Notion Workers – Full Tutorial 2026

[00:00] Learn the revolutionary potential of [00:01] notion workers. This course provides a [00:04] comprehensive guide on how to leverage [00:07] this powerful feature to automate tasks, [00:10] connect external data, and build complex [00:13] business logic within your workspace. By [00:16] the end, you'll have successfully [00:17] deployed your own custom workers, [00:20] regardless of your programming [00:22] background. Notion recently released one [00:25] of its biggest updates ever, but very [00:28] few people can actually use it. I am [00:31] talking about workers. Basically, [00:34] workers are the missing puzzle piece in [00:36] notion. They allow you to do practically [00:39] anything. Bring data from external [00:42] applications into notion. Sure. Build [00:45] complex business logic and codify kind [00:48] of any workflow that you need. No [00:50] problem. But there's one issue. You need [00:54] code. Now, in 2026, that of course [00:57] doesn't mean you can't use them if [00:59] you're not an engineer. That's exactly [01:01] what this video is for. Over the next [01:04] few minutes, we will go over everything [01:06] you need from why workers matter to how [01:10] to set up your notion for workers all [01:13] the way to of course building your [01:14] actual first worker. And I'll promise [01:17] you, even if you've never touched a line [01:19] of code and also haven't used AI to [01:22] create code, by the end of this video, [01:24] you will have not one but three workers [01:27] shipped. It's an interactive one, right? [01:29] So, we have also a workbook to go along [01:31] and exercises so you can follow along [01:33] and yeah, just do the thing. Since this [01:35] is a bit of a longer one, we've [01:37] separated into five big chapters to make [01:39] it easier to follow along. First, why [01:42] workers matter, right? bit of a look [01:44] behind the scenes and where in your [01:46] business you might want to use them. [01:48] Second, the general setup that you need [01:50] to do both in notion and on your machine [01:52] to be ready to build your first workers. [01:55] And then number three, worker walk. So [01:58] let's start with chapter one why. In [02:00] terms of the why, it's basically giving [02:03] notion superpowers, right? If you wanted [02:05] to do it in one sentence. And with [02:08] superpowers, I mean that while notion is [02:10] an amazing tool, right, to build custom [02:12] systems for your team, there are kind of [02:14] like three typical use cases where you [02:17] will run into, you know, limitations [02:19] with notion. The first one is you need [02:22] to get external data into notion. Say [02:25] you have an external CRM and need to [02:28] pull it in or Google calendar in your [02:30] meetings should show up in notion or [02:32] it's a personal use case, right? and you [02:34] want to get the data from your ring, [02:36] right? Or like any kind of health [02:38] metrics into it. No matter what, you [02:40] have data outside of notion, you get it [02:42] need to get it into it. If you're lucky, [02:44] notion has a pre-built connector, but [02:46] even if they have it, like for example [02:48] for Jira, right, it might not do what [02:50] you need it to do. So that's kind of [02:52] area one where you run into limits. Area [02:55] two is the opposite one, right? And so [02:58] like here, the third one, it's sort of [03:00] pushing data out, right? Maybe you have [03:02] your information in notion, but it needs [03:04] to be elsewhere, right? You want to [03:05] publish a page to WordPress. You want to [03:08] sync something to your newsletter tool, [03:10] right? Or you want to push updates to [03:12] Slack. Again, the idea very often is, [03:15] right, that notion is this central hub [03:17] in your workflow. There's a bunch of [03:18] tools still around. It means you have [03:20] data flow in and data flow out. And then [03:23] the third aspect is within notion or [03:26] within notion the hub and that means [03:27] running complex workflows where the [03:30] business logic of notion's default setup [03:33] is just not there. Sure, we have built [03:35] and automations, we have buttons, we [03:37] have things to do like simple logic [03:39] steps. But when it comes to complex [03:41] stuff, for example, let's say, hey, [03:43] whenever I create this type of a [03:45] project, check our processes, figure out [03:49] what's the current list of tasks for [03:51] this type of project, create them and [03:54] assign them to the right responsible [03:56] person. Right? This kind of complex [03:57] business logic can't be, you know, [04:00] easily done within notion. And up until [04:04] basically workers, the answer to all of [04:06] this was well, we need an external [04:08] system. We need external often no code [04:12] tools to close the gap. That's why we [04:15] need workers, right? They fill this same [04:16] need. Getting external data in, getting [04:19] data out of notion and creating complex [04:21] business logic. Now, to be fully honest, [04:24] when notion workers first came out in [04:26] early alpha, um now months ago, um I [04:29] looked at it and I was like, "Okay, this [04:31] is super cool, but I don't quite know [04:33] where I would use it yet." And the [04:36] problem was this old story, right? If [04:38] that if you have a hammer, everything [04:40] looks like an A. So, I was looking at [04:41] this and I was like, well, okay, all [04:43] these three areas, right, where workers [04:46] are great. I have no code tools to do [04:48] this, right? Why do I need a worker? And [04:51] it's true, right? the ven diagram of [04:53] worker can do something notion uh sort [04:55] of no code true can something right [04:57] that's pretty much a perfect overlap to [04:59] a to a certain degree but the actual [05:02] question is around right it's like why [05:05] would you ever bother with no code tools [05:08] if you can create a worker for it to be [05:11] honest I think no code tools are dead [05:14] there's no reason for them anymore with [05:16] few exceptions that I just wrote a [05:18] LinkedIn p like a newsletter and blog [05:20] post on that link below But the TLDDR is [05:22] like why do we have no code tools? It's [05:25] so we can very fast deliver solutions. [05:28] Um it's much cheaper than traditional [05:30] software to develop and um you kind of [05:33] like have this like way for like [05:35] non-engineers to do it. All three of [05:37] these statements are no longer true uh [05:40] due to AI, right? With the right AI [05:42] workflows, it's very much possible to [05:44] create code faster than you could [05:48] manually click together a make or n word [05:51] flow. It's also because of that, right, [05:53] much cheaper to do it that way. And if [05:55] you particularly use notion, right, to [05:57] host it, you can execute much more [06:00] complex logic where it would be really [06:02] expensive to do this with like a hosted [06:04] solution with make or with nad. And then [06:08] of course, right, this like [06:09] accessibility, that's what this tutorial [06:11] is for, right? I promise you, even if [06:12] you have no idea of coding, by the end [06:14] of it, you're going to have your first [06:16] workers deployed, showing, right, that [06:18] you don't need to learn a no code tool [06:20] anymore. You're simply going to need the [06:22] learn the principles of, you know, [06:24] building these systems and you can go [06:25] ahead and do it with the help of AI. [06:28] Now, in order to get started, we [06:30] basically need two things. You need [06:32] either the clawed app or codeex [06:35] installed. And I'm going to link to both [06:37] of those below. Uh you can also use [06:39] cloud code of course in your terminal, [06:40] but if you're a beginner, right, I [06:42] highly recommend you either start with [06:44] um one or you start with one of these [06:45] two apps because they just make it a bit [06:47] easier and more accessible. So download [06:50] that. Step number two, go into your uh [06:53] finder, right, or your your general [06:55] folder and on the top level of that [06:57] folder, right? So in your user, uh [07:00] create a new top level folder that you [07:01] can call either claude or codeex, right? [07:04] depending on which of those two tools [07:06] you use. That's it for now, right? [07:08] That's where inside of it, we're then [07:10] going to go and build like our different [07:12] workers. But just like top level folder, [07:15] app downloaded. [07:17] And then we want to install the notion [07:18] developer tools. Now, we're going to go [07:20] in and this is also linked below in the [07:22] description and the companion post for [07:24] this, but basically there's one specific [07:26] command that you want to run in order to [07:29] install um the notion CLI. And the [07:32] notion CLI is just a way for you to [07:35] authorize and connect to notion because [07:38] when we're building workers, right, [07:40] we're not doing this in the notion [07:41] interface. We're doing this through code [07:44] and through our code. Why do we need to [07:45] tell notion, hey, we're actually allowed [07:48] to do this. So that's basically what [07:50] we're preparing here, right? Installing [07:52] these developer tools. You can do this [07:54] in two ways, and I'm going to show you [07:55] in a second. You can either go and open [07:56] the terminal on your machine, right? And [07:58] simply type in this command, right? copy [07:59] that command in and hit enter or you can [08:02] just ask right cla or codeex in the app [08:05] and that's what we're going to do in a [08:06] moment to install this for you right [08:09] super easy so yeah just like as a quick [08:11] heads up now make sure you install this [08:13] and then let's jump right in and take a [08:15] look at it after publishing the initial [08:17] version of this tutorial I got some [08:19] feedback that we forgot some helper [08:20] setup steps so just a quick insertion [08:22] here if this is the very very first time [08:25] that you use your machine for any kind [08:27] of you know code work. There might be [08:30] two additional setup steps that you need [08:32] to take and they depend a bit on which [08:34] machine you're on. I'm always on Mac OS [08:37] so apologies that Windows part is a bit [08:39] shorter but basically on Mac OS you [08:41] probably want to uh add the command line [08:43] tools um again you can simply ask claw [08:46] code or codeex and we go through that in [08:48] a moment right to do that for you but [08:50] you also see the command here on screen [08:53] and for windows it's a little bit more [08:55] involved because a lot of the notion [08:57] tools most notably the command line that [09:00] we're going to install in a moment they [09:01] are currently as of the release of this [09:04] video only available for Mac OS and [09:08] Linux. So you can directly run them in [09:10] Windows. Luckily there is a workaround. [09:12] It's called WSL. Uh basically creates a [09:15] small Linux environment inside your [09:17] Windows where you can run all these [09:19] things in. Again, right, you can simply [09:22] ask uh your AI tool of choice uh and you [09:25] know tell it, hey, we need this also to [09:27] run it. It's probably going to figure it [09:29] out on its own to be honest, but still [09:31] better if you can give it a bit of [09:32] guidance here. And then we also need [09:34] NodeJS which is a way for your machine [09:36] to run JavaScript which is very relevant [09:38] because a lot of the stuff that we are [09:40] going to use is built on top of that. [09:43] Again happening all behind the scenes. [09:45] Uh and you can simply ask your AI tool [09:47] of choice uh to install it. Right? We're [09:50] going back to the regular tutorial in a [09:52] moment there. You'll see then you know [09:54] just how to interact with um either [09:56] claude code or codeex to ask it to do [09:59] these things. Just remember right before [10:01] we dive into it um ask it on uh on Apple [10:05] right to get you um the command line [10:08] tools and NOJS and if you're on Windows [10:10] ask it to get you know WSL and then [10:13] NodeJS so that everything runs smoothly [10:17] starting with claw code if you have the [10:19] claw app installed and open it you can [10:21] in the settings right switch away from [10:22] chat or co you want to move all the way [10:24] over to code. Uh this is one of my many [10:28] Mac subscriptions that are running in [10:30] parallel. And then here, right, just to [10:31] talk quickly about the interface. [10:33] Basically, here's where you're going to [10:34] like just chat with the tool. Then here [10:37] you can add a folder. So you want to [10:39] make sure that you click on plus to add [10:40] another folder and pick this top level [10:42] folder, right, that you just created in [10:44] the previous step. On the on this side, [10:46] right, you can see what model are you [10:48] using, right? So in general, you want to [10:49] use Opus 4.7. Depending on how quickly [10:51] you burn through your usage, right? You [10:53] want might want to like reduce it from [10:54] extra high to high. But let's assume [10:56] you're not working too much. So, I like [10:57] to keep it on extra high if we can. [11:00] Don't enable fast mode. It burns through [11:02] credits a lot faster. So, you you want [11:05] to avoid that. Uh and then um here on [11:08] the bottom left, right, is another thing [11:09] to know about is sort of like auto, [11:11] right? So when it comes to clawed code [11:13] since it has control over your machine [11:15] and can store things right it's the [11:17] question of like you know like do you [11:19] want it to do that automatically or [11:20] should it ask for permissions now to be [11:22] honest nine out of 10 times I run mine [11:25] in bypass permissions which basically [11:27] means it's not going to ask anything the [11:29] reason for that being is that um I am [11:32] still learning coding concepts right so [11:34] Claude will know a lot better so the [11:37] chances of me catching a problematic uh [11:40] thing right are quite low. So a lot of [11:43] the work comes into our planning and [11:44] making sure we explain C cloud exactly [11:46] what we want it to do. Um on auto mode, [11:49] right, it's not going to uh do as much. [11:50] It's going to ask you for permissions [11:51] more often. So it's a good idea to start [11:53] on sort of ask permission, get a feeling [11:55] for what kind of things it typically [11:57] wants from you. And if you feel like, [11:59] hey, I can never, you know, review this [12:00] anyway. I just keep pressing accept, [12:02] except accept, then you can move on. For [12:04] now, we're going to stay on auto mode, [12:05] which is sort of like a compromise here. [12:07] Um that being said, right, this is AI. [12:10] it runs on your machine, right? So if [12:12] you give it access to your folders, you [12:14] need to be aware of that it can make [12:15] changes to it, right? So we want to [12:17] always be very precise with what we do [12:19] and most importantly always run it on [12:21] the smartest models because those are [12:22] the ones that have the highest chance of [12:24] interpreting the things that we ask them [12:26] the right way. All right, so with all [12:28] that out of the way, let's start by [12:30] asking it to install the notion CLI. [12:33] This is my simple promp for this, right? [12:34] Can you please install the notion CLI on [12:36] this machine? Here's the documentation. [12:37] And this is the general pattern that you [12:39] will follow for all of these things [12:41] where you're going to go to the place [12:43] where the documentation is and you [12:44] simply going to give it to it and it [12:46] will interpret that. And again all these [12:47] links are in the companion blog post [12:50] down below. And then I'm simply going to [12:52] hit enter. Now while this is running [12:54] here let's quickly take a look at how [12:55] this would look like in codeex. Now [12:57] codeex is basically the companion or the [13:00] equivalent to claude from openai. And at [13:03] the time of the recording of this video, [13:05] right, both of these interfaces might [13:06] look a bit different, but the principles [13:08] still hold true, right? You kind of want [13:10] to look for the area where you can put [13:12] this in. Now, with claw, right, at the [13:14] time of recording, you have these three [13:16] different options of chat, co-work, and [13:17] code, and you need to go to code. With [13:19] codeex, it's just one interface, right? [13:21] So, there's no differentiation between. [13:22] So, you just go to the default [13:24] interface. I've also selected my folder, [13:26] right? So, same principle here. It's [13:27] called Codex. Uh, permissions, right? [13:30] Same here. can sort of like give it full [13:31] access or default permissions or auto [13:33] review. Um, and then the model, right, [13:35] which in this case, right, 5.5. Again, I [13:38] like to switch it to sort of like high [13:40] just so we have a bit more confidence in [13:41] what it does. Apart from that, same [13:43] principle here, right? We're going to [13:45] simply run our command asking it to [13:47] install this. I've actually already [13:49] installed uh the CLI on my machine. So, [13:52] this is what I get back from CL, right? [13:54] It tells me, well, everything is ready. [13:56] Uh, one good practice when you install a [13:58] new app, right, is kind of asking, hey, [14:00] what are the things that I can do and [14:01] what are some of the commands that we [14:02] could run just to get a bit of an idea [14:04] of it. Um, and then it kind of lists [14:06] these things, right? And all of these [14:08] commands are commands that you would [14:09] type into your terminal, right? So on a [14:12] Mac, you would open your terminal, [14:13] right? And you would now type, for [14:14] example, ntn write doctor, right? That's [14:17] what was one of the the commands that we [14:19] that we saw here, right? That to check [14:21] the health on all configuration and so [14:24] on. [14:25] Um you can always do that right and the [14:27] easiest way is simply to ask broad to [14:29] walk you through this because some of it [14:30] it can just do for you right so this was [14:32] another example right we could simply [14:34] have typed the the cur command right uh [14:36] into our uh machine notion cloud was [14:39] able to do this for us the things where [14:42] we have to do something ourselves are [14:44] stuff where we need to com on top of [14:46] that we need to sort of like complete [14:47] like an action on our machine so n login [14:50] would be an example right so again you [14:52] can simply ask claude or co-work you [14:54] uh can you help me um use ntn login and [14:59] uh authorize [15:00] authorize this machine right and we kind [15:03] of tell you hey here's what you should [15:05] do and sometime it's going to tell you [15:07] hey I'm going to run this for you [15:08] sometime it's going to tell you hey open [15:10] your terminal paste this in can [15:11] basically just follow the steps what [15:13] will happen is you will take ntn login [15:15] right you will go to probably to your [15:17] terminal ntn login you will hit enter [15:19] and it's going to open a popup right [15:20] that says hey you know like please [15:22] select your notion workspace is where [15:24] you want to authorize it. So if you're [15:26] in multiple notion workspaces, you'll [15:28] have to pick the one where you want to [15:29] install workers to and then you simply [15:31] confirm and then you authorize. There is [15:33] one more step that you need to take [15:35] before you can then actually start [15:37] building your workers and that is you [15:38] need to activate them in notion. So for [15:40] this we actually need to go into the [15:41] notion UI. We need to go to our settings [15:44] and then under features you want to look [15:45] for workers. Now in order to use a [15:47] worker first you need to be on at least [15:49] a business plan, right? business or [15:51] enterprise and you need to be a [15:53] workspace owner to enable it. Otherwise, [15:55] you need to ask a workspace owner to do [15:58] this. Once the owner has enabled this, [16:00] then you can choose, hey, can all [16:02] workspace member do this or only admins [16:04] or only certain groups? That's kind of [16:06] the prep step. But once you're done with [16:08] that, you're good to go, right? CL or [16:10] codeex installed, folder set up, um [16:14] installed the notion CLI, logged in to [16:17] the notion CLI, and activated workers. [16:20] That's all the prep. So now let's [16:21] actually start building. As of the time [16:24] of this recording, notion basically [16:26] gives us three different types of [16:28] workers that we need to know about. And [16:30] even if we get more types, right, these [16:32] will probably be the most important [16:33] archetypes. We'll always expand, write a [16:35] blog post with the latest information on [16:37] them. But basically, we'll have sort of [16:40] like eventdriven workers, right? Which [16:42] means, hey, something happens. Let's [16:44] trigger this flow. Then we have sort of [16:47] like a scheduled data sync or a sync [16:50] worker. This is a worker that runs [16:52] regularly and regularly pulls [16:54] information from a third place. Right? [16:57] Sort of when it comes to one of our [16:59] problems, right? This very clearly [17:00] solves the problem of hey we need to get [17:01] information from an outside tool into [17:04] notion. This is a very special type of [17:07] worker because this worker as we will [17:09] see later actually creates and manages [17:12] its own database. So it's very unique [17:14] because it creates this immutable [17:16] database in notion right where like some [17:18] of the entity properties cannot be [17:20] changed right so you you can safely sync [17:22] data in and don't need to worry about it [17:23] someone modifies it and then last but [17:26] not least we have workers that extend [17:28] the abilities of your AI where you can [17:31] give a custom agent literally a tool [17:34] right and say hey custom agent now you [17:36] can talk to my newsletter page right or [17:38] now you can publish to WordPress and [17:41] it's sort of an on the agent to call on [17:43] them. We'll go over all of them in a bit [17:45] more detail, right? But I wanted to sort [17:46] of like explain, hey, these are the [17:48] three types of workers. We're going to [17:50] build one example for each and then you [17:52] understand exactly how to use them in [17:54] what situation and what the peculiar [17:56] things about them are because they are [17:59] of course in some overlaps, right? So [18:01] like the sync one, that one runs on the [18:03] schedule, but you can also hack your [18:05] your web hook one, right? to run on a [18:08] schedule if you kind of want scheduling [18:10] but you don't want a worker that manages [18:12] its own database. But, you know, no need [18:14] to, you know, figure that out just yet. [18:16] Just want to give you an idea. Hey, this [18:17] is the map, right, of what we're going [18:19] to look at in this course. And now, [18:21] let's start with the first one, right? [18:23] Because it's the simplest to understand [18:24] one. Something happens in notion. We [18:26] want to automate some cool business [18:29] logic before we continue with building [18:31] the specific worker. Um, a quick kind of [18:34] thought on general setup, right? This is [18:37] a beginner tutorial for workers. So I'm [18:39] not going to go deep into like the exact [18:42] development workflow, right, for our AI [18:44] enabled um uh development, right? And we [18:46] are using to build our internal tools or [18:49] like more complex client um scripts and [18:52] uh software. Basically, if you're [18:53] interested in seeing that uh let me know [18:55] down below in the comments. Here we kind [18:57] of want to focus on the 8020, you know, [18:59] like just building your first thing, but [19:01] having a bit of structure, right, will [19:03] help later down the line. So we know [19:05] that we've already in our folder right [19:07] set up claude or codeex and one of the [19:10] first things that we need to get used to [19:12] right when as non-developers we start [19:14] creating these is this is just a folder [19:16] structure with files on your local [19:18] machine of course you can back it up to [19:20] GitHub right but for now it lives here [19:23] on your local machine and that means we [19:26] want to within that right create folders [19:28] um for every single project that we're [19:30] developing right so in our case we're [19:32] going to create three workers so every [19:33] single worker right would be one top [19:35] level folder for example project tasks [19:37] the hacker news sync that we're going to [19:39] build in a moment and the wiki enricher [19:41] and within every single um uh project I [19:46] like to start with a selection of four [19:49] files right so this is basically sort of [19:51] part of my setup prompt um and uh you [19:54] can see like a simplified version of [19:56] that down here right uh you can either [19:58] you know screenshot that from here you [19:59] can check out the link in the in the [20:01] description right for the blog post or [20:02] in the setup template there you have in [20:04] full. But basically, we want to tell um [20:07] Claude or Codex, right, to always set [20:09] these up and make sure, hey, we want a [20:11] readme, which sort of explains the what [20:14] and why of what's going on here. We want [20:16] learnings, so as this goes through and [20:19] you know, your AI figures out, ooh, [20:21] actually the notion API works [20:23] differently, right? So, it writes it [20:25] down so next time it doesn't have to [20:26] figure this out again. And the change [20:28] log, so of what changed now, none of [20:30] these will be actually filled out by us, [20:32] right? All of these will be maintained [20:34] by the AI but it sort of automatically [20:36] builds the context so that it gets [20:38] better next time because the same [20:41] principles right that you have elsewhere [20:43] for AI still apply here create context [20:46] make sure that for every single session [20:47] that you work with it you manage your [20:49] context window and you don't balloon it [20:51] and then if you go to your next session [20:53] you have still that information [20:54] externalized right so you don't have to [20:56] explain yourself every single time [20:58] always again so this is the structure [21:00] that we'll use so basically after you've [21:02] created you know, your main project, you [21:04] can ask um, you know, Claude or Codex to [21:06] uh, set up the structure like this. And [21:08] then with all of that out of the way, [21:10] we're ready to actually start building. [21:13] Now, if you've been watching this so far [21:14] and think, well, this is great, but I [21:16] don't have time to build workers or I [21:18] need my whole team to be able to build [21:21] workers. Well, then let's have a chat. [21:24] As you might know, I'm on a mission to [21:25] build Europe's number one notion [21:27] consultancy. And that means also helping [21:29] you both with AI flows in and around [21:32] notion and of course with workers. [21:34] Whether you're looking for a full-scale [21:36] transformation of your existing setup, [21:38] right, where we go in and help you bring [21:41] your notion and the automations and [21:43] business logic to the next level or [21:45] rather a targeted workshop, we have [21:47] something for you. Just check out our [21:49] link in the description or send me an [21:51] email and we can figure out the details. [21:53] But for now, let's take a look at how to [21:55] actually build that first worker. And [21:58] here is that cool business logic. Our [22:00] first worker will be triggered on a web [22:02] hook. That means something happens in [22:04] Ocean and then it will spring to life. [22:07] The specific example that we're going to [22:08] choose for this is we're going to have a [22:10] project and we're going to have a [22:11] project of a specific type. For example, [22:14] a new marketing campaign. Whenever that [22:17] happens, what we want to happen in the [22:18] background, right, is we want to say, [22:19] "Hey, new marketing campaign, let's [22:21] create these default tasks." This has [22:24] been basically been an issue, right, [22:25] with notion ever since Notion came [22:27] around. We need to create these default [22:29] task entries. And there's a bunch of [22:31] workarounds, but none of them are really [22:33] like scalable, right? We always told [22:35] clients, hey, the the ROS raw is the [22:37] best solution. Is this where you have a [22:40] dedicated process database? We'll take a [22:42] look at it in a second. uh and then [22:44] basically you have a no code automation [22:45] that does this. Now we can finally use a [22:47] worker right so all within notion and [22:49] the flow is simple right new project is [22:51] created of a specific type we're going [22:53] to look up what is our SOP for that [22:56] situation and then we're going to create [22:57] a set of tasks specifically for that [23:01] right and we can level this up like you [23:03] know populate specific due dates right [23:05] we can level it up by assigning the [23:07] right person to these tasks where the [23:08] world is your oyster once you have this [23:10] basic flow ready just before we dive in [23:13] right what are some other examples where [23:15] you could use these web hook workers, [23:17] right? One could be for recurring tasks, [23:19] right? It could be so, hey, whenever I [23:21] check off a recurring task, right, make [23:23] sure that you create an next one. It [23:24] allows you to do much more advanced [23:26] logic than our previous workaround which [23:29] relies on notion custom automations. But [23:31] yeah, this is a great one. So, let's [23:34] take a look first of how the setup is in [23:36] notion, what we need there because we [23:37] have a little bit of prep and then we're [23:39] going to look at how to build a workout. [23:41] Also, quick side note, since this is [23:43] meant to be an interactive uh tutorial, [23:46] right? We have sort of a workbook with [23:48] all of these elements linked down below, [23:50] right? So, make sure to go there. Uh you [23:52] need to subscribe to the newsletter. You [23:54] can unsubscribe immediately if you don't [23:55] want our tips, but then you get uh a [23:57] list sent with all our templates and [23:59] resources and also right this one here [24:01] specifically so you can follow along and [24:03] don't have to build the, you know, the [24:05] preset yourself. But yeah, so let's go [24:08] and check out notion. fairly simple [24:10] setup in notion, right? This is not for [24:12] like pretty uh notion setups. If you [24:14] want to see that, right, we have a full [24:15] tutorial both on how to manage your [24:17] projects in notion, how to build your [24:18] business setup in notion, how to build a [24:20] CRM, and we have a bunch of follow [24:22] alongs. For this one, we just want to [24:23] focus on having the backbones here. So, [24:26] basically what we need is we have our [24:27] project database, right? And this [24:29] project database has tasks linked, [24:31] right? So, I can see, hey, here's my [24:33] project, right? And here are my tasks. [24:34] And currently, we don't have tasks [24:36] linked to it. Ideally what I want to [24:38] have is like whenever I create a new [24:40] project with a type new marketing [24:41] campaign right or marketing campaign I [24:44] want to make sure that it prefills tasks [24:47] and so for this setup to work I need [24:49] actually a dedicated process database [24:51] right this is separate from tasks I have [24:52] one global task database that contains [24:54] all the actual action items and then I [24:57] have this process database which lists [24:59] out hey these are the ideal steps that [25:01] needs to be taken and in there right we [25:03] have a type so that means we can have [25:05] multiple processes for like you know any [25:07] number of projects in here and we can [25:09] say hey these are sort of like the seven [25:11] standard steps right that need to [25:12] happen. You can think of this like [25:14] living documentation. It's one of the [25:16] big benefits of notion because it means [25:18] hey your documentation always influences [25:21] the execution because every time right [25:23] we create a new project it will look at [25:26] these seven tasks and create them and if [25:29] we update our documentation right [25:31] because we say hey actually now we [25:32] should make have nine steps for the [25:34] immediate next campaign that gets [25:36] launched right these nine tasks get [25:38] created so you don't have this problem [25:40] that you often have in documentation [25:42] right where you update your [25:43] documentation and no one follows it [25:44] because in the actual daily execution it [25:47] you know doesn't really carry over or [25:49] the other way right that you have now [25:50] your list of tasks always in your system [25:53] but you never update the documentation [25:54] right it's inevitably linked so uh that [25:57] being said right just more as a side [25:59] note um we have here our database for [26:01] processes we have the name of the task [26:03] right the project type I like to order [26:05] them so we know in which order they [26:06] should be loaded into the system um and [26:09] then we already have like sort of like [26:10] two properties for some more advanced [26:12] elements right we have a due date offset [26:14] So we can say hey based on when the [26:16] start date of this campaign is right [26:18] after 3 days this task should be due [26:20] right after 7 days this should be due. [26:22] So you can sort of like set dynamic [26:24] milestones and we also have already [26:26] roles assigned to this right. So you'll [26:28] notice on projects we'll have different [26:31] roles across the board. Currently it's [26:33] all me but basically we can do then some [26:35] cool things like having the worker go in [26:37] and be like okay um who is the project [26:40] manager on this one? It's Matias. Let's [26:42] assign that task to Matias, right? So [26:44] you can see you can build really [26:45] elaborate systems. Again, the world is [26:48] your oyster, but this is kind of our [26:49] base information, right? Now we want to [26:51] create a worker that says, hey, okay, [26:52] whenever we have a new campaign here, [26:54] right? Or let's say whenever we set a [26:56] campaign to in progress, maybe let's do [26:58] that. Whenever we set a project to in [27:00] progress, we want to make sure that it [27:03] checks what's the project type, gets [27:05] these tasks, and assigns them. all these [27:07] setup steps, right? Just make sure that [27:09] later down our life is easier. So, here [27:12] I have my I'm back in Codex, right? I [27:15] actually already have uh created my my [27:17] main folder, right, for this specific [27:18] one, right? Auto task creation. And now [27:21] I'm going to ask it, hey, can you please [27:22] set up the um files, right, that I want [27:25] here um and keep them updated throughout [27:27] our session. And then once done, please [27:29] also load the skills, right? If you've [27:32] watched some of my previous video, you [27:33] know how important skills are for AI [27:35] because skills are basically just small [27:37] instructions that teach UI a specific [27:40] thing. And I always like to think of it [27:42] like, you know, my AI as a poke, right? [27:44] I can teach it something new. And we [27:46] always want to teach it the right set of [27:47] skills for a project. When it comes to [27:50] building notion workers, that's of [27:52] course, you know, how do you entangle [27:53] them? And luckily, notion is creating [27:55] these skills for us. So, all we need to [27:57] do is point it at the right address. [27:59] Again, link for that down below in the [28:00] description. and it's going to set [28:01] itself up. Now even though skills are a [28:04] concept that comes initially from [28:05] anthropic and so from the company behind [28:07] claude works exactly the same way in [28:09] codeex right also there you would tell [28:11] codex hey please set this up codex won't [28:13] have a clot in defile right but sort of [28:14] in general instructions um but same for [28:17] the other types right read me learnings [28:19] change log and of course um the worker [28:22] skills so while this is working we can [28:25] already talk about the next part and [28:28] that is then explaining the actual thing [28:31] you're trying to [28:32] When it comes to that, I highly, highly [28:34] recommend that you are going to get [28:36] yourself um a dictation tool. It's so [28:40] much better talking to AI rather than [28:42] having to type this all out. My two [28:44] favorites are Monologue if you are on a [28:47] Mac or Whisper Flow if you are not. Um [28:50] link to them uh down below in [28:52] description. Both amazing, really, [28:53] really cool tools. Um work also on [28:55] mobile, right? And they will level up [28:58] your AI skills so much. We just ran our [29:01] first ever live workshop, right, with a [29:03] small group of people and it was so cool [29:06] to see everyone talk to AI and the [29:08] feedback throughout was that hey, now [29:10] that I talk to it, it's so much better [29:12] than before. So, take the serious, [29:14] right? Get yourself that it's really [29:16] worth it. So, uh, with that out of the [29:18] way, right, AI was working in the [29:20] background. So, now I'm just going to [29:22] start dictating and tell, hey, this is [29:25] what we're now actually trying to do. [29:27] Hey, Claude, awesome. Thanks a lot for [29:29] the setup. Now I need your help and want [29:31] to just describe what I'm trying to [29:33] build and then I need you to go over [29:35] this and ask any kind of clarifying [29:37] questions that you, you know, need to [29:39] have solved before we actually dive into [29:42] building this. Um, here's what we're [29:44] trying to do in Notion. We going to have [29:47] a database with our projects. And um [29:50] whenever there's a project with a [29:52] certain type and that is set to the [29:54] status in progress, I want to uh have [29:57] the worker check a process database that [30:01] lists out a sample list of skills uh [30:04] sorry of tasks and those tasks uh should [30:08] then be created for that specific [30:09] project. Um there will be some um a bit [30:11] of advanced logic right. So tasks uh in [30:13] the the process list they have a they [30:15] have a due date offset. So we want to [30:17] make sure that hey based on the project [30:19] start date you know plus 3 days for the [30:21] due date plus 7 days for the due date [30:23] and we also have a role um and on the [30:26] project we have a notion person property [30:28] with that role. So we want to make sure [30:30] we assign the task that we create in the [30:32] task database to that role. So overall [30:35] three databases projects, processes and [30:38] tasks. When project hits certain [30:40] condition, check in processes and create [30:43] entries in tasks related to that um [30:47] project. [30:48] And you see like if I had to type this [30:50] all out, that would take forever. It's [30:52] also totally fine that I had to correct [30:53] myself in there, right? Like AI is very [30:55] very smart. So absolutely no problem [30:57] about that. Now before we're going to [30:59] send this off, uh I'm going to actually [31:01] quickly go to notion. I'm going to grab [31:02] the links to um my uh notion entries and [31:07] paste them in here. That's going to make [31:09] it so that Claude can look at them. In [31:11] order for that to work, you need to of [31:13] course make sure that your Claude is [31:16] connected to notion uh ideally through [31:18] the MCP. That's the easiest way. So, if [31:20] you haven't done that yet, here's how [31:22] you can set this up. You want to [31:24] navigate to the settings for your claw. [31:27] Showing you codeex in a moment. uh and [31:29] then uh find currently it's called [31:31] customize and then connectors otherwise [31:33] you can just ask claud hey where do I [31:35] add an MCP and we'll tell you where the [31:37] current version is because sometimes [31:38] they shuffle the UI around but basically [31:40] under connectors right you have notion [31:42] here and then um if you have it [31:44] connected already if not right you're [31:45] going to click on plus um browse [31:47] connectors you're going to select notion [31:49] you're going to select the workspace [31:50] that you want to connect to and then [31:51] you'll see this now I have pretty much [31:54] on most of these uh always allow so it [31:56] simply does the job right particular for [31:58] like all the read own tools. I don't [31:59] want it to ask me before it reads a page [32:01] in notion. You can of course change this [32:03] also to always confirm. Um if you want [32:05] to in codeex uh you would do this in [32:08] settings and then MCP servers and then [32:10] fill out the values here. Basically, you [32:11] could also ask, of course, as always, [32:13] AI, invite, hey, how do I set up the [32:15] notion MCP? Um, but as I was navigating [32:18] here, every less we don't even need this [32:20] because remember previously, we already [32:23] logged into our notion workspace through [32:25] the CLI, right? When we typed in ntn [32:28] login. So, the tool is already [32:30] connected. We don't actually need the [32:32] MCP. So, you can ignore that part. [32:34] Actually, still the database links, [32:36] those we definitely want in there. So [32:38] sending this all off for it to work on [32:41] it again. And the key thing here is that [32:44] we ask it to ask us questions. If you [32:46] want, you could actually switch um into [32:49] plan mode, right? So Claude has this uh [32:51] built in where um if you basically [32:53] switch it to plan mode, there's some, [32:55] you know, prompting in the background [32:57] going on from Antropic that says, "Hey, [32:58] you know, don't do anything yet. Just [33:01] figure out what how you can best achieve [33:03] this and ask user questions." It's a [33:05] great sort of like start and entry [33:07] point, right? So there's nothing wrong [33:08] with switching from auto to plan mode. [33:10] Uh going through that for a bit. Um I [33:12] personally right for our proper [33:14] development flow I have a set of skills [33:16] and most notably the grill me skill [33:17] right not for me from an amazing dude [33:19] called PCO. Link into that also down [33:21] below which just forces the AI to ask [33:23] you really a ton of questions. Great for [33:26] more complex builds. But for this here [33:28] right we it's a very this is a very [33:30] straightforward simple process. we'd [33:32] probably be fine even without the AI [33:34] asking us anything and just having it, [33:36] you know, oneshot this. But still, it's [33:38] good practice uh to whenever you build [33:41] something to explain what you're trying [33:42] to do to the AI and then have some way [33:45] or shape whether that's plan mode, a [33:46] specific skill or what we did here, [33:48] simply typing in ask me any kind of [33:50] clarifying questions so you know exactly [33:52] how this is supposed to work. Um to make [33:55] sure you align, right? Because hidden [33:57] assumptions is the single biggest reason [33:59] why AI things fail. So here we see our [34:02] questions come back. Okay, how should a [34:03] worker get triggered when a project is [34:05] in progress? That's great, right? So I [34:07] didn't even actually tell it that. So we [34:09] will do the web hook on status change, [34:11] right? That's the whole point. Remember [34:13] three types um of uh systems and you see [34:16] also you could do a sync but that's M [34:19] does it own managed database. It's again [34:21] it's awkward because we're writing to [34:22] something existing. So the AI is pretty [34:24] good at that and we don't want um a tool [34:27] on um because then we would need to run [34:29] an agent every time we want that and [34:31] that costs agent tokens, right? Which [34:33] are very expensive. So we want to do the [34:35] web hook. Uh which project types should [34:38] it trigger works for all types. Yes. And [34:40] then in the process row it can simply [34:42] check um whatever is in there. That [34:44] means what we're building here will work [34:45] also if the team adds 10 more types and [34:48] 10 more processes which is one of the [34:50] reasons these automations are so [34:51] powerful. Right? You need to set it up [34:53] once, but then everything else happens [34:55] in the UI, right? You can create three [34:56] more project types, um, three more [34:58] processes, and it will work just without [35:01] having to touch this. All right? If it [35:03] already has tasks on the process [35:05] process, don't recreate them. Yeah. Um, [35:08] so this is generally a great idea, but [35:10] it requires us to have sort of like an [35:12] some logic of how to match tasks, right, [35:15] based on the name or based on an ID. So, [35:17] we can see how we want this. For this [35:19] one, we're actually going to do it [35:21] simpler. We're going to say, hey, if we [35:22] set this to in progress, we want it to [35:24] create it every time, even if it creates [35:26] duplicates for your production, right? [35:28] You probably want to want to do this. Uh [35:30] um if you do this actually for your team [35:32] and, you know, have some sort of ID on [35:33] your tasks. But here, um actually, let's [35:36] go with this. Let's see what it comes up [35:38] with as a solution. And still create [35:41] task just leave the owner empty. Yes, [35:43] that makes sense. or actually um now [35:46] here um always uh assign this then to [35:51] the owner of the project or as a [35:54] fallback to the person who created the [35:58] project right I never want to have empty [36:00] tasks or unassigned tasks in my system [36:02] there should always be someone who's [36:04] responsible so now same game right going [36:06] to wait for this a bit and come back [36:08] building with AI is a lot of answering [36:11] questions you know bursts of focus and [36:13] then waiting for a moment, right? [36:16] Usually I would go now and do a sort of [36:18] like cycle through a few projects, [36:20] making sure that there's always the next [36:22] questions ready when I come back, but it [36:24] also is fairly fairly quick here, right? [36:26] So, let's see whether it has any other [36:28] things. And we see as we're doing this, [36:30] right, it's going to write these details [36:32] into the docs into the pages that we've [36:34] set up, which is why it's so useful [36:36] because as soon as you give AI a bit of [36:38] structure, like in this case, the [36:40] decisions, the change log, right, it's [36:42] going to make have a lot easier time, [36:43] right, to keep track of this. And then [36:46] if we have to restart the session at [36:47] some point, right, because the more we [36:50] do, the more the context will app, the [36:51] worse the performance of the AI will [36:53] get. Um, we don't have to run through [36:55] all the questions again. here. For now, [36:57] our context uh should still be fine [36:59] because we didn't have now a lot of [37:01] conversation with it. But if you feel [37:03] yourself, you know, getting back and [37:05] forth with it for a while until you [37:06] align on the kind of workout you want to [37:08] build, um then it's best to wrap up. [37:10] We'll do like an example of sort of like [37:12] a wrap-up uh later. But for now here, [37:14] I'm going to tell yes, uh please go [37:16] ahead and start um building the um the [37:21] worker. Let me know if you need anything [37:24] from me. Right. And again for all of [37:27] this we're in auto mode. So for the most [37:28] part it will just execute but some uh [37:30] commands will have to um still approve [37:33] manually. If we were to switch to bypass [37:35] permission it would now simply like run [37:37] until it thinks hey I need some some [37:39] input from my TS or some UI actions [37:41] right for example if it starts doing [37:44] this and it checks hey are workers [37:45] active in the workspace right the setup [37:47] set before and you haven't done that yet [37:49] it will tell you hey you know can you [37:51] please go in the UI and do that because [37:52] it can't take all exceptions. Same with [37:54] the NTN logger. So, just going to pause [37:58] here quickly and then come back when [38:00] Tino tells us again it's ready for our [38:02] input. So, here we are again. It needs [38:05] some input. It needs to it wants to [38:07] check, right? Whether it to uh see [38:10] whether um web hooks actually enabled [38:12] it, right? We always tell you in the [38:13] chat what it sort of like uh wants to [38:16] figure out. Um it tells me, hey, auto [38:18] mode um doesn't allow me to do this [38:20] without you approving it. So, you know, [38:21] you need to click once. Um, and [38:23] basically what it's doing here, uh, so [38:25] we're going to say yes, deploy in the [38:27] documentation from GitHub, right? The [38:28] workers, uh, notion still has some alpha [38:30] flags in there. So, for example, hey, [38:32] you know, workers might not be enabled [38:34] or the web hook capability might not be [38:36] enabled, but these like everything from [38:39] this tutorial should be uh in your [38:41] workspace as of today. Um, again, as [38:44] long as you're at least on the business [38:46] plan. So we still going to go ahead and [38:49] test this but for our next worker uh [38:51] this is one of the things where we could [38:52] tell it ahead of time right um by the [38:54] way we have workers in here um web hook [38:57] books work yada yada so it doesn't have [38:59] to go through this again this is what [39:01] the learnings are also for right so that [39:03] it remembers this and doesn't have to to [39:05] worry so it worked for quite a while [39:07] right more than eight minutes in a go [39:10] now it's pretty much ready it tells me [39:12] it's currently blocked on task number [39:14] three it needs an integration token [39:15] right so even So we have of course um [39:18] logged in here right to do actions in [39:20] notion. So that in order for the worker [39:22] to do something that sometimes still [39:25] needs a proper integration. So it also [39:27] gives us already the URL where we can uh [39:29] do this right. So this is one way [39:30] although notion just updated this. [39:32] Notion just has now a new developer [39:33] portal and link for that also down [39:35] below. So we want to go there create an [39:38] API token right um name it accordingly [39:41] and then granted access to these. So [39:44] let's go ahead do that and then I also [39:46] show you what the best way is to give it [39:48] this token because of course you could [39:49] now paste you know your API token in [39:51] here. It always recommends you not to do [39:53] it, right? It's not a good security [39:54] practice to sort of expose your API [39:56] tokens here in the chat. So it's going [39:58] to talk about something called an EMD [39:59] file. Again, even if you never heard of [40:01] that, I'll just show you like how it [40:03] works and you can follow along. So no [40:05] worries there. In the notion developer [40:07] portal, you want to go to connections [40:08] and then here create click on new [40:10] connection. Then you can choose which [40:12] workspace this should be installed. And [40:14] that's perfect. We're going to call this [40:15] YouTube um workers uh demo, right? Going [40:19] to create an access token. Uh create a [40:21] connection. And now we have it here. And [40:23] we can copy it and add it. We can also [40:25] change the permissions here. Right? So [40:26] if we say, hey, we have a worker that [40:28] doesn't need to make actual changes, [40:29] right? We could give it only read uh [40:31] permissions or if it needs to interact [40:33] with comments, right? So you see there's [40:35] a few extra settings, but for like the [40:36] default ones, this is totally fine. And [40:38] you can click on copy here. In order to [40:41] create your ENV file, we need one more [40:44] sort of engineering tool or code related [40:47] tool. And that is a way to actually you [40:49] know edit these files. Now the one that [40:51] I would recommend is that you go with VS [40:54] code. It's sort of like uh that's a [40:56] great starting default like very broadly [40:58] used um developer tool. If we were to [41:01] write code, right, this is actually [41:02] where we would do this, right? So when I [41:04] did my web development like very basic [41:06] crash course a few years ago this is [41:08] where you would go and would create your [41:10] HTML and CSS files. Um now we're [41:13] basically just using it as a glorified [41:14] file browser here. So after you install [41:17] it, you basically open a folder, right? [41:19] So you'd go to the file folders and you [41:21] then um select the one that you created, [41:23] right? So this is the autotask creation [41:25] folder in our claw project and you see [41:28] all the different files it created here. [41:30] And among others, right, the read me, [41:32] right? Um that we told it to set up with [41:35] uh it stuff. And this is also by the way [41:36] the easiest way to check out well what [41:38] is actually the readme, right? What did [41:40] the AI decide to to document here, [41:42] right? So here um we say part of the [41:44] YouTube demo workers project live build [41:45] on screen no secrets or tokens on screen [41:47] right I told it to hey please when when [41:49] I share my screen avoid printing out any [41:52] any secrets right that we should not [41:53] have in there um we can also check what [41:56] the learnings are right so this is what [41:57] it what it learned um when when doing [41:59] this right so choose a web hook trigger [42:01] um gate result right is enabled so this [42:04] is how we can just see what happens and [42:06] follow a bit along um now in order to [42:09] give it now the secret what we're going [42:10] to do is we're going to click on create [42:11] a new file that and we're going to uh [42:13] type in env. Um that's sort of our [42:16] environment variables. Uh it's like how [42:18] we hand uh our code usually things that [42:21] we don't want to expose anywhere. And [42:23] then we're simply going to call this [42:24] notion. I'm going to just type it, [42:26] right? It's just a text editor. So we're [42:27] going to call this notion API token [42:30] equal to and I'm going to paste in my [42:31] token. Now in this case, right, I'm [42:33] exposing my integration token on the [42:36] screen. That's a bad idea. You don't [42:37] want to do this. So before this video [42:39] goes live, I'm going to go back to the [42:40] integration. I'm going to rotate this [42:42] key so it no longer works. But this is [42:44] basically the secure way, right, to hand [42:46] that to your AI rather than just pasting [42:48] it in the chat. So that being done, we [42:51] can now go back to our cler and tell, [42:53] hey, you now have an env token. So back [42:56] here, I'm going to just tell it, all [42:58] right, I set up a token for you in an [43:01] env file. Uh there's also one project in [43:04] the demo task in the demo projects [43:06] database that has a type set. You can [43:08] basically use that for testing. All [43:10] these entries are demo entries [43:14] and then set up have a look again. Now [43:17] you also need to make sure that your [43:18] integration token can actually access [43:20] the databases right that it needs to [43:22] interact with. In order to do so, you [43:23] want to open the database, click on the [43:25] three dots on top right corner, [43:26] connections, and then add connection. [43:29] And here, look in this case, right, for [43:31] uh YouTube workers demo, right? That's [43:33] the one that I created. And add this to [43:34] the page. And it needs to have access to [43:36] all the data sources, right? That's [43:38] supposed to access. So if you have also [43:39] like a top level page that contains all [43:40] of them, you can also give it access to [43:42] that. And then with notions waterfall [43:44] permissions, right, they all will be [43:46] shared. So we are nearly ready to test [43:48] our first worker. uh cloud is going to [43:50] ask us, hey, should we push this, right? [43:52] Are we ready? So, yes, we want to push [43:54] and deploy, which means it's going to [43:56] give the worker that is installed on our [43:58] workspace the integration token so it [44:00] actually can do what it's supposed to [44:02] do. Quick note, you can also confirm [44:04] that your worker is now uh in the [44:06] workspace by going to your settings, [44:08] then to workers, and then here, right, [44:10] where we initially turned them on, we [44:12] now see we have our here auto task [44:14] creation, right? The one that we just [44:15] enabled. You can also then disable it [44:17] from here or delete it from here, right? [44:18] if for whatever reason uh you forget how [44:21] to to access through clot but yeah [44:23] that's where you'll see that it exists. [44:25] All right, our automation is ready which [44:28] means we need to now just make sure that [44:30] it's properly wired up because remember [44:32] we triggered via a web hook. So we need [44:34] to still set up that web hook part and [44:37] cloud actually is going to explain to us [44:38] what we should do here right we should [44:40] create a new automation on our database [44:43] um make sure that when the trigger is [44:44] set to status right we want to ping a [44:46] URL and it gives us this URL um in a [44:49] text file right so we can ask it to [44:50] print it out here or we can go now back [44:52] to VS code right and sort of inspect [44:54] this file the reason why it typically [44:56] prefers to give it to us in a file and [44:58] not print it out is that of course [45:00] anyone with that URL can ping your [45:02] worker right a book can also come from [45:04] outside of notion. One way to increase [45:07] the security here is to simply um add a [45:10] secret, right? So we can ask um Claude [45:12] here in a moment if we wanted to, hey, [45:13] can you make sure that there's also like [45:15] a header secret to authorize. So just [45:17] when we ping this web hook, just in case [45:19] someone else comes across this URL, they [45:21] can, you know, just keep triggering [45:22] tasks, right? It needs a password as [45:24] well. So that's generally a good [45:26] practice. So I'm going to uh I would [45:28] recommend, right, that you go ahead and [45:30] try this for yourself, right? If it goes [45:32] to that and doesn't suggest it [45:33] proactively, ask it to give you um you [45:36] know a secret that you can put in a [45:37] header. Very important. We'll talk about [45:39] that in a second. But in this case, [45:40] right, let's quickly go to VS Code, [45:42] right? And here, let's look for the file [45:44] that it said it created right here. Web [45:47] hook URL text. Perfect. So, here's the [45:49] web hook URL that we need. So, I'm going [45:51] to copy that and then I'm going to head [45:53] back over to Notion to set up the actual [45:56] trigger. So in notion now for the [45:59] automation right let's click here you [46:01] know um trigger uh auto tasks uh [46:05] basically just follow what claw told us [46:06] right but here in the automation we say [46:08] okay whenever the status is set to um in [46:12] progress then we want to trigger this by [46:14] the way if you want a full tutorial on [46:15] how notion automations work right I have [46:17] a full guide for that also down below [46:18] it's even with workers it's still well [46:20] worth learning those because kind of [46:22] they're they're the pre-step right often [46:24] times a simple automation will be enough [46:26] and And so you want to know, hey, when [46:27] do I use a worker? When do I use this? [46:29] And as you can see, you very often will [46:31] need normal automations to trigger your [46:33] workers. So send a web hook and we're [46:35] going to send this web hook to the URL. [46:38] Now, in case what I highly recommend, [46:40] right, you set up a secret, right? What [46:41] you would do is you would say add a [46:42] custom header and then whatever it tells [46:45] you to put in as the key, right? So like [46:46] X header secret, something like this. [46:49] And then here in value, that's where the [46:51] password would go. And that would then [46:53] allow you right to make sure that hey no [46:55] one outside of your workspace can [46:56] trigger your worker if they randomly [46:58] happen to figure this out right which is [46:59] still it's very unlikely but as a [47:01] security best practice a good idea. Um [47:04] you don't need to send anything else [47:05] right that's enough. Um this is also one [47:07] thing to keep in mind sometimes when you [47:08] work with claude or codecs um and set up [47:11] one of these webbook triggers it's going [47:13] to ask you to hey and set this in the [47:15] body right or have these values sent [47:18] through. If that happens to you, you [47:20] need to just tell uh Claude or Codeex [47:23] and that actually with a notion [47:24] automation, you can't set any values [47:26] except for the header, right? The only [47:28] thing you can send is um the actual, you [47:31] know, header uh elements and then the [47:33] content is like predefined by notion, [47:34] right? So, we can't write something [47:36] custom in there. So, the easiest way is [47:38] to simply just tell it to like, you [47:39] know, have a look otherwise at the ID or [47:41] have a look at the payload once uh and [47:43] digest this and then figure out how to [47:45] use that. So yeah, in this case, we're [47:47] just going to click on enable and then [47:48] we can uh test this. Another way to of [47:51] course send these web hooks, right, is [47:52] through button actions, right? Just so [47:53] you've seen it. Um, right. So if you [47:55] want to sort of like trigger this [47:56] manually, you could also uh look for [47:58] like the the button here. Where is it? [48:00] Yep. Let's call this and let's call this [48:03] create tasks. The nice thing is this [48:05] will work regardless of whether you know [48:07] the uh status change triggers it or this [48:09] triggers it, but it's both the same. So [48:11] here we say okay, also send web hook [48:13] right. Same idea. Um, all right. And [48:16] save. Now we are wired up here. Now what [48:19] I'm going to do is I'm going to go back [48:20] to claude before I click or trigger [48:22] this. Tell it everything is ready and [48:24] that it should please start watching the [48:26] first run just so we see whether [48:28] everything works or whether something [48:29] needs to be, you know, troubleshooted [48:30] because again sometimes it might miss [48:33] something in the first run and having it [48:35] watch this first run just makes sure we [48:37] get any kind of debugging that's [48:39] necessary immediately done. So, and [48:41] doing so again super simple just asking [48:43] it can you please watch the worker and [48:44] we see yes um it's going to watch this [48:46] for the next 5 minutes. So I can go back [48:48] here now and either you know use this to [48:51] set it to in progress or the other one [48:53] and then we will go back to our claude [48:56] and see whether that catches us and whe [48:58] something happens. And of course here in [48:59] notion we can also investigate you know [49:01] are these tasks being added and that [49:04] looks pretty good right so here we have [49:06] our entries right uh we have them all [49:08] set the status is set the due date is [49:10] set based on the offset right that we [49:12] created of course they're all assigned [49:13] to me because I'm filling all the roles [49:14] in our demo project but still pretty [49:17] neat so let's go back to demo projects [49:18] and actually let's create another one [49:19] sort of like second campaign campaign [49:24] second test let's just do that right [49:26] let's keep it simple uh here Again, uh [49:28] what we would we can test this one sort [49:30] of like what happens if we don't have [49:32] something set, right? Let's sort of like [49:34] test this as well, right? So here I'm [49:35] going to set this to in progress. This [49:37] one shouldn't get any tasks because it [49:39] doesn't have a type, right? And on third [49:41] test, that's the event and let's just [49:43] make sure we get all our properties here [49:44] that we need. Right? So like I want to [49:46] see the status and I just want to see [49:48] the project type. Yeah, perfect. So [49:50] here, right, this one is not uh in [49:52] progress, but this one will get now new [49:53] marketing campaign and here I'm going to [49:55] set it in progress. So if I investigate [49:57] this, I should not see any tasks. Yes, [49:59] perfect. Zero tasks. If I investigate [50:01] this one, right, this in a moment should [50:03] get the tasks uh from the worker. Let's [50:06] see whether it runs all through. And [50:07] then we can have a look right at our um [50:11] um plot and see whether it investigated [50:14] anything from those runs. All right, [50:16] tasks did not come through in our third [50:18] test, which is even better because now [50:20] we can troubleshoot. Right, this might [50:22] always happen, right? You might set it [50:23] up and okay, it doesn't do what it's [50:25] supposed to do. What do we do? Well, we [50:27] ask AI. Now, at this point, right, our [50:29] context window is getting quite happy. [50:30] I'm still going to do this for now in [50:32] here, but this is at some point, right, [50:34] you might want to take the decision to, [50:35] hey, okay, let's actually document [50:36] everything and then do troubleshooting [50:38] in another window. But for now, I'm just [50:40] going to ask it um um okay, the [50:43] background watcher just ran again, [50:45] right? So, that's good. Um so, I'm going [50:47] to ask it now here um again with [50:49] dictation. [50:51] Actually, I tried to do another few [50:52] tests. I created one project without a [50:54] type and set it to in progress and I [50:56] created another project that is has the [50:58] type and I set it to in progress. It [51:01] didn't get any tasks after that. I [51:03] clicked the button that I created that [51:05] triggers the same web hook another two [51:07] times and also it didn't get any tasks. [51:09] Can you check the logs and see what [51:11] might have gone wrong there? Um and then [51:14] yeah identify whether there's any [51:15] specific fixes that we need to do or [51:17] what the cause is for that. Uh once we [51:19] have this kind of bug resolved and fixed [51:21] then we can document everything here in [51:24] the project. So we have a complete [51:25] handoff for next session and wrap this [51:28] up for now. And then this is sort of [51:30] like otherwise right if you wanted to [51:32] wrap this up now or like you know like [51:34] um started in a new session debug [51:35] hunting you would simply ask it to hey [51:38] you know um we actually want to do a few [51:40] more steps but the context is getting [51:41] quite heavy. and you just write yourself [51:43] sort of like a handoff document and drop [51:45] it into the project and then kind of [51:46] just print out for me a prompt that you [51:48] want me to use to start in the new [51:50] window. So yeah, we see four new web [51:52] program runs all exit zero and now it's [51:54] going to read the logs and that's the [51:56] great thing right notion workers [51:58] automatically log um what they do uh so [52:00] you can see okay where is something [52:02] going wrong for now right uh you always [52:05] need to use cler codeex to access that [52:08] what you could of course also do if you [52:10] wanted to see those runs in notion you [52:12] could create a database for worker runs [52:15] and then as part of your setup right you [52:17] could tell cloud always hey and make [52:19] sure that whenever you know a worker [52:21] runs it logs sort of those details there [52:23] if it makes it easier for you to see [52:25] that. All right, so let's see what it [52:27] found. Um, all four runs quite [52:29] correctly. Perfect. Second test, no [52:32] type, has no project type. That is [52:33] correct. Um, third test. Ah, it has no [52:36] start date. That's the one that [52:38] surprised you. Um, if it has a project [52:40] type but no start date, it treats it as [52:42] mandatory. So that's good to know, [52:44] right? Um, that's sort of like a [52:45] behavior where we didn't quite know what [52:47] to do yet. So now we want to say okay [52:49] yes we want to have a fallback even if [52:51] there's no date we want to sort of like [52:52] fall back um to today cuz maybe someone [52:55] didn't fill it out correctly. But of [52:57] course that's sort of the strength of [52:59] building this in code in a no code tool. [53:02] You can also have all that logic but the [53:05] more edge cases you have the more [53:07] complicated it gets. And at some point [53:09] you're then like hey team okay we can't [53:11] handle all these edge cases I need you [53:13] to you know put the data in correctly or [53:14] it will fail. with code of course you [53:17] also will have an issue at some point [53:18] but it's much easier to do this for [53:21] example you could simply say hey okay if [53:22] that's the case actually also please set [53:25] that right if you know for example based [53:27] on other documentation in your system [53:29] you know this type of project should [53:31] always run 8 weeks and if the person [53:33] hasn't you know set that yet just take [53:36] today and that or you know ping [53:38] somewhere write update a status add a [53:40] task to someone in the system to do this [53:42] right any kind of business logic you can [53:44] codify here in your worker to make sure [53:47] that goes through. So, uh, fixed it, [53:50] right? And now we can kind of, uh, check [53:52] this again to see whether it works. By [53:54] the way, one quick trick here. If you go [53:56] down below, you see how we're currently [53:58] running. And you see if I hover over [54:00] this, I can see my usage. I can also [54:03] just hover over it. Same idea. Uh, so I [54:05] see like, okay, on my plan, I currently [54:06] have sort of like 30 um% right of my uh, [54:10] 5 hour limit here use because I have a [54:12] few queries running in parallel on the [54:13] side on top of this one. and we see the [54:15] context window of this specific query. [54:17] Currently sits at 135,000 tokens. Now, [54:20] in theory, we have 1 million to 2 [54:23] million uh of context window with um [54:26] OPUS, but um based on uh my own [54:29] experience and what I've seen uh other [54:31] people do that, particular ones that [54:33] like do proper engineering with it, um [54:36] 120 to 150 seems to be the kind of sweet [54:40] spot um where you want to think about [54:41] handing it off, right? So 135K, that's [54:45] still good, right? We just need to [54:46] create the documentation now. We don't [54:47] need to do hard work. But if you were at [54:49] this point now and say, hey, okay, [54:50] there's a whole new logic step that we [54:52] need to build. We need to do a few more [54:53] questions back and forth. That would [54:54] definitely be a signal for me to say, [54:56] okay, let's hand off to a new session. [54:58] Um because while it can process more, [55:00] right, you're diluting this right now. [55:02] There's all this information that floats [55:04] with it. You're using up your limits [55:05] much quicker. Um your response quality [55:08] will go down. So it's best to have it [55:10] externalize what happened so far and do [55:12] that. But yeah, for now it deployed [55:14] everything. So we can go back and test [55:15] it. Here is our fourth test. Right, we [55:18] can set this to in progress. I'm going [55:20] to uh investigate. Right, this also [55:22] doesn't have a a start date set. So [55:24] let's see whether that runs through [55:26] correctly. It also doesn't have an owner [55:28] set. I don't think that should be an [55:29] issue because we told it, hey, if [55:30] there's no owner, fall back to the [55:32] person who created it. And yeah, indeed, [55:33] it's running through correctly. But you [55:35] see already, right? Like there might be [55:36] depending on your system a few things [55:38] that you might need to keep in mind. So [55:39] that's why it's important to test to [55:41] iterate and then work together with [55:43] claude and the logs to figure out okay [55:44] where does something go wrong. But this [55:46] is perfect. So now let's go back simply [55:48] to claude and tell to you know like [55:50] great job. Please document everything. [55:51] Wrap this up and we can start a fresh [55:54] session. Amazing work. This uh works now [55:57] flawlessly. So at this point I would say [55:58] let's make sure we document everything [56:00] thoroughly in this project so that next [56:03] time when we come back to it we know [56:04] exactly what the state is, what our [56:06] decisions were and if you want to expand [56:07] on it like how we can do that. Once [56:09] you've done that we can then uh wrap up [56:12] this session here. All right and then [56:15] we're going to send this. Now one of the [56:17] things that it keeps asking me right is [56:18] whether it wants me to commit this [56:20] because that's because I'm using Git and [56:22] GitHub to kind of back up everything and [56:24] do version control. But again, this is [56:26] something for more like the proper [56:28] engineering flow also like the the vibe [56:30] coding engineering flow that we are [56:32] following internally again right if you [56:34] would like to see like a more detailed [56:35] tutorial like a more advanced one on [56:37] that let me now know down below in the [56:39] comments we might also run a workshop on [56:41] this actually at some point um but yeah [56:44] so um it's going to go ahead and do that [56:46] and that means our first chapter is [56:49] wrapped up at this point you should have [56:51] your first working worker right that [56:53] gets triggered by web hook that does [56:55] something in your workspace, you can [56:56] codify any kind of complex business [56:58] logic into it. And yeah, then we're [57:00] ready for the next one. Onto our second [57:02] project, sync workers. As a reminder, a [57:06] sync worker is one that that you set up [57:08] and say, "Hey, on a cadence every 10 [57:11] minutes, once a day, once a week, go to [57:13] this external data source and pull the [57:16] data into notion." The specialty about [57:19] this one is that this creates this this [57:21] kind of work. If you call it a sync [57:23] work, right? If you define that uh it [57:25] means it needs to manage its own [57:27] database. So rather than creating a [57:30] database like we had already right with [57:31] our tasks and etc right it will create a [57:34] new one. There's currently no way to [57:35] tell it hey I actually have this already [57:37] please just update that. Hopefully that [57:39] comes in the future. For now this will [57:41] create a new one. So you might have to [57:43] migrate other data in there or like you [57:44] know shuffle things around a bit. The [57:47] cool thing about this one is that the [57:49] database that it creates is managed by [57:52] the worker which means every property [57:54] that is managed by it cannot be changed [57:57] in notion. You can't change the values [58:00] of the entries. You can't change the [58:02] property types etc. it's sort of [58:04] protected in the UI which means right [58:07] for this oneway sync you're always [58:08] guaranteeing that this data looks [58:10] exactly like in the external data source [58:12] and you're not risking that hey you know [58:14] there's something wrong with the with [58:17] the setup um you can of course add your [58:20] own properties to it so for example uh I [58:23] have a database that syncs my new setup [58:25] from convertkit in there and I have all [58:27] the you know properties from the worker [58:29] for the details and analytics and then I [58:31] can relate it to my actual content and [58:34] can add additional properties if I [58:35] wanted to. We'll see in a second how [58:37] this works. The important thing to know, [58:39] right, is that this is creates this [58:40] oneway sync with a manage database. Now, [58:43] just as a side note, as of filming this [58:45] tutorial, this is the only way to inside [58:48] of the worker declare a timer, right? A [58:51] so-called chron job. That's basically [58:53] whenever you say, "Hey, every 15 [58:55] minutes, every 30 minutes, run this." In [58:57] a normal worker, right? one that doesn't [59:00] manage its own database, you can't do [59:02] this. But there might be plenty of [59:04] situations where you want that. For [59:06] example, maybe you want to create a [59:08] worker to build actually a two-way sync, [59:10] right? This is nothing that we're going [59:11] to look at here. This is a lot more [59:13] complex, but you might want to say, hey, [59:14] I want to update be able to update also [59:16] data in notion and then it should update [59:18] that in the external source, right? We [59:19] want it in both ways. Then you can't use [59:22] a sync worker right now because remember [59:24] you can't change values in in your [59:26] system. The workaround for this just so [59:28] you further ones would be to take the [59:30] first type the web hook worker and then [59:33] in notion you can fire a web hook on a [59:36] schedule right so you can say hey every [59:38] 15 minutes fire this web hook so through [59:40] that you can basically you know build a [59:43] chron job sort of like a chrome job at [59:45] home um to make sure that your worker [59:46] does that so basically if you say hey I [59:49] want something a worker that runs on a [59:50] schedule but I don't want this oneway [59:52] sync database and managed stuff um go [59:55] with approach number and simply trigger [59:58] a web hook on a SC cadence. But for this [60:00] example, what we're going to build is a [60:02] system where we're going to pull an [60:04] external data source. I think Hacker [60:06] News should have a good one. We'll check [60:07] that out in a second. And basically, you [60:09] want to have always the updated list of [60:11] top hacker news stories in our system. [60:13] But this could you be your external CRM, [60:15] right? So if you um manage deals in [60:18] something like HubSpot or so, you could [60:20] pull it in. If you have a product and [60:22] want to pull from your product usage [60:24] data into notion that would be a great [60:26] use case or again pretty much any [60:28] external data source for personal ones a [60:30] lot of people think Spotify or like you [60:32] know like the health data from Uring [60:35] wherever you have data externally and [60:36] want to make it accessible this is the [60:38] way so fresh window in clawed uh code [60:41] here again right just going through the [60:44] flow now and I'm describing how this [60:46] works so I'm putting on dictation and I [60:48] explain what I wanted to do this is a [60:50] new folder in right separate worker, [60:52] separate folder, data sync. Um, and it [60:54] doesn't have the data structure yet. So, [60:55] I'm going to ask it to do that. All [60:58] right, cloud ready for another project. [61:00] Uh, this time we want to build uh [61:02] another notion worker, a sync worker. In [61:05] order to set this up, I would first like [61:06] you to create um besides the readme [61:10] which is already in this project, our [61:12] learnings MD and our um change log. So, [61:15] just so we can keep track of everything [61:16] that we're doing as we go through this [61:18] project, right? Please keep this always [61:20] up to date. Once that's done, then uh [61:23] I'd like you to ask me questions about [61:26] the project that I'm planning to [61:27] implement here. Right. What we basically [61:29] want to do is we want to use notion's [61:31] sync worker option to pull an external [61:34] data source. I'm thinking of pulling in [61:36] hacker news. I think they have an open [61:38] API that we can pull. We could pull that [61:40] once a day to create a database to [61:44] update uh that database in notion. um [61:46] the worker will have to have its own [61:48] managed database. If I recall correctly, [61:50] I've already set up a sort of like [61:52] parent page uh where you can drop that. [61:54] But yeah, that's so far what I have in [61:56] mind. Again, please ask me clarifying [61:58] questions so that we know exactly what [62:00] we're working on. And then once we're [62:01] ready with everything, we can go to [62:03] implementation. You should have already [62:05] um access and log into um notion via the [62:09] CLI. It should already be confirmed. But [62:10] you can check that if you want, right? [62:12] Again, it's like you hand off a project [62:14] uh to a human on your team. You just [62:16] give them some context. And if you do [62:17] this often enough, you then wrap this in [62:20] skills, right? So you don't have to [62:21] repeat yourself all the time. So just [62:23] going to fire this off, right? And see [62:26] what it comes back with. So the usual [62:28] game, it's come back comes back with a [62:30] few questions for us. Which hacker [62:32] newsfeed should the worker hold in [62:33] roughly how many stories? U let's do the [62:35] top front page. 30 sounds totally fine. [62:38] Although let's actually just do the top [62:39] 10. And this is where like other comes [62:41] in, right? Sometimes these options are [62:42] very short and they don't really [62:44] describe what you want. Um so it's a [62:45] good idea to then use other than just [62:47] dictation. In this case um let's do uh [62:50] front and only the top 10. Right? How [62:54] should notion database behave from one [62:56] daily run to the next? It will yeah we [62:58] want to mirror the current list. Right? [62:59] So this is now the question right? Your [63:01] sync worker can either just replace [63:03] things right. So it just shows the [63:05] latest 10 or if you always want to copy [63:07] them, right? Um it can uh do that as [63:10] well, right? And uh accumulate. But here [63:11] we want to replace. So I just want to [63:13] see always the latest top 10 stories. I [63:16] basically want to have hacker news in [63:17] notion. How cool would that be? Where [63:19] should we fetch the data from? Um I [63:23] don't know, right? Like this is a [63:24] question. I I I don't know which one is [63:26] more efficient. This will happen quite [63:28] often, right? where um sometimes when [63:29] you build something more complex, uh [63:31] Claude is going to ask you engineering [63:32] questions. And in my proper engineering [63:35] setup, I have kind of like a a long list [63:38] of or like a thorough context document [63:40] that kind of walks it through the [63:42] decision tree where I spell out very [63:43] clearly, hey, um you are the senior [63:45] engineer. I'm kind of product, right? If [63:48] you look at it in a traditional company [63:50] structure where it's always like there's [63:51] someone, right, who figures out, hey, [63:53] this is what we should build and this is [63:54] what the users are asking for. And then [63:56] there's the engineering side which [63:57] actually makes it happen. The relation [63:59] here if you as a non-engineer build with [64:02] claude or codeex is very similar right [64:04] you are stepping into the role of [64:06] product. And so sometimes you need to [64:07] remind claw that hey when it comes to [64:10] engineering questions you probably can't [64:12] answer them. So it's also important that [64:14] you shape your you know sharpen and [64:16] train your own perception and muscle [64:17] here that you don't give input on things [64:20] that you can't right. If you want to, [64:22] you need to get the context, right? You [64:23] might open a new window and ask, you [64:25] know, Claude, hey, what are the [64:26] trade-offs between the these so that you [64:28] can make a judgment call or in something [64:30] like this, right? I'm basically just [64:31] going to tell it um you are the expert [64:33] on this, right? It's your call. I don't [64:35] really care. Whatever you think is more [64:37] efficient. [64:41] And then beyond title source and the [64:43] hacker link, hacker news discussion [64:45] link, what else do we want? Um, sure. [64:48] Let's do the the the count. Let's do the [64:50] author. Um, [64:53] post the time. Let's get it all. Right. [64:55] Let's get it all. Uh, and then I don't [64:57] actually have no idea how this API looks [64:58] like. Is there also a summary also uh or [65:03] the actual uh content? No worries. Uh, [65:07] yeah. Not but just curious, right? So, [65:10] like how that uh how that looks like. [65:12] And then we're going to submit. Oops. [65:14] Did manage to submit without clicking [65:15] the button there. We want to submit our [65:17] answers. Apologies for just flashing big [65:20] on your screen. So yeah, and then again [65:22] wait a bit, right? As you see, this took [65:24] it five minutes for the first one. So [65:25] again, you might uh not need to do some [65:27] other things in between like circle [65:29] between a few workers that you're [65:30] building. Uh it's a kind of new way of [65:32] working where you invest a lot of time [65:33] into the planning and the execution is [65:35] taken care of by the system. Similar to [65:38] actually what we did uh in our [65:39] development guide, right? If you're [65:40] curious on how to build notion [65:42] workspaces with a similar AI first [65:44] approach, check out our CRM guide down [65:46] below. All right, that was quick. After [65:48] answering these questions, it decided it [65:50] has everything it needs. I simply went [65:51] ahead and built this. So perfect. Um [65:54] asking again whether I should deploy it. [65:55] So yes, I want to deploy it and run it. [65:57] Uh and then see how this looks like in [65:59] the workspace. And you see right like [66:00] this 10 minutes that was just popping up [66:02] there, right? So it's been working [66:03] basically on this now for for [66:05] consecutive time of 10 minutes. If we [66:07] check our context window where we're at [66:09] um blah blah blah, right? 86k. And that [66:12] shows you right you can basically for [66:14] these simple workers you can oneshot [66:16] them in one session. If you do more [66:18] complex integrations you might want to [66:20] do a few rounds and then go into proper [66:22] development. But again for something [66:23] like this it's so accessible now to [66:25] create these even as a non-engineer. [66:28] It's actually crazy right um and this is [66:30] not 10 minutes of active time right this [66:31] is sort of like 10 minutes of clots time [66:33] and then a few minutes here and there [66:35] for us to answer questions. All right so [66:38] it's deployed now let's check in notion [66:40] how it looks. All right, and here we go. [66:43] Our database with the top 10 from [66:44] hackernews of the day. And you see, [66:46] right, this is what I meant. This work [66:48] database is synced. So while I can click [66:50] into this and open it and then of course [66:52] I could sort of like oops create a [66:53] template or I can also of course type in [66:55] the page body because that is not [66:56] synced. I can't change any of the [66:58] values, right? I can't say actually this [66:59] is rank one, right? Like I can't click I [67:02] can only copy it. Same for this. And [67:04] this is sort of the power of these [67:05] databases, right? Where you protect them [67:06] from accidental changes in the system. I [67:09] can then go ahead, right? And I can now [67:10] add new properties to this. So I can [67:12] say, okay, you know, let's, you know, [67:14] move this a bit. If for whatever reason, [67:16] right, we wanted to link this to our own [67:17] internal database, right? For like some [67:19] insights, we could do that. Or we could [67:21] add a, you know, a tax property, right? [67:23] And then if we weren't resetting this [67:25] every day, right, we could now create [67:27] tax here. And we could, for example, [67:28] create a notion agent that actually goes [67:31] in and reads the blog post, right? Um, [67:34] and then adds the text to this to make [67:36] sure like we have the context from it. [67:38] Or we could create a worker, right, that [67:40] actually now as a follow-up, right, goes [67:41] in, takes the content from this post and [67:45] drops it into the page body, right? [67:46] There's so many extensions to this, but [67:49] that's the basic premise of a sync [67:50] worker. We can then tell it like how [67:52] often it should refresh it, right? In [67:53] this case, it's every uh every day, but [67:55] you could run this, of course, right? If [67:56] it's a sync for like data that changes [67:58] often, every 15 minutes uh much more [68:01] frequently and have everything in here [68:03] in motion in basically more or less real [68:05] time. On to worker number three, giving [68:08] AI tools. This is all about upgrading [68:11] and upskilling your notion custom [68:13] agents. Because while they're great at [68:15] doing the natural things within notion, [68:17] there's basically two areas where giving [68:20] them tools can be really impactful. One [68:23] is expanding their capabilities. So kind [68:25] of the things that we talked about [68:26] before, right? Pulling data in, pushing [68:28] data out. If you want to interact with [68:30] an agent while doing that, right? This [68:32] is what you need. For example, let's say [68:34] you have a data warehouse. somewhere, [68:36] right, for your company with tons of [68:38] information and you want you have a [68:40] custom notion agent where that people [68:41] can ask any kind of data query so they [68:43] don't have to, you know, go to a data [68:45] analyst so they can focus on actually, [68:47] you know, cleaning the data and prepping [68:48] it properly. Well, you can give your [68:50] notion agent a tool that allows it to [68:52] call your data warehouse and do that. Or [68:55] you could give it a tool that helps you [68:57] automatically, you know, like create a [68:58] newsletter, update a blog post on the [69:00] web, etc., right? all things where [69:02] otherwise you would have to first [69:03] interact with AI in your workspace and [69:04] then still manually take those actions. [69:07] That's kind of one of the big blocks. [69:09] The other big block where tools for your [69:11] notion agents are very relevant is to [69:13] optimize for cost, right? Because when [69:16] an agent runs, it will consume tokens [69:18] and the more information it needs to [69:20] consume to do something, the more [69:21] expensive it gets. Often times though, [69:24] parts of what it does can be outsourced [69:26] to deterministic code. For example, [69:29] let's take dduplication. If you want to [69:32] dduplicate uh your notion database, you [69:35] could of course say, "Hey, agent, you [69:36] know, query this, check for similar [69:37] names, etc. But then it needs to go [69:40] through all of that, right? That would [69:41] be immensely expensive. It would be a [69:43] lot more efficient to give your agent a [69:45] tool, a bit of code, right, that [69:47] basically runs through database and [69:49] based on code, deterministic logic, [69:51] right, determines, hey, you know, these [69:52] are people with the same name, these are [69:54] people with the same email, right? These [69:55] are duplicates." And then the agent just [69:58] sees the analyzed duplicates and then [70:00] can use its reasoning step to say okay [70:03] how are we going to merge that together [70:04] right then you kind of get the best of [70:06] both worlds code to do the heavy lifting [70:09] and then agent reasoning to do the [70:11] actual thinking job of it. So in this [70:13] case, we're going to do like a very [70:15] simple um um example, right? Just so you [70:18] have built your first tool for a worker. [70:20] We're going to create a custom agent. [70:22] And this custom agent um will be able to [70:25] fetch the Wikipedia API, right? So to [70:27] just make it a bit easier to get some [70:28] structured data back and then populate [70:30] that in our notion database. So same [70:33] idea as before, right? We're going to [70:34] clot, we're going to describe what we [70:36] wanted to build and then we see like [70:38] what is this the unique element, right? [70:40] If we need to, you know, give this to a [70:42] worker in notion, you know the drill for [70:44] these prompts right now, right? [70:45] Basically going to tell it, hey, I want [70:47] to set up a custom agent tool as a [70:49] worker. Um, this is what it wants to do. [70:52] What I just also described here, right, [70:53] here's a database that these things [70:55] should be pushed in. And if you have any [70:56] questions for me, right, again, please [70:58] ask me. So you also one thing that you [71:01] see is um since we are working in a [71:03] folder, right, where there are other [71:04] folders already with workers, uh, the AI [71:07] is smart, right? So it's going to check [71:08] out what it did before there. So some of [71:10] the learnings even though they're kind [71:12] of like constrained to an individual [71:13] project will carry over and if you build [71:16] more and more workers you can of course [71:18] start to abstract learnings right. So [71:19] you might create a big learning stock [71:22] and then occasionally ask you know [71:24] claude to go through the individual [71:26] projects that you created and sort of [71:27] figure out hey what of this knowledge [71:29] should we abstract away so that it's [71:30] just easier when we build new things. [71:33] Again, for these simple ones, that [71:35] doesn't really matter. But let's say, [71:36] for example, you integrate this with [71:38] custom tooling from your company and [71:40] your internal API has some quirks, [71:42] right? Very helpful that you don't have [71:44] to reexplain this every single time. So, [71:46] what should the tool actually do when [71:47] the agent calls with a topic? Um, tool [71:49] fetches Wikipedia detail, creates a row [71:51] in a glossy, yep, and returns a [71:52] confirmation. Exactly. Right. The agent [71:54] also doesn't need to do the right step. [71:55] That's why we want to optimize the [71:57] tokens as much as possible. So, I can [71:58] just tell the agent, hey, can you fetch [72:00] me details on this? and it does it as [72:02] efficiently as possible. Which Wikipedia [72:04] endpoint? Um yeah, that sounds good. Um [72:08] let's do queen structure data ID for [72:10] glossery, right? Let's follow the [72:11] recommendations. And how should we [72:13] target the glossery database? Um we want [72:16] to write to an existing database, right? [72:18] So important. So basically asking, hey, [72:20] should this be a worker managed one? So [72:21] like the sync one, no, we don't want the [72:23] sync um manage one. We have a database, [72:25] right? It should use the existing one. [72:27] Um that's why it's so important right to [72:28] go through these questions so that the [72:30] AI actually ends up building what we [72:32] want um rather than something else which [72:35] also right as a general principle the [72:36] more complex the thing that you're [72:38] trying to build the more time you want [72:40] to spend on this plan so we're rushing [72:41] through this here right because it's a [72:42] demo it's a tutorial it's meant to give [72:44] you a bit of exposure to your first [72:46] workers when we build more complex flows [72:49] right for client projects or internal [72:51] ones it might well take half an hour an [72:53] hour two hours right and to go through [72:56] this grilling session and figure out hey [72:57] what is it exactly that we need what are [72:59] the rules what is version one right that [73:01] we want to do in the beginning what are [73:03] things and features that we you know um [73:06] you know will ship later so yeah uh [73:08] let's let this run through see whether [73:10] it has any other questions and otherwise [73:12] it's probably going to go ahead and [73:14] build this directly while we're waiting [73:16] for a moment and sort of like 7 minutes [73:19] later we see it's basically done right [73:21] it's sort of like updating its [73:22] documentation as always and it tells me [73:24] hey the the deploy right this block [73:26] because that's why we we're still on [73:28] auto. It can't do everything without our [73:30] our input. So, it's going to finish [73:32] writing this up and it's going to ask [73:34] me, hey, should I should I execute this? [73:35] And again, let's just check just check [73:37] like how much context it used, right? Uh [73:39] still well below our 120 to 150k uh cut [73:44] off for this oneot setup. So, pretty [73:46] perfect. Uh one option of course right [73:48] is you can always if you run into [73:50] context limits and usage issues you can [73:52] go down from you know extra high to high [73:55] or medium but in particular right if you [73:57] don't have too much experience with [73:58] guiding AI I recommend to like you know [74:00] increase the setting there to make sure [74:02] you have sort of guidance from the [74:04] system on what you're doing. A good 15 [74:07] minute later, the deploy and everything [74:09] came through uh and all the updates. And [74:12] now I'm simply asking it to, hey, can [74:14] you give me a prompt that I can paste [74:16] into notion AI to set up the agent for [74:18] that? Now, this is sort of like a trick [74:20] that sometimes comes in very handy [74:22] because we just gave this session here a [74:25] lot of context, right, on what we're [74:26] trying to build. And now on the notion [74:28] end, I don't have a custom agent for it [74:30] yet, right? Many times you might already [74:32] have a custom agent that is set up with [74:34] prompts and skills and has everything it [74:36] needs to know and you just need to give [74:37] it a tool and expand it. But in a [74:39] situation here where you sort of first [74:41] create a worker and then say who okay [74:43] now I need an an agent to handle it. One [74:45] of the easiest way to do this right is [74:47] basically ask the AI here to give you [74:50] sort of like a prompt, right? So we can [74:52] just copy what it gave us here and [74:54] because it wrote the code uh [74:56] instructions and knows exactly like what [74:58] it's supposed to explain. So, I'm going [75:00] to copy that. I head back over to Notion [75:03] where we have our demo research base. [75:04] And I'm going to ask Notion AI here, [75:06] hey, you know, can you help me set this [75:08] up? So, never run this on auto, right? [75:10] Always on the max model. Um, I'm just [75:13] going to explain this. All right. I need [75:15] your help to create a notion custom [75:18] agent for me for this YouTube demo [75:20] purpose. Basically, we've created a [75:21] worker that is a tool for an agent to [75:23] call Wikipedia and enrich these research [75:26] topics. I've also asked Claude to just [75:28] build this worker to give us a little [75:30] bit of context, right? So you know [75:32] exactly what to build. Basically, I only [75:34] ever want to chat with this custom agent [75:37] um to do this, right? Because you just [75:38] want to demo uh what is possible uh [75:41] here. So nothing too fancy. I will wire [75:43] up the actual tool, but if you could [75:45] create the agent for me with all the [75:46] instructions, that would be amazing. And [75:49] yeah, just leave it and uh have it run [75:52] cuz that's the easiest way to set this [75:54] up. Of course, we could go in the UI. We [75:56] could do this on our own. Also, a quick [75:59] side note, um, this alone, right, this [76:01] is a demo use case. You probably [76:03] wouldn't create this as an actual custom [76:06] agent in your space, right, to just call [76:08] Wikipedia and pull this out. So, you [76:10] know, think bigger here. Think about [76:11] like other external data sources that it [76:13] might want to pull from, right? Maybe [76:14] you need to make some sales decision. [76:15] Maybe you're not syncing a external data [76:18] source into notion with a worker type [76:19] two. you just want sometimes your agent [76:22] to be able to to access that data or to [76:24] push somewhere. So no, don't treat this [76:26] as a best best use case to actually [76:29] create a notion agent. So let's wait for [76:32] this one to run through and then come [76:33] back and wire everything up. So our [76:35] agent is ready. Right here are all the [76:38] details. Now again, right, this [76:40] typically wouldn't be an agent. Just as [76:42] a side note, I hope Notion changes this [76:44] in the future, but right now you can [76:46] only attach tools to a custom agent, not [76:49] to your personal one. Otherwise, an [76:51] interaction type where you say, "Hey, [76:52] I'm only going always going to open this [76:53] in chat, right?" This should not be a [76:55] custom agent. Usually, this should just [76:57] be a skill for your personal one. So, [76:59] you don't have to pay for tokens. But [77:00] there's one exception right now, and [77:02] that is if you need to attach workers [77:03] because they're only available here for [77:04] now. Um, in order to do that, right, um, [77:06] we can go down here now and click on add [77:08] connection. and the connection will show [77:10] up down here, right? As a tool. One [77:13] little quirk here. Um, when you go here [77:16] and try to look for workers and you [77:18] don't see it, it might be because you're [77:20] not a person who deployed the worker. [77:22] Right now, only the person deploying the [77:23] worker can also attach it to agents, [77:26] which means you can't yet build sort of [77:27] tools for your team and then they attach [77:29] it to whatever agent they want. They [77:31] would always have to share that with you [77:32] and do it. There's no way to sort of [77:33] like add permissions yet. Hopefully, it [77:36] comes in the future. Also just a side [77:38] note why we didn't do a good job naming [77:39] this here but you know that's for the [77:40] future. Let's connect this and then when [77:42] we connect it we see okay there's a [77:45] worker and there's one tool. So one [77:47] worker can contain a whole section of [77:49] tools which is relevant. For example [77:51] let's say we go back to our data [77:53] warehouse example and we say okay this [77:55] agent should have you know like 15 [77:57] different SQL queries that it can run [77:58] against our data warehouse. In that case [78:01] you wouldn't need 15 different workers. [78:02] You create one worker and then that [78:04] worker would have 15 different tools and [78:07] then you can also like on a one-off [78:08] basis decide hey should this agent have [78:11] access to all these tools only to some [78:13] and if we update this uh worker in the [78:16] future right should it automatically get [78:17] new tools or do we want to have to opt [78:19] them in so yeah connections basically [78:21] work just the same way as custom MCPS [78:24] here. Um one more setup step right I [78:26] always start running my agents on Opus [78:28] 4.6 just to make sure that you know if [78:30] we test this on a or 4.7 because if the [78:33] smartest model can't you know handle [78:36] your instructions and your tools then [78:37] you have a problem then you need to go [78:39] back to a drawing board but once it [78:40] works we can then go back and make some [78:42] simpler changes to it. All right, let's [78:45] save this and then let's test it. Can we [78:47] test your new tool with a fun query on [78:52] Wikipedia? [78:53] You pick whatever you like and then we [78:57] have it run and now right we'll see like [78:59] how this works and how it goes and what [79:02] it decides to call from. One interesting [79:05] aspect notion yeah we need to like [79:07] confirm two calls if we run them like [79:09] this. If they run on a schedule right it [79:11] will run without confirmation. Um we can [79:14] then also see like what tool it called [79:16] right in the moment. So it sees okay [79:17] runs that via the tool. Okay tested on [79:19] kabibara awesome. Um, let's see the [79:22] glossery entry, right, that we have [79:24] created for that. Let's see. Is that the [79:26] link that goes to it? No, that's the [79:29] Wikipedia link. Let's go back to it and [79:31] figure out where that is in notion. Ah, [79:34] there we are. Right. So, our copy driver [79:35] is in there, right? We see the summary [79:37] I've got from Wikipedia, the URL, um, [79:39] the thumbnail, right? And the last [79:41] enrich step. So, ran through perfectly. [79:43] Right. Now you could again expand on [79:45] this, have it do more, but if you follow [79:48] along so far, you've successfully [79:49] deployed your third worker. So now you [79:52] have a worker that runs with a web hook. [79:53] You have a worker that runs on a [79:55] schedule, and you have a worker that can [79:57] be invoked by an agent. So much for our [79:59] big notion workers beginner tutorial. [80:03] Congratulations, right? Even if you've [80:05] never touched a line of code before, you [80:07] now have three shipped workers that can [80:09] actually do things in Notion. From here [80:12] on the world is really your oyster and [80:13] the only limit is your imagination. What [80:16] do you need to happen when you know [80:18] business processes run through notion? [80:21] What is something that you wish could be [80:23] pulled in or pushed out of notion? [80:25] Whatever it is, you now have the tools [80:27] to build it. As always, if you have any [80:29] questions around this, let me know down [80:31] below. If you want our help as notion [80:33] consultants, you know where to reach us. [80:35] And other than that, I highly, highly [80:37] recommend that you check out this video. [80:40] In this video, we explore how you can [80:42] teach AI hard things, most notably [80:45] things that you are not really good at [80:47] yet. It's one of the most complicated [80:49] things to do, but also if you figure it [80:50] out, right? If you can use AI to cover [80:53] up your weaknesses, then you have a new [80:55] superpower. So check this one out and I [80:58] will see you in a few