Transcription
I let my co-founder roast my Claude code setup live in this video, and he didn't hold back. So, 5 months in, I thought I had Claude Code pretty dialed in. But it turns out my claw.md file is missing the most important thing. My folder structure is making my life harder, and I'm still manually triggering a whole bunch of tasks that Claude could be doing automatically for me.
So, in this video, we're going to go through the whole audit, the specific fixes I'm making this week, and the one rule Gail says every business owner should add to their claw.md so it maintains its own documentation as you work. We're Mark and Gail from Authority Hacker, and we teach real-world CLA to busy online business owners.
All right, Mark. So, we are reviewing your setup today, and let's just start with like, what are you using code code for day-to-day? Like, I want to, you know, I want to see what you use it for and how you use VS Code, how you do all that stuff, and let's try to modernize your setup a little bit.
Okay. Sure. So, I do a lot of one-off kind of marketing businessy type tasks. So, think if we're doing a promotion with a different community or a JV webinar or a launch or something like that. It tends to be quite unique. So, I'll use it for kind of organizing the project and for, you know, creating emails, creating marketing materials and that. I use it a lot for kind of troubleshooting AI when I'm trying to set up. A good example would be the Meta Ads Conversion API or their new Meta Ads CLI, like figuring out how to set that up because it wasn't super obvious.
Building some basic automation things, for example, welcome email sequences to new members, which, you know, it's mostly creating the content for that rather than the full end-to-end automation. Uh, I've got some regular processes, I would say I do, you know, almost every week or every couple weeks, things like uploading the podcast. So, creating thumbnails, titles, descriptions, you know, chapter markers, like all those kind of things. I do, uh, our email newsletter, some customer support things. Again, that's not really well kind of integrated into all our tools. I'm still kind of copy-pasting.
That's why I want to dig. That's why I want to dig. [laughter] In our, uh, AI accelerator for the plus members. Um, when people apply for that, we review them manually. So, I'll go and kind of use AI to help me. I mean, I still look at it, but I use AI to help me kind of draft the responses to that and whatnot. Uh, all of the changes on our website, so new landing pages, you know, changes to the sales page, things like that, I do using cloud code as well.
Do you run everything in VS Code? Is everything running through VS Code?
99% of the time I use VS Code. So, I do have the Claude desktop app, and I do have Claude code set up on the same folder working, you know, with all the same information. I just find myself not using it that much. Maybe sometimes when I...
I don't even know why, to be honest with you. The main times I use it are when I have so many tabs open in VS Code that I'm like, "Ah, I don't want to lose everything." I'll just use Claude desktop for a while.
Interesting. Okay. Okay. Cool. I think we'll talk about that. I think it's quite interesting to talk about the desktop app versus VS Code as well, because things have changed quite a bit recently.
Yeah. Uh, and actually, my recommendation is most people start on the desktop app right now. I'm actually going to update our course soon based on that.
And the reason why is because you see how, like, let's say you're working in VS Code now, and your folder is called "code," but that means you're basically dumping everything you do in this folder, right? And eventually, this folder gets messy, and we'll talk about organization, etc. But obviously, you're going to have a lot of your stuff in there. But like, the reality is, like, your job is many things, right? Is some support, some website stuff, some marketing, and within marketing, there is email, and there is ads, and there is, you know what I mean? Like, you could kind of like break it down.
And, and the, you know, where the code desktop app wins over VS Code is in VS Code, you're working in that one folder here that you open. So, you kind of have to dump everything. Whereas on the Claude desktop app, you can open multiple folders, and they all run in parallel, and you can essentially just chat with either folder depending on where you click. And what that means is you can kind of be more granular, and then you can do more customization in your code.md and in how things work in the context, etc. And you can start organizing things a little bit differently. And so, I tried to do that. I mean, I set up VS Code before the Claude code and the Claude desktop app was really a thing. I tried to do that with this areas section here.
Yeah. Uh, so I have my...
And you were opening the subfolder?
Initially, I was, but then I got lazy and stopped doing it. So, for things like podcast or even marketing in general, email marketing, I would have like different claw.md files here.
And then in my root claw.md folder...
It would have, it would kind of like mention the different directories, and, you know, some of them have area context, some have different claw.md files. So, I mean, that was the thinking. Like, honestly, it's not well-maintained. And...
And then you end up with like, 60 like windows of VS Code all floating on top of each other, and your desktop becomes overwhelming, and you're like, "This was supposed to be simpler." And...
Yeah. And I think what's happening now, so I'm 5 months into it, something like that.
Yeah.
Is it's still doing a really good job. Like, it, for whatever reason, it understands mostly what we're doing. But I feel like, I feel like we're starting to get a bit of drift.
In that, you know, some things we do in the business have changed a little bit, and I haven't necessarily found out where that's still referenced in my file structure. And yeah, it's just, it's not too bad now, it's still like very, very good. But I feel like if you fast forward another six months, some things might start to get too far out of sync, if that makes sense.
Which is why I'm changing a little bit my opinion. And it's like, initially when we talk code, it was like, "Oh, work within that one folder and keep it organized." But now I feel it's almost better to kind of split it up. And that's where the desktop app makes sense, because then you can make multiple folders, and you can chat with either, right? You can have an email marketing thread go at the same time as a support one, etc. And they all run within their different folders. So, if you...
I don't quite understand. And do you want to show us that, though? 'Cause I can only understand that.
I don't really have the same job as you. So, I don't have as many folders and levels. But like, for example, this is my code desktop app, right? And you can see that, for example, I have the website folder here, and I have like my prompting lab, which is like my historical catch-all thing. And this is my project for my tasks on Notion that I shared on another podcast. Right. So, if I wanted, I could like start a new task on the website. So, I could say, "Change the H1 on homepage to," and let me switch to Opus 4.7, "and then to Hello Mark." Uh, don't worry, I'm not going to commit this. But the point is, like, I can launch this task, and then I can start another thread on prompting lab, and I can say, "Prepare a newsletter sharing the updates at Google IO yesterday." And it's going to call the right skill, etc. The point is, now I have a thread going on in my prompting lab folder.
How does it know which one to use?
The way it works is when you make a new session here, you can pick a folder.
Okay, so you pick each...
And then you can order, you can essentially group by project. And what that does is it just basically creates these groups per project. And so you are able to create smaller folders that contain more specific things, and then each one can have its own code.md, can have all of that, etc. And then the idea is you can just launch...
As many as you want in different folders, but that still operates within one interface, and that's kind of nice. Basically.
That interface. Yeah, I like that a lot. Yeah.
So, do you see the value now when you connect the different folders, etc.? And so, you'll be able to be a little bit more granular, and then you'll be able to spend more time on your code.md for things that matter, and then you can still reference files outside the folder. Like, Claude will still be able to go outside the folder, and you could have like a common ground place where you have some context that it reads, basically, right? So, yeah, that's pretty much, uh, that's pretty much why I think this is good.
What desktop app? Can I just ask, 'cause like when it launched, it wasn't very good at, like, your files?
There's two ways to do that. So, basically, the first one is you can actually click on files here, and then you can see your files. So, I have a little bit of mess, don't be mad. But like, you have this basic file reader, which is like, okay, but not great. But the point is, let's say you want to go deeper. You want a better, uh, file editor, right? You can always right-click here and open in VS Code. And what that does is this opens a new window of VS Code exactly in that folder, and you can resume that thread if you want to. And then you basically are working in the same folder, because remember, whether you work in VS Code or you work in the Claude code app, as long as it's working within the same folder, it's the same thing, right? It's like, it's not the client that matters. That's a really important thing, because one of the questions we get asked a lot, and like, to be honest, I was asking that question in the beginning, is like, "Which should I use, VS Code or..."
You don't have to choose. Yeah.
Claude code, or even like, if you're looking at something like using CodeX from OpenAI, it's a different system, but...
You're not really locked in in the same way, 'cause what you're saying is like, all the files and folders are local, so you just use them in a different app.
And the point is that the way I would say for most people, it's easier is now to kind of connect all your folders to the Claude desktop app, start here. And you're like, "Well, you know what? Like, let's say I wrote me a newsletter. It's not finished now. It's working. But let's say I wrote a newsletter and I'm like, 'Oh, I want to have the thing where I can highlight text. It is what I highlight. I can give feedback, etc.' Well, no problem. Right-click, open in VS Code." And then I can resume the same chat in the VS Code extension, like by going to the chat history, and it will show there because it's in the same folder. And then I can just open the markdown file and just start highlighting text and just commenting on this. And when I'm done, I close VS Code and I resume back here. And that chat is also updated because it's the same chat, right?
Right. Okay. So, yeah, I'm thinking how I would use this then for my work. Like, would you go granular? Like, would you have, you know, a podcast folder, an email folder, or would you have like one level up, like a marketing folder?
Probably a marketing folder. Probably I would do a marketing one, a support one. I, you see that even with all your stuff, it's still decent at dealing with it. So, it's like, if you keep things like as broad areas of interest, I think that's fine. Even if you have a small business, like, it can still be okay to do everything within one folder. But like, the point is, as context grows, this becomes a problem. And so, you kind of need to kind of address. But it's not really a problem, because you know how we teach people to have these areas, right? Like, you showed. Well, you can essentially just branch them off as their own folder, or even just not even move them. You can literally just open them as a folder. So, you just make a thing, and instead of opening prompting lab here, I can just open, uh, if I go in code, let's go prompting lab, and then I could go into an area for example, uh, and then I could open, let's say, website or social content. And boom, that becomes a new folder that runs, even though it's a subfolder. And then, yeah, and that was what I thought I was doing initially in VS Code.
But now you can do it within one window.
Yeah, exactly. Because what happened in reality was I just, I was too lazy to go file open and, you know, and have like many windows at the same time. It's confusing, and you have to, uh, etc. So that simplifies everything, and that's the idea. And then you can have a CL.md that goes live in that social folder, and boom, it can use that context directly. And so that's why I kind of like the idea of like, you can start here, you right-click, you open it, you go, I mean, you see I have many code editors, but if you just have VS Code, it, you just show VS Code, and then you pick that one, and it will just, you find your same setup, and you can resume the chat that was running in the Claude app, and that's kind of the beauty of the ecosystem is it's folder-based, and so the chat history is also folder-based, which means that it will actually work in VS Code.
Yeah. Okay. So, I mean, action for me after we finish recording. [laughter] This is like, I don't actually think I need to create new folders. I think I just subfolders. But if I use them to the cloud app, that's going to give me a nicer, easier-to-use interface in order to access them. And to be honest with you, for a lot of those tasks, I'm thinking, you know, podcast one, there's a lot of image generation, and for thumbnails, it's nicer to use the Claude app because it fully renders, like, in the chat flow, the images, whereas VS Code, you kind of, you can still open it in there, but you have to kind of click to open it, and it's, it's not, not as smooth, not as smooth of an experience.
I agree. So, let's say, for example, like you have this view that you can select here in diff, and diff basically shows you what the agent removed or updated on the file, right? Um, so let's say it created a file. You...
Does that mean, stupid question, does that mean yes? It... [laughter]
And the point is, like, you know how on VS Code you can highlight the text and you can comment, and it's kind of missing in the desktop app. But the thing what you can do is you can open one of these diffs, and you can start selecting the text like that, by clicking on this plus thing and then dragging down, and then you have this comment thing, and you can just say, like, let's imagine this is actual copy, you can say, "Make it shorter." For example. And what that does is it starts creating these comments, and you can have multiple of them, right? So, you can just pick that, and you can pick this, and say, like, "Make it longer." And then you can just go through all your comments, comments.
And then you just send it, and it will just process them.
So, you can, that's quite nice. Like, honestly, even just having like the green and red highlight, like the old and the new, like, it's much nicer than the way it works, it's just that's displayed in VS Code.
Yeah. So, it's like you can do that. It's not perfect, but it's one way you can do that. So, it's kind of like, you either open the markdown file in fire, and it's like, you can't really do anything, or you open the diff, and then you can do that, and you can just queue your comments like a Google Doc.
But if you, let's say, create a newsletter, for example, then it's kind of nice. You get, like, basically, you get a full green new file, like, new markdown file created, and you can go and comment on each line and drop your comments. It will just know what to do and do it, basically. Um, so you get the idea. Like, that's how it works. So, yeah, that's pretty much how it works, and it's a good starting point. I think there's still things missing in the desktop app, but it's kind of nice when you kind of, like, run it like that. Like, you remove everything, and it's like, literally, you just have this one thing, and you can just go on the side and see all your threads here, basically. Like, it's very simple.
I really can't emphasize enough, like, for me, in this position, like, there is no downside to use. Like, I have them both open right now. They're both accessed in the same folder. Like, it's not like a one-way door type decision that you're locking yourself into one app or one ecosystem.
Yeah, you can just start on the desktop, and then just jump onto the text editor only when you need the complexity and the advanced text editing features of it. That's kind of the beauty. So, let's go back to your setup and keep going.
So, I mean, I've got my claw.md file open here. Do you want to have a look through that and tell me, uh, if I'm, I'm setting this up correctly? So, it's, I, I'm thinking, or I was told somewhere, probably by you, like, try and keep it as short as possible, because this is the, this file is sent secretly, like, in the background of every single chat I'm sending to provide context to Claude. So, I have a couple sentences about what our business does, you, me, our product, who the key people are, a little bit like about a brand voice, just avoid, you know, cheesy words, like, be practical, uh, that kind of thing. My, never use m-dashes.
Writing style is just, don't use m-dashes. [laughter]
Honestly, this works. Like, I don't know how, but like, it knows the writing style, it gets it, and it's never uses m-dashes.
I'll show you a better way.
Okay. I've then [snorts] got my directory structure. So, on my left here, I've got, you know, the projects folder, the dump where I'm, you know, dragging and dropping in files to work with. I've got my areas, and I got data, uh, which I've started populating, but I haven't really kind of like solved this data, uh, thing. My idea was have, you know, like all of our podcast transcripts, or, you know, data, uh, things like design for the website, like all of our logos, icons are here, or photos of you and me that we use regularly across thumbnails and different things. I wanted to have them all there so they're in an easy-to-find place. That I have in my cloud.md file, the MCP servers, which I have access to. So, I mean, initially, it just, you know, I was asked it to interview me, and it kind of created a few of these. You can see here, for example, the Meta Ads one, where I've added this later, and I've just told it, "Hey, go update my claw.md file," and it's written, to be honest, I think more than it needs to have in in here.
Yeah, because I was going to say, the MCP servers, they already loaded in context as well, or somehow? And so, like, I think what would be more interesting is like, when you wanted to use these tools. Okay.
Um, so it's like, because like, it can discover what these tools are and what they do quite easily. Like, it's in the description of the tools. But however, it doesn't know when and how you want to use them. And so, like, that's kind of the context that's interesting. It's like, "Ah, like, use Notion when we want to save some notes, and use Bento when we want to email our subscribers that are not necessarily, uh, customers, for example." You know, that kind of stuff is not baked into the tool description. It's interesting.
But explaining that Bento is email marketing is not needed. It knows what it knows what it is.
Okay.
Because you're connected to the MCP.
I still, I find myself like, you know, a boomer, I'm still like, really directing it quite closely. "Use the Bento MCP to upload this email." Rather than, "I don't upload this." Yeah.
Because you have not inferred in the content that like, "Hey, when we send newsletter, we send it to Bento." For example. And so that is not in the context. And if it was baked in, then you would not need to babysit it as much, because it would infer it, right? Um, and so, so it's more like, it's still interesting to describe the tools you have, but in a different way, not what they do, but like when they are relevant, and who, like, you know, who uses them, in what context, etc.
Um, so...
That's useful.
And then the last section is my area context, which is just like, here's all the different areas you had. Yeah.
Yeah. Different other cloud MDs. Yeah. And then specifically, like, we had this marketing strategy one, 'cause I'm doing a lot of marketing stuff. It refers back to that, I think, quite frequently.
Yeah. Um, just for things like, you know, what's the price point of our product, who's our, uh, you know, things like that. How do you...
How does, to be honest with you, it's been a while since I updated, but it hasn't changed like a [snorts] lot in the last six weeks or so. So...
All right. Like, this is okay, but like, you, it's still not a lot of context, basically. So, yeah, the question is like, how do we give more context without bloating the code, right?
I know the answer to this. Yeah.
It's progressive disclosure, which is you have another markdown file, which is linked to from your cloud.md file. So, when it needs to, it can go there, right?
So, so how...
Which is kind of what I tried to do with area context. But...
So, there's two problems. One, it's like, for example, the writing style, etc. You need to write a lot more about like, how you like things written. The context of the company, as well, we need to give a lot more context. Who is the target customer? That would be very handy to have it. So, it's like, what you're missing is, you're missing a documentation folder, or some kind of like wiki that Claude maintains itself. Um, and so, what you need to go ahead and do is be like, "Hey, like, we're not going to put everything inside the cloud.md, but we're going to link to documentation files that it can go read when it needs." So, it's like, "Hey, if we're working on something that reaches the audience, go read the file that describes what the audience is." But it will not read it when you do something else, right? If we go, if we write some copy, just go read that guide on like, what our voice is, with the examples and everything. And again, it can be a quite decently big markdown file, but it will not be read when you're doing something else, like support or something.
My idea for that was to kind of put that in the data folder and kind of reference that there. I don't know if that's the best place for...
But there's another part that you're missing in this is like, okay, the documentation folder matters, but the problem is you told me it's not updated, right? You haven't updated the marketing strategy, etc. So, what you need to say in the cloud.md is rules. It's like, "Hey, every time something new comes up that is different from what you've read in the documentation, as you did the task, offer to update the documentation." And so, what's going to happen when you finish a task? Be like, "Hey, I noticed that we did this thing, and it's different from what I read in the file when I was doing the task. Do you want me to update it?" And you'll say yes, and it will just go ahead and update your documentation, and your context is going to keep updating with minimal work on your end, basically.
Okay. Yeah, that would be super useful, 'cause I mean, I was thinking like, it would update it automatically, and I would be like, "I didn't want that to happen in case it like, you know, goes a bit rogue and updates in a way I don't want." But that's, yeah...
Yeah, you just say, you proactively offer to update the documentation, and it will just, and you tell which file you're targeting, and what you're going to change. And at the end of a task, when you say something like, "Hey, you know, I was reading this file because the code.md told me to read it, and now we say the price is this much, when actually my documentation said it's this much, should we update the price on the documentation?" You say yes, and we just go boom, and just go update all the files, basically.
And so, you kind of want it to maintain its own mini wiki like that, and you want it to take care of it, because the problem with documentation is always outdated, and it's not going to be perfect, right? Like, sometimes you'll need to have a bit of a bass, etc., on cleaning things up, but it's going to help a lot in not, because the problem is probably inside your folder now, you have a lot of stuff that's getting older now.
And then as it's getting older, the quality of what you get is going to drop because it's going to be outdated.
So, you need to update your CL. You need to make a documentation folder. So, like, literally make a new chat at some point, be like, "Hey, we're making a mini wiki that's just for you. Make sure you update your code.md that you know that." And then let's kind of like, you know, try to go through the workspace and try to build this wiki based on what we have, and build it together. Spend half an hour doing that, and then update your code.md pointing at each documentation file, and then change your code.md to tell it to update things.
I'm thinking that principle of asking it to prompt me to, if it wants to update things, can be applied in other areas as well. For example, uh, like any task I'm running regularly, specifically, I'm thinking like, uh, create things, email marketing, and podcast publishing, especially like thumbnails, titles, that kind of thing. It can kind of update its skill prompt, if you want to update the skill based on what I've changed or something like that, rather than, 'cause when it doesn't prompt me, I'm like, "Oh, I'll figure it out next week or something." And I never go back and update.
Just... The way you do it is like, you kind of need to give a clear signal that the task is done. Let's say you worked on the newsletter, and it's like, you get a, you're like, "Okay, we're done now." And you say, "Hey, when you were done, and we've just run the skill, you review the chat, and then you review the skill, and you propose edits to the skill that match the specific use cases." And so, then the only thing you need to say is like, "We're done." And then it will know to do this process when you say, "We're done." You know?
Even that, I think I would probably wouldn't do all the time. I might have it ask me, "Are you happy now that we're done?" And then like, it will look for that, like, trigger before it then does that prompt or something like that. But the point is, like, you just need to signal that this is the final version of the task, because otherwise it might think you're iterating again, and so it will not know when to trigger that. But yeah, that's kind of like a flow that you can have, and it's just minimal work as well. And you can see how the goal is to kind of like offload all the tedious stuff to the model and have it do it rather than do it yourself. So, it's like a thing that I have with a lot of people I talk to is like, instead of putting just context in your cloud.md, you kind of want to put rules, rules that build context, basically. And you go one level of abstraction up, and you just tell it how to maintain its own context. So, that's, in my opinion, that's kind of like one of the biggest ones for you is like, I would do that. Actually, one thing that you can do as well is you can add some rules that kind of like offset the weaknesses of the models. So, did this happen to you, for example, that sometimes you're talking to Claude code, and it's like, "Ah, go and type this in your terminal," or like, "You should go and copy-paste this in this file," etc., where you can tell it, "It can clearly do it."
Always. And I just reply, "Do it yourself."
So, again, these are the kind of rules you probably want to put in your cloud.md, which is like, "If you can do the thing, don't ask me, do it." [laughter] And it's like, that's the kind of rules that I like to put in there. Same, I always tell it to always verify and correct its work. So, after it's done with the task, it should review it, and then apply fixes if it needs to apply fixes. It sounds obvious, but like, you kind of reinforce these behaviors by putting these one-liners into your CL.md. So, I have these kind of like workflow rules. And so, like, find all the annoying things that happen in terms of the workflow that you have with Claude, and that's what your Claude is for, is just putting some kind of like high-level workflow rules. So, one thing as well that happens quite often is, you know, I dictate my prompts, and Opus 4.7 likes to take things very literally, and so like, it just does what you say, nothing more, nothing less. And so, I have instructions to tell it to infer the meaning of my prompt, and kind of go beyond what I just said, to behave a little bit more like Opus 4.6, for example, right? And then, if it's not sure, to ask me, and to use the question tool. And so, find the things that irritate you with your workflows, or that makes it stop when it shouldn't, and put very high-level prompts to tell it not to do that, basically. And that, that will make your experience significantly better.
Okay, great. All right, let's talk about skills now. So, for those who don't know, skills are basically saved workflows, as a mix of markdown and scripts that allow you to repeat the same workflow and to document it in great detail so that you can get a consistent output every time. So, you can see that this lives under the dot Claude folder on Mark's computer, under the skills subfolder, and then below that, then you have a folder for each skill. And then if you open one of them, you'll see you have at least a skill.md, and you could have other folders, basically.
Uh, yep. So, skill.md is the kind of like trunk of the workflow that gives the steps to follow, and potentially links to other documentation. Again, progressive disclosure. So, you give the high level, it goes and reads some documentation when it needs, if it needs to, and the model decides. So, let's talk about some of the skills you use, and see how do you use them.
Okay. So, I mean, the most complex skill I would say I use is the podcast publishing one, um, where I'm analyzing the transcript of this, and I'll be doing this after our call or our recording. Then it's generating a number of thumbnails, YouTube thumbnails, as well as like title ideas and stuff. It's a bit of a back-and-forth process there to kind of get something that that fits. Like, it's not fully automated. I still need to kind of guide it and steer it a bit.
Why do you have to guide it? Like, what do you have to get about it?
It's just, I mean, it doesn't hit, it doesn't give me three perfect ideas every time. Like, I have it give me 10 ideas, and then I select three that I think are going to work better, 'cause some of them are like...
I agree. This is the kind of stuff where like, you still need some kind of human taste, and you could improve it, but I don't think you could make it perfect, basically. Uh, so it's like, fair. It's not...
You know, the single biggest thing that improved that was switching from Nano Banana, uh, image model, GPT Images, too, like, it's completely transformed it. So...
It's a good model. GPT Images, like, very, very bullish on that model. Yeah. If there was a way to upload that to YouTube to get to upload to YouTube, that would be amazing, but I don't think there is. I was just thinking of the email broadcast one, which I got open, actually. So, if you want to do a promo for, you know, a new skill that we've released inside AI accelerator, or a new course, or thing. Um, I often use this, and it will write it, plus, uh, upload the email to Bento, which is our email tool, as a draft ready to send. There's a few quirks, like it doesn't do the preheader text and things like that, but it does save some copy-pasting.
Why doesn't it do the preheader text?
Don't think it's possible in Bento to do the preheader text. So, I edited the skill.
I don't have one open, actually, but it, maybe this one, this one has it. No, I don't think so. So, I edited the skill where it will give me all this, you know, say, "Hey, I've uploaded this," and then it'll like, give me the text to copy-paste into the preheader. So...
Not too bad.
Uh, and so, what's frustrating you with these skills? Like, do you think it could be better? Is there anything that you still do manually that could be automated?
So, the, what's happened is like, the processes which I run, you know, every week are pretty decent, I would say. You know, there's minimal copy-pasting. What's happening though is there's a lot of like, kind of small, you know, "Oh, it only takes five minutes" type tasks, which I do, which I'm automating parts of in Claude code, but I still end up copy-pasting a bunch of stuff. And my barrier is really the setup, the integration setup with all the tools. It's like, "Ah, it's going to take me half an hour to figure out how to connect my Google." Uh, so here's a good example. We have a group, right? And it's for our plus members only. And so, people have to apply to join that. Um, and I have to manually invite them to WhatsApp. Like, there's no way around that at the moment. When they update their WhatsApp number, I get a notification. It saves it in a sheet. And then this is able to, because I've connected Google Sheets into Claude through the desktop app.
Yeah, it's much easier.
It can only read. It can't write. So, it can read the Circle ID and check if they have access, and it can give me the name and the email from the Circle ID. Um, and I can say, "Okay, yeah, this..."
But you cannot update the spreadsheet saying you processed it, right?
Yeah. So, at the end, I have to just go copy-paste that manually, and it's like, yeah, again, it takes 5 seconds, 20 seconds each time I do it. But, you know, it's one of those things that...
So, this one is particularly annoying because Google sucks at making these APIs easy to access. If you've never done it, it's done through the Google Cloud Console, which is kind of like the enterprise API type stuff. And that's kind of the only way to have read and write access to a lot of the tools. You can get access to your entire Google Workspace, which is awesome. Uh, and then if you connect it through the Google Workspace skill, workspace skills, sorry, then it's really quite powerful, actually. You can really do a lot with this, but it's, it's like a 15-step process of a confusing dashboard to click around, download some JSON files, and then link to them. It's very complicated.
I think what's happened as well is they've obviously changed the like menu layouts over time. So, whenever you ask AI what to do, it's like, "Oh, click on..." You know, this doesn't exist.
You know, there's a solution, right? Which is, there's actually CLI, which is short for command-line interface, which is essentially a terminal tool that allows Claude or CodeX to operate the cloud console, like the confusing dashboard. And so, it can go do most of the work for you. So, if you actually connected the Google Cloud Console, and then you just need to authenticate, right, with your account, and just give it full access, and then you can just tell it, "Hey, I want to like, create a new API key, you know, for this thing, uh, to connect it." And then it's going to do most of the work, and it's just going to give you a URL to click on, and then you're just, I need to download this JSON file that they require, and just tell it where it is, basically, you know, copy the path, or even drag and drop it into your code, basically. And that's the only thing that you will have left to do manually. Claude can do all the API setups for you, and you can do all of that. So, it's like the high...
Like, first of all, can we just talk about what a CLI is compared to an MCP? Like, it's essentially, like, if you don't care about the technicalities, they both let you access all of the functionality of a system.
Yeah. But CLI is like, better or more efficient in some way?
Yeah. It's a terminal tool. It's meant for humans, in principle, but just agents are very good at using it. They're very good at terminal, and so you can connect that, you can authenticate yourself, and then anytime you need an API from Google, you can just say, "Hey, can you add the API to my project?" And it will just do it. Um, so this goes 15 minutes.
This goes on to my next. Yeah, that's my next question. Is that a bollig to set up?
It's a 10 to 15 minutes, and there's literally two pages to open rather than 15 to 25 with the manual setup. [laughter] So, that's the solution. And then you can update your skill to update your spreadsheet, so you don't have to go and open it and mark everyone who has added. If you use the code desktop app, uh, you could have it use computer vision and add people through the WhatsApp app, if you want. So, you can take over your mouse and go and add people as well, if you want. Not super, super clean. The computer vision is like mediocre, I would say, with Claude, but potentially you could launch that, go for lunch, come back, and it's done, or it deleted everything. But yeah. Yeah, I haven't been impressed with computer vision. I mean, I haven't used it recently on Claude, but, you know, 3, 4 months ago, like, it just was super slow to do anything. So, if I'm at my computer, it's, it's not speeding anything up.
Especially when we first launched, there was like 50 people to add or something, but now...
You know, it's like, one every couple...
But that spreadsheet should update itself. Like, that's not good.
Yeah. Yeah, definitely.
Any, any other skill you want to show?
Yeah. Yeah. So, I have this skill which does plus member application. So, when before people join AI Accelerator Plus, uh, we screen everyone, and they fill in a form on our website, and then, uh, I get the application and...
At a time of my choosing, so not automatically. I'll go through and review them.
And so, I, I have a skill. I type slash plus applications.
What was happening?
As well? Why did you have...
What was happening is it worked, seemed to work quite well initially, but I don't know if the Notion CRM thing that we built started getting lots and lots of names on it or something, but it would do like plus applications, and it's supposed to look for, you know, new or not processed, whatever the initial, um, category is.
And even if there's like two people in there, it would be like searching through like all the other names, you know, spend like, really a long time just trying to try, trying to find people. So, I'm using the Notion MCP. I think I can't even remember when I think I set that up in the Claude.
The desktop app. I can see because if it says Claude AI, then this is the desktop app. So, you can see in your tool calls, it says Claude AI, which means, yeah, that's good. That's good to know, 'cause like, one thing I did wonder is like, how the hell does have access to my Notion? 'Cause I didn't remember setting that up in VS Code. If you also do type slash MCP, it will show you which ones are connected through the Notion app. And no, if you type just in in the... Yeah. Slash, just type. Yeah. Search MCP MCP servers.
Press enter.
And you can see, see local, Claude AI, etc. You can see which ones are connected. Claude AI is the desktop, remember.
Yeah.
So, you can see what is connected through. And then you can see your Gmail is connected, your calendar is disconnected, and Stripe is disconnected. Uh, and you have Circle twice, I think. No, you only have it once. Sorry. But it's disconnected.
No, no, it's twice. Oh, you have it...
Twice. And, okay. So, again, context optimization, if you wanted to. So, let's go back to the Notion workflow. So, I think, yeah, you have a skill problem. First of all, like, the fact that you cannot find the pending stuff, I guess it's either an MCP program or skill program, but I can see the return is like, this heavy JSON, basically. I can see I can see the tool. So, I don't use the Notion MCP. I use the Notion CLI again. And what's really cool is it, it's more efficient, basically. It just gives less big payloads to the agent. And same for pages, right? When it reads a page, it, you know, Notion has all this formatting, etc., that ends up being just lots of characters to express that, but it now gives the content of a page in markdown, for example, when the agent goes. So, it's very, very efficient. It's almost as if you had local markdown files, but it's kind of nice on Notion. So, it's the NTN, the NTN extension. So, that's why it's called the NTN CLI.
Okay. N...
So, sorry, N, like the letter N, TN...
TN CLI.
Okay.
And, uh, if you search for that, you'll find it, and it's, it's very good, basically. Um, so you install it, you activate it, it's already activated on our workspace, and you give it access to things, and you remove the Notion MCP, otherwise it's going to default to that.
Okay.
And then you tell it in your cloud.md, be like, "Hey, when you interact with Notion, use the NTN CLI." For example.
So, question for you. This, because I'm going to be using, uh, the cloud desktop app quite a lot more now after this conversation.
You can still do that.
Should, should I be setting up? Does it matter where I set the MCPs up? Like, do they need to be in, you know, the claw.ai?
It works anyway. So, the desktop app, it's the same.
Like you can see how your desktop app MCPS are showing up in your VS code right now.
>> Yeah.
>> And vice versa. If you have a C MCP setup in the folder that you're working, it's going to read that file and also use that. Um, so that's why it's really interchangeable completely and it's like there's no issue with switching. So, yeah, set that one up and I guess it's like you need to make it sure that it reads the statuses properly.
So what does it do? It gets the data. Then what?
>> It gets the data and then we go through this process. It makes a recommendation about like, uh, you know, should we approve this person, get a call, ask for more information?
How does it do that?
>> I can't even remember. So I mean, what happens here? Uh, I mean, it's a very simple skill. It's just a skill.md file, right? Um, okay.
>> And we automatically enrich the application with data inside Notion. So when it pulls the Notion page, it's pulling the context of the application plus
>> what AI has already found out through its research and its assessment. Right? So there's that that assessment, but it's never automatically approving someone or anything like that. It's just presenting it. And honestly, mostly it's kind of like my read on this person, like how we want to proceed.
How does AI enrich the lead?
>> I don't know. You set that up. [laughter]
>> Ah, it's my workflow then. It's just it's a Gemini uh search basically.
>> I mean, we did that a while ago. I'm sure we could do better now, but
>> Yeah, I can see. Okay, cool. So, it just basically just decides. I didn't realize it was still on. [laughter] It's okay. Uh, so okay, so it tells you that and you just get that and then you just decide basically.
So do you decide in code or do you go back to Notion?
>> Yeah, no, it gives me a few options and then it'll write the email based on that and it does a pretty good job of doing that and I just kind of copy that copy paste that email.
>> Um [laughter]
>> What? Well, this is so this is where it's a little bit complicated. So what was happening was we were initially emailing these people and saying, "Hey, thanks for your application. We want to get on a call. Here's a calendar link."
>> But then we started having these deliverability issues because what happens is a lot of people send out, you know, spam emails or uh, you know, cold outreach saying, "Hey, you know, let's get on a call so I can sell you something."
>> But because they hadn't emailed us, you know, we were sort of essentially cold emailing them from the mail provider's perspective. So
>> they were not getting through. So what it's set up to do is it the form is built on UForm. So UForm will send me but the reply address is the customer. Uh, so then I reply to the UForm email. So it's like a re-email rather than a new email. And I don't include a Calendly link. I say I ask if you want to get on a call in the next few days. Um, and then usually they say yeah, like when I'm free. And then
>> What does why does that stop you from automating this? It doesn't. [laughter]
>> That's the question. It's like all of this for that for nothing. The point is like you should have your Gmail connected. It should send that email rather than you doing it and copy pasting.
>> This is the same situation as before is like, oh, how do I connect my Gmail to this uh write maybe read only? Maybe it's read only. Yeah. Okay. But again, it's the same Google Cloud Console thing.
>> It wasn't just that. So, it's also how do I tie the email I received back to this application? It's not as if it's reading that email and replying. It's reading the Notion and then it needs to find that email, which obviously it's going to have the same email address, right? So, it should be easy to do.
>> Oh my god, it sounds very hard for AI. [laughter]
>> I know. I know. I know. It's it's just it's that like mental barrier of like, oh, I'm going to have to like spend a few minutes like focusing on this and I just haven't done it yet.
Okay. So, it can do all of that. I think we could improve our research and most importantly, this skill should be scheduled. You should not run it.
>> Um, what should happen is
>> potentially
>> scheduled or triggered.
>> Is it important to reply to people fast and how fast you respond right now?
>> No, usually in like, uh, you know, the next working day. Um
>> So I think scheduled is better because then you can batch all applications at once.
>> And so what you probably want to do is you want to do a batch run as a routine, right? So on the cloud desktop app, you can schedule tasks. Um, so let's go back to mine actually and it's super easy. First of all, like the newsletter on Google IO is ready. [laughter] Actually made me a newsletter while we were talking, but and that's one of our skills in AI accelerator. But the point is you have these routines here that you can create and so you can just create a new routine, run it locally, run it in whichever folder you want. So again, if you've broken down your folders, you can do that.
>> Ju sorry, you're going fast here. So just so I understand, a routine is a scheduled task and there's local routines which require your computer to be on and cloud routines which don't, but some cost to that or something?
>> Yeah, it costs more money and it's like, uh, you need a cloud environment set up. It's for developers. It's not for you. Don't use it and put in auto mode so that you know it can
>> like it's not for me specifically.
>> No, it's not for anyone listening to this podcast, to be honest. Um, and it's like I don't think I would recommend you do that. Like run things locally for now. And you can set a schedule. You can say every day at 9:00 a.m. And the point is like the prompt would be like, go through all pending applications and prepare, like if you don't want to send the email, just say prepare the drafts. And then what's going to happen is you have that thread waiting for you in your threads on code and you know, you wake up in the morning, be like, hey, I reviewed them. I prepared the drafts, etc. Are you okay for me to send them? And you're like, oh, maybe send the first two, but don't send the last one. I don't think we should approve him, for example. And you're done. And the idea is like the thread has run already. Uh, and you're just kind of like, you input the last prompt that sends the emails. Uh, and you can do the review quickly, like your manual review as you do it. And so that's another perk of switching to the desktop app is you have these scheduled ones that you can do. And you don't even have to open this interface. If you just make a chat and you say, "Let's make a new routine." And you chat on what you wanted to do and which skills you, it will just create it for you. So you can just have a chat on like, hey, every morning I want you to do that and every evening I want you to do that, etc. And use this skill and everything. It will write the prompts for you. It will do everything. And so that's the other thing that you need to think about is like, not only how do I make skills, but also which skills should run on their own. And you can run them up to the last approval point. So they like you could say, for example, like imagine for example, we have a a skill section on our circle, right? What you could do is you could say, every time G posts a new skill, it's only me posting there, right?
>> Prepare the draft for the email for the email list.
>> And then you just basically you get it ready and then you just can open the markdown file. You're like, yes, no, or like you give some, be like, okay, cute on bento, and that's done and you don't have to think about it.
>> Yeah. Yeah. I see. I see. So, okay, this is cool. So [laughter] like to use an analogy here, it's like, you know when you like sometimes feel like not going to the gym.
>> It's like the only thing you need to do is not go to the gym and work out. It's just put your shoes on or like get your gym kit ready or put on whatever.
>> Some of these tasks like I know AI can do it really easily, but it's like it's just that like mental hurdle of like starting it that's uh that's there's some friction. But the point is you can do it manually and then be like, hey, okay, let's schedule it and then don't don't put the last publishing step, like make it stop at that point so that you retain all the control and then you can run it like 10 times and then when you're like fully confident it works every time, you never change anything, you're like, okay, now you can do the last step, just update the automation.
>> But just to ask, so the routines are triggered at a set time, whereas the one with circle you're talking about, wouldn't that be better to just like as soon as you post it be uh
>> Potentially, but again, is it released that time sensitive?
>> No.
>> Like not really, right? If it checked once a day
>> or like whatever. The thing is like Anthropic is a little bit sneaky because they kind of limit you. You have limits for routines.
>> Um,
>> Let's show me show you. So, if you actually go into settings and then usage, you will see that you have uh daily routine runs. And you can see you can run 15 a day, which is fine for most people. Like it means you can run 15 automations per day. And it'll be annoying for that.
>> Does that mean if I'm running it weekly, I you know,
>> you use one for one day.
>> Okay. Right. So, you know, let's say I had seven routines all week on different days. That's one routine of my one of 15.
>> Yeah.
>> Okay.
>> So, it's not it's not that bad. Basically, uh, you can automate a lot of stuff and provided you hit that ceiling, we can totally vibe code an alternative to this if you want to. Um, and it's not it's not super difficult basically.
>> Okay. So if you're using the hourly routine, then you're going to use more.
>> Yeah. That and for most of the time that's like that's when Collect is good because Collect doesn't limit you at all on automations. You can do as many as you want.
>> Um, but yeah, you can see how now you can start thinking of like not being the trigger for these things anymore, which I think is kind of the next step for you is like, how do I like what happens systematically and I can have AI get to the point where I'm needed without me thinking about it. And then you'll just see these threads pop up in your code desktop app in the right subfolder and you'll just have these little notifications like I have here for uh the newsletter, for example. So you can see for example, I have this little dot here next to prepare newsletter, that means it needs my attention. So you'll get a thread like that basically. And so if these things get ready as well, you're more likely to actually do them rather than skip them because they're just kind of cued and waiting for you.
>> Yeah.
>> And I think more likely to do them quickly as well.
>> Exactly. And then it removes the need to think about it in your head, which kind of frees mental space, which is kind of like, you know, switching to the desktop in my opinion, it will free mental space for you because you won't always stare at VS Code all the time and it's
>> that mental space that is my blocker right now because you know, this allows me to do so much, but it's like that there's also so much more to like kind of just think about and you know, make sure it's happening type thing. So yeah.
>> So
>> I'm literally going to when we're done with this, I'm going to take the transcript of this call and be like, make me a list of things I need to do in cloud code. Now.
>> I'm curious to see how people like this episode though. Let us know in the comments. Uh, if you think this is interesting or if you're more like that, if I should bring more people to grill on the podcast or something.
>> Yeah, we could do a series of [laughter] like G roster your cloud code setup. Uh, and yeah, it'd be interesting to see how different businesses use it as well.
>> Cool. But yeah, so I mean, I think that's the main stuff for you. It's like basically make your skills do more, connect all the tools, then automate them up to the point where it gets published so that you have no drop in quality. You can always give your feedback like you always do, but it just does it. Uh, Claude MD should have more rules uh and should also have you need a wiki that has a lot more context. So it goes and reads the context it needs for the task at hand, but it doesn't read everything.
>> Actually, one question I had, what's the best way to build that context? So you know, take the writing style example. Would that be just to like look at writing like emails and things that we put out before and reverse engineer it or
>> Yeah, I know how you are. You'll be like, I want it to be perfect on day one. Don't try that. Uh, all that matters is that you just want a document that says writing style and yeah, you can give it three emails and be like, hey, write a writing style document based on that. And then
>> okay.
>> Because it has the rule of updating its documentation in the cloud. MD, it will whenever you give feedback on an email it wrote, it should offer you to update the writing guidelines and these writing guidelines just get better, same as your skills as you use them basically.
>> Yeah. Okay.
>> And so that's why these kind of like high-level rules matter more than hard facts almost. You want to tell it how to think and what are the big high-level frameworks.
>> How to think, but how to improve as well. I think that's the key. But that's the point that it should think of improving itself and it should think of like doing the task if it can do it rather than telling you to do anything, etc. And so like that's that's what your cloud MD is for.
>> Yeah. Awesome.
>> All right.
>> It's actually super useful podcast for me for a change. [laughter]
>> All right. Well, I guess uh I guess we're good, right?
>> Yeah. So, uh, you know, we've obviously been through a lot with Claude Code here. If you're not currently using Claude Code or you need some help getting set up initially, you can actually go on to authorityhacker.com/learnclaudecode and you enter your email there and we're giving away the first two lessons, which includes the setup of Claude Code from our Claude Code course, which is inside our product AI Accelerator. So authorityhacker.com/learnclaudecode and you can start getting set up there uh as well. So, thanks Gail for uh roasting my cloud code setup today. That was a humbling uh but a very ultimately very useful exercise. So, thanks for that and thank you for watching this podcast on YouTube. If you have any questions about your own cloud code setup, leave a comment on YouTube and we'll go through those in the next few days after this podcast is out and we'll try and uh, you know, give you guys some help and feedback on that as well. And also let us know as Gail said in the YouTube comments, if you would like more episodes like this where we kind of go into the kind of practical behind the scenes setup of how businesses business owners are using AI in different ways. And yeah, of course, subscribe so you don't miss the next episode. We'll see you next week for another episode of the Authority Hacker.