actually verified as met. So it's not like working until it thinks it's probably done, it's working until it proves it can be done. And this is the feature that really helps you get the most out of everything that I've told you inside this training.

So if we look at that terminal bench score before, the terminal bench that went from 4.6 to 28.3, measuring precisely this skill, right? Staying on track for a long job with many steps without losing the plot. So the goal mode is where you get to use that. Now, if you're wondering, okay, what is the prompt?

Let me show an example. So in this example, you can type forward slash goal, and then when you do forward slash goal, you just click on it like so, and then it selects that as the goal system. And then you can say, for example, a working client tracker that can open up in my browser. I need to be able to add a client with a name, what work I'm doing for them, the deadline, the current status.

I need to update the status and delete finished clients. Everything shows in one clean dashboard, blah, blah, blah. And that's the sort of stuff that you can use. Then you can walk away, you can come back to the finished work, and it just basically works on a repeating loop until the job is finally completed.

Now, I was talking about the schedule tasks before as well. As you can see, that just ran in the background. So one cool thing that you can do with GLM 5.3 is you can have multiple different projects working in the background at the same time in parallel. So you can have a team of agents getting stuff done instead of just one at a time.

So if we have a look at this, for example, we looked at that schedule task and we actually ran it earlier as a test, and you can see that it came up with the research and then gave it back to us. Now, it's actually found relevant headlines that have just dropped. So you can see here, it's found the headlines for the AR automation news that was scheduled in. So for example, like August 13th, August 15th, August 11th, it's actually given us the dates of the stuff that just dropped.

Now, if we actually have a look, it created the page as well. So we've got the headlines here. With the breakdowns, we have a full list of all of the news headlines from this week, and it's basically ordered it as well. And this page looks super nice as well.

So it's created a website using the schedule tasks that we did before. And that's just another example of how good this stuff is. Now, when you actually look at this, you probably notice like the web design is not amazing. It's not gonna blow your socks off, right?

Now, if you have a look, for example, at the systems that we have inside our website here, what we've actually done is we've got a skill that organizes everything that we create, all our guides for all our tutorials into a beautiful guide like you can see right here. And so this is using a skill. And if you're, for example, building our websites or apps or anything like that, that's why you wanna have these skills where you tweak and you tweak and you give feedback to your AI agents, and then eventually they create something awesome. So if we have a look, for example, inside this section, we have all of these skills that we've created, and then we can just make sure that when we are using any of these skills, they're detailed, we have a look at the skill created, and it's just ready to go.

The reason that I recommend that is it's just gonna save you so much more time and it's gonna be way more useful and powerful. So for example, when we created that guide that I just showed you about DeepSeek Harness, which you can see right here, it looks beautiful, it's super nicely created, right? Now, when you're looking at this, you might be wondering, okay, how do you call the skill? So you would just say like use your guide skill to create a guide on this, and then it will pull in the guide that we've got right here, right?

And then you can see that's fully done. And you might be wondering, okay, why am I showing that inside Cloud? Because that's typically the agent that I use, but Zedcode works exactly the same. You just say, use your skill for this, use your skill for that, and then you've got everything custom trained inside one place.

Now, one thing that I also like to do with any of these sort of models when they come out is I like to plug it inside an agentic operating system. And this is like the highest level of using something like GLAM 5.3 or any sort of AI agent. So if you're using an app like Zedcode, the more AI systems you use, the more powerful you can become with AI. But the problem is like if you've got, for example, Zedcode in one app, you've got Cloud Code in another app, you've got Obsidian inside another app, and all your agents, even chat jeopardy in a browser tab or whatever, is it becomes super messy, right?

Because then you have all of these different apps that are not linked together, and they're all scattered around, and you're switching from one tab to another tab, you're constantly context switching, and also it's very time consuming to work that way. Whereas what I like to do is I like to have a mission control like you can see right here. And this way we have all of our agents and all of our agentic systems inside one place that we can come back to at any time. So for example, we have the memory system that I was talking about before, but it's visualized and it's beautiful.

And I can see all of the memories we've created over here and the graph here. Now, the great thing about this is like, bear in mind, GLAM 5.5 will come out, then GLAM 6, and then Cloud Fable 5.5 will probably come out one day. Then we'll have, you know, GPC 6. And so the problem is if you're like chasing all these models, you know, it's very hard to keep up with Zedcode.

They're constantly creating amazing new models. So how do you get around that? How do you avoid shiny object syndrome and stay focused and also build something that lasts instead of something that just breaks on every single system? So what I like to do is I like to have this agentic operating system with Claude, with Hermes, with anti-gravity.

And then when I'm using something like GLAM 5.3, I will plug it into this system right here. So we can select GLAM over here, and then we have Claude code inside our agentic operating system. And we have GLAM and the brain plugged in over here, right? So we can use it whenever we want.

And so that way, everything, number one, is synced into our memory system. And number two, everything is just one click away. I don't have to switch tabs, don't have to switch apps, don't have to think about anything. Everything is just ready to go right there.

And I like that much better because it's going to save you time and everything's needless. Now, how do you build something like this? Literally all you do is you would go into a new task site. So you'd be like, build me a mission control dashboard where I can control Z code and all my other agents in one place.

Start with Z code and just make it like a beautiful mission control. Use the CLI inside Z code, and also have a section for the obsidian memory as well inside the dashboard. Now that will take about 20 minutes to build out. So you probably want to get a preview of that right now, but that's how you can begin to build the dashboard.

If you just want to get my system with everything ready to build, because I've spent like three to four hours a day building something like this, you can get that inside the AI Profit Boardroom, link in the comments description, or just go to the AIprofitboardroom.com. But that is like the top, top level for using agentic operating systems, right? For becoming powerful with AI. And the great thing about that is you got the memory system, you got your agents inside one place, it's super productive.

And then also, for example, we can build custom workflows. So we can have a video agent over here, we can have an SEO agent over here. And this is literally just like a nice UI for the SEO skills you saw before. So any sort of skills or any sort of workflows you run, like if we have a look, for example, at this automation.

this AI automation news workflow, super boring, right? Super boring inside a little workflow like this. Doesn't look nice, isn't fun to come back to. We can check the conversations, but it's kind of hard to switch between them, right?

Whereas for example, if we have an agentic operating system like this, well, we have the automation for news over here, and this just runs on a schedule. So I don't have to click this, I don't have to manually open it up, but I've got this beautiful UI for the same automation. I've got all my history of the automation runs here, so I can flick between the different days and when it ran and what it did and everything else. And then also we have all of our things linked together.

So for example, I was talking about the SEO content workflow before. So I can link the automation for finding news to the automation for creating SEO content in one button. So if we click on publish to WordPress, now I actually publish an article directly to my WordPress on that topic, which is amazing. Right, that's just gonna save so much time.

So for example, if we have a look at DeepSea V4 Pro Crush's Fable 5, if I want to publish an article on that, I just click this button and it will start publishing content directly to my WordPress website. How amazing is that, right? That's like the most advanced layer for that. And it goes outside the scope of this course, but I just want to show you what's possible because that's the top layer.

So the bottom layer is like you create a basic website. Then you can, for example, create apps and that sort of thing. You can create weekly planners, you can add skills, you can add automations. You can, for example, use long horizon tasks with the forward slash goal mode.

You can have your agents running inside multiple conversations. And then you can build a UI around this that runs locally, and you can have all your automations inside one place. One other thing that I want to show you with a GLM, and this is pretty cool, is that you can plug it into a different harness. So a harness is just the hands and the body for an AI brain.

So if you think about it, for example, GLM 5.3, itself is a model, and that is a brain behind the actual system. Then we have the app itself. So if we look, for example, Z code is a harness, this course focused on mostly. Hermes is another agentic harness, and you can plug in your coding plan and GLM 5.3 or GLM 5.2 directly into a harness like Hermes.

What does that mean? That means that you get an agent that you can basically run with the brain. And the difference between harnesses is the way that the information is presented. So for example, here, you'll see that we've got bot mode with Hermes, and this allows us to create different team members with GLM 5.3.

So we can click on new bot here. We can create a new bot like a researcher, give it details of what its job is, whether it's active and everything else. And then we can switch between these different chats here, and it's like using GLM 5.3, but the difference is that it feels and it looks like WhatsApp. So for example, our content writer, and we can just message it like so.

And it looks and feels like WhatsApp. We've got all our chats here. I think that's pretty amazing because you can basically use GLM 5.3, but you can use it like WhatsApp, or for example, Telegram, and you just have your different chats organized in one place. So that's really cool as well.

And then for example, you can even build like voice agents. You can use Hermes Oracle for grabbing the latest news, and then you can publish content to WordPress. We've got the articles that we've published over here, as you can see, and it's all ready to go inside one place. That's amazing, I think.

And then also, for example, you can give GLM 5.3 access to different APIs. So for example, if you wanted it to do outreach for you, you could give any API keys that you have, like for example, Hunter. Hunter API can generate leads for you. You can give that API to GLM 5.3.

You can just plug it inside the chat here, and then you can say, okay, generate a list of leads for me or whatever. So there's some really, really cool stuff that you can do with this. So that's a lot of the system right here. Here's some more examples of what you can build with this.

So we have a build where basically we took our customer feedback. So you can gather whatever you've got sitting around, like support emails, reviews, survey answers, notes from calls, and put it into one folder locally. And then from here, you can use a prompt like this, so read every piece of customer feedback in this folder, group the feedback into themes, count how many times each theme appears, and then build a single webpage that shows the top complaints, ranked by how often they come up with the top three quotes underneath each theme so I can see the real words customers used. Also list the top three things that customers praise.

Don't finish until each piece of feedback has been read. And then you can see here that it's basically taken that information, started building with it. We're using GLM 5.3 on Macs with full access. And then it creates a beautiful webpage like this that basically analyzed all of the top things that our customers gave us in terms of feedback, right?

And so it's really good for actually taking the data that you have, analyzing and organizing it in a way where it's very easy to read and understand everything. Another example here is a planner built around your week. So you could build a weekly planner with this. So we said, go build me a weekly planning page.

I open up every single Monday morning. It shows the seven days of the week. I can add tasks to any day. Mark them as done.

Drag leftover tasks to the next day. Shows me how many tasks I finished this week compared to last week. Everything's safe so it's still there. Next Monday, keep the design calm and simple, plenty of space, nothing cluttered.

Don't finish it until adding, moving, completing and saving will work completely. So it started off coding with this. We've got the left-hand side with the chat, the right-hand side with the preview. And then it actually built out the weekly planner as you can see right here, right?

So it can create some really useful, actionable stuff. You can even, for example, build out apps. So this is an app that we built out where basically we can add customer jobs, customer details and everything else. How did we do that?

So basically we said, you know, goal equals build me a mobile app. I can add to my phone, home screen and use offline. It's a job tracker for my window cleaning business. And you can customize this to whatever you want, right?

So on the main screen, I see today's jobs with customer name, details, et cetera. I can tap a job to market complete. I can add a new job. And that basically went off and built us a mobile app.

As you can see right here, that actually works. It saves data. We can store it on our phone. It's ready to go.

So apps, weekly planners, websites, data analysis, the Zed code is very good at all of these different things. And also if you create an app, you will basically, what you can do is you can say like, give me the simplest way to open this on my phone from my computer in plain steps so I can follow without any technical knowledge. It will give you a local address. You open that address on your phone, you add it to your iPhone and then you name it, tap it, and then it can just run locally as well, right?

Now, that would not be for like a phone on the, you know, an app on the app store, but just for like building an app that runs locally on your phone, which is pretty nice as well. Also, I want to teach you a few things about like the tokens inside the system as well, because there are three things stacked here that most people never notice and together they can roughly double like what you do. So the first is caching. So Zed code actually reports a cache hit rate above 98%.

Caching means that when the model rereads context, it has already seen, like for example, your projects, your files, your second, third and 10th message, that repeatedly reading counts for less against your quota than reading it fresh. So Zed AI actually puts the practical effect at around 30% more effective tokens. That means that you get more tokens out of what you're doing. So how do you take advantage of that?

Well, if you stay inside one project and keep talking, so if you're like, if you've got a project open like this, as you can see right here, then you can basically stay inside that project and keep talking rather than starting fresh projects constantly. So a long conversation inside one project is cheaper per message than like 10 short conversations across 10 projects because of the cache hit rate. And that's the opposite of what most people assume, right? Second cool thing, and most people don't realize this as well, is the off-peak rate.

So basically usage outside peak hours counts as half rate against your quota. So if you want to make sure that you don't hit token limits when you're coding with this, their peak hours run on China time with GLM 5.3. And so weekday afternoons over there, which is a fairly narrow window. But if you're in the US, the UK, or most of Europe, the vast majority of your normal working day already lands in like the window that's half rate, right?

Without you doing anything at all. So you basically double how many tokens you get if you're like in the US or Europe and you're coding with this. So if you happen to be working during that peak window and you've got a big job that isn't urgent, you probably want to schedule it for later instead because then you can get double the amount of tokens. And the third is the model's own efficiency.

So basically fewer tokens per task, then comparable models by quite a wide margin. So when you stack the caching, the off-peak rates, the efficiency, et cetera, you can get way more tokens out of the process as well. And you can be way more efficient with it as well. Now, also you can swap the engine.

So like I was talking about before, if you already use Cloud Code, GLM 5.3 can plug straight into it. So Zed AI built their own endpoint to be compatible with Anthropx format. So you can actually point Cloud Code at GLM 5.3 with a few settings and no special software, right? Same with OpenCode, which had GLM 5.3 live on launch day with the 4 million token context window.

So if you spent like months getting comfortable inside one of these tools, you don't have to throw that away to try this. Like you can swap the engine and keep the habits. And that's also why people test these models at all, right? Switching takes about 90 seconds.

So trying a new model costs you almost nothing. And that compatibility also opens up the part of this that most people never reach, which is, you know, setting up skills and setting up automation. So you stop typing the same thing over and over again. So if you create a plain text file in your project folder called instructions, you can write down the rule you want followed on every project and job in that project, your brand colors, your tone of voice, your rule.

Every page, for example, must load fast on mobile. And you rule that maybe nothing gets marked finished until it's been tested. Then you can start every prompt in that project with read the instructions file first. And that's a skill, not a fancy one, but that's just an example of what you can do right here.

Also, you might wonder, okay, where do I start with this sort of stuff? You know, what should you actually automate with Zcode? You know, what are the best things that you can get started with on this? So I wouldn't recommend like automating the interesting work or new ideas that you have.

I would automate the work you'd describe as, I have to do this again every single day, right? So anything you've typed more than five times can become a skill file. Anything that you do at the same time every single week is a scheduled jobs. Then that way, the boring repeated stuff is where you get the most value from this because that's the stuff that's eating your week and your time up.

Now, I also want to cover some mistakes that I see people make with these models because avoiding these will put you ahead of most people. So for example, mistake number one is prompting it like a chat model, like short, vague, one-line requests. This model was trained on jobs, not questions. You can give it a job.

Mistake number two is leaving the effort on low for builds that need something much better. So you might think, ah, save tokens, I'll use it for low effort, but you'll just get fast, shallow output and then blame the model. So you won't max for like anything, the building that's actually like complex or sophisticated. Mistake number three, not giving it context.

So you definitely want to build that memory system that I talked about before. It will just save you so much time. And also, you know, when you give it examples of what you're actually doing, you can use that context to personalize everything to you. Also, mistake number four is no verification line.

So when you're coding with this sort of stuff, you want to make sure that every real prompt ends with something the model can check itself against. Mistake number five is starting over instead of iterating. So when something is like 80%, right, just give it feedback on how it can improve, what it can do better, what it needs to do next, right? That's way better than like starting over and over again, which some people do, because they're like, oh, I just need to start again and, you know, start from scratch.

But it's like, usually you just one or two steps from completing the project. Also, bear in mind, this doesn't create images as well. So GLM 5.3 doesn't create images for you. Just one thing to know there.

You might also like wonder when should you use this? So I would recommend that you use this basically as a sub-agent. I don't think it should be your main agent. It's a great model.

I mean, it's a fantastic model, but I think it's better for like sub-agent work. So for me, for example, personally, I'll get Claude to orchestrate it. And then the great thing about that is like, I can have Claude control GLM 5.3 for the tasks where I need to use a lot of tokens, but I don't want to max out my Fable 5 limits. And that's a great way to use it.

And you can just go to Claude and say, hey, use GLM 5.3 for this or for that. And that saves a lot of time as well. And yeah, that's basically it for GLM 5.3. That's the full course.

So just to recap, we've covered everything you need to know about the model itself. We've covered how to build it with agentic operating systems. We've covered skills, automations, tasks, how to build and automate anything, the best ways to use it, how to get Claude to orchestrate it, how to plug it into Claude as well. And then also how to use it with the goal mode.

So we've pretty much covered everything that you need to know about this sort of stuff. If you want to get more training on agents and AI automation in general, feel free to check out the AI Profitable Boardroom. Link in the comments description or go to the AIProfitableBoardroom.com. Inside the community, you can ask questions to get help and support in real time.

Loads of great members inside there. There's always someone online ready to help you whenever you need. You can also direct message me personally as well so you can get help and support whenever you need as well, which is great. And then also inside the community, you can meet cool people.

You know, it's a really great way to network and that sort of thing. Inside the classroom, you can get access to all of our training. So we've got new daily updates over here. We've got our agent OS system and you can also grab the agent operating system that I talked about before that orchestrates your agents.

You can see all of our new training. So we cover pretty much everything that's useful that you can use here. And then inside the calendar, you can jump on weekly coaching calls, get help and support in real time, ask questions, share your screen. Inside the map, you can meet people in your local area who are building out with AI agents like you.

And this is all available inside the AI Profitable Boardroom. Link in the comments description or go to the AIProfitableBoardroom.com. So that is how to build and automate anything with GLM 5.3. Powerful model.

And the other thing that I would say here is it's just getting better and better all the time. If you saw the first version of GLM last year, not very good at all, but it's improving at a rapid rate. And I've noticed that with all the Chinese open source models. So who knows where this goes?

The other thing that you'll find is GLM 5.3 is open. So you can get the open weights for it, but you can't run this locally. It's way too big as a model to run locally. So if you're watching this and thinking, oh, I'll run it locally and that sort of thing, you need a ridiculous number of GPUs.

So it's always better to run it with the API, cloud hosted. It's going to run faster. It's going to run smoother. And that's much more reliable for most people watching this.

So thanks for watching. I'll see you in the next one. Cheers, bye-bye.