In this episode, Amir takes us through how to use cloud skills to build digital employees.
We go through AB testing idea agent, marketing insight agent, and then we build one live together.
You're going to learn about what cloud skills is, why it's the biggest thing that happened since sub-agents, and how to actually build them yourself.
Amir, what are we learning today?
Today, we're going to talk about cloud skills.
I'm going to tell you what they actually are, how they're different from projects and sub-agents in cloud, and why this matters and how you can actually apply for work.
Okay.
And by the end of this episode, are we going to be able to apply cloud skills?
A hundred percent.
I'm going to show you.
First, I want to talk about what it actually is and why it matters, but I'll show you how to use existing skills in cloud that they just came out with and how to create your own and how to apply for your work.
So whether you're in marketing or in data analysis or any sort of document creation, you can actually use skills to do that.
Cool.
Let's do it.
Cool.
So first thing is I want to talk about cloud.
So not a lot of people are familiar with cloud projects and I want to talk about what that actually is and why it matters and how it's kind of related to skills.
So, within Cloud AI specifically, you can actually create projects and they're essentially workspaces with a set of custom instructions.
So this is a system prompt And it has relevant context, memories, and tools.
So say, for example, you're part of a broader marketing team and you want to create a project that will have a set of instructions to analyze marketing data, for example, or generate a newsletter.
And you want it to connect to specific tools, have relevant context and files.
So this could be a glossary of terms you use within your organization, your brand guidelines, depending on the task it is that you wanted to do.
And then also have memories generated from the chats that you have within that project.
So it's really great for collaboration with other team members.
Now you can also use it yourself as well but really the ability for you to create repeatable tasks and do a certain set of instructions with external tools and data.
So, if I'm in a marketing team, this is something that I want to look at and essentially, youth and Claude, share with my team members and create projects around it.
All right.
So the only thing with projects, I would say that it's important to One work with your team members to actually refine the system instructions and then always have relevant context files.
As your business changes, the data changes, you need to update it and you have to go back and constantly update these context files.
And I'll talk about why context is important in this specific session and kind of how it ranks up against skills.
Now, the next part of it is subagents.
With subagents.
This is more relevant in Cloud Code specifically, and I actually use subagents in Cloud Code to spin up multiple agents.
And multiple agents are really great at breaking down complex multi-workflow tasks into individual tasks with specialized agents.
So what does that mean?
Say, for example you're building a very complex feature and you want to delegate the front end to one agent and the backend to another.
So within the chat, you can actually spin up these agents to say hey Claude, create an agent that will work on the frontend using this set of rules and then create another agent to spin up to do the backend for it.
And what's interesting is the context is isolated to that conversation window.
And so whatever context is provided or gathered in that conversation is actually used as an input.
But those agents have a set of system instructions as well.
Now, where things get interesting is skills.
And I want to talk about kind of why this actually matters.
Skills are automated workflows and tasks that you can apply globally at a project or individual level.
So, whether you're an existing project, you have a set of system instructions.
You can use skills which is an add-on or augmented skill set within that project or individual chat.
And it can do a set of tasks, create documents, create PDFs, analyze documents.
It can actually help build MCPs for you.
You can use skills to create other skills or create visual art as well.
Now, when do you actually use this?
It's for very specialized tasks based on the constraints and guidelines and steps built by you, the expert.
I think it was Kaparthi a couple of days ago.
He had an excellent analogy where it's like AI is essentially your coworker or someone that like, reports to you.
You want to train it, you want to build the guidelines.
You know this is not verbatim, but like basically, what he was trying to say was yeah, it's someone that you work with and you can kind of build the constraints around it and guidelines on how you want it to respond to you.
And this is kind of similar in some nature.
You can create.
For example, let's say you are a paid media expert and you run campaigns for your clients and you want a very detailed analysis on your visit to booked appointments and what the conversion rates look like and how that attributes to the different channels you have and what's performing better than the other in terms of campaigns.
You can create a skill that can follow a set of custom instructions but also scripts that you can build out yourself to analyze that data.
And I want to circle back later on why that actually is important.
But what's also interesting is that it actually only loads context when it's relevant to the task.
So when a project, oftentimes you have the LLM that's determining which context to retrieve and add into the conversation window and reference it.
But in this instance it's only based on the judgment of the task whether or not it should pull relevant context.
And it's just relevant to exactly what you want to get done.
So I would say the key takeaway here is that it's repeatable instructions, it's laser focused on a set of tasks, it pulls in context as is needed and it has the ability to run scripts or run code to perform specific functions.
Why this matters?
Because there's a paper great paper out there called talking about context, rot and how essentially talks about how to do effective prompt engineering and how the right amount of system prompts from you know, very detailed to vague, and the right amount of context has a huge impact on performance.
And as you add more context, you essentially could be you know, I don't want to be quoted on this but potentially be like degrading performance from the LLMs and likely to lead to more hallucination.
Sam Altman, the co-founder of OpenAI, just said that it is the era of the idea guy, and he is not wrong.
I think that right now is an incredible time to be building a startup.
And if you listen to this podcast, chances are you think so, too.
Now I think that you can look at trends differently to basically figure out what are the startup ideas you should be building.
So that's exactly why I built IdeaBrowser.com.
Every single day, you're going to get a free startup idea in your inbox, and it's all backed by me.
High quality data trends.
How we do it, people always ask.
We use AI agents to go and search.
What are people looking for and what are they screaming for in terms of products that you should be building, and then we hand it on a silver platter for you to go check out.
We do have a few paid plans that take it to the next level.
Give you more ideas, give you more AI agents and more almost like a chat GBT for ideas with it.
But you can start for free, ideabrowser.com.
And if you're listening to this, I highly recommend it.
It makes sense, right?
Exactly.
The more context you have, the less likely you are to hallucinate.
Well, yes and no.
The more context you have, you're less likely to hallucinate, but the right amount of context.
So it's kind of like a coworker.
Do you want to give them all the information or just the right amount, so that it doesn't bombard them to get the right task done?
Exactly.
So that's what I would really bring you down into.
So I'm going to go through some examples, but I want to talk about the importance of skill and why it's actually solving a real problem that I have faced myself.
So, with custom skills, how it works is that you essentially create this Markdown file that explains exactly what the skill is and what it does.
And you can actually create reference files that it can reference back into for additional context.
So say, for example, you create a skill that applies XYZ's company brand guidelines to presentation and documents.
And this overview.
So essentially you know the skill overview has a set of tasks and instructions.
That follows, but you can also have an additional document as a reference.
That is an example existing brand guideline document that I can reference.
And it's not, it's only pulling it when it needs to.
You can take it another layer and you can essentially create custom scripts as part of that skill.
Now there's a great documentation by Anthropic on this and they talk about kind of how to write good skills and descriptions.
But what's interesting is that When you are using a cloud project and you have MCPs or tools connected, connectors connected, the LLM is determining which tools to call based on your instructions and how to perform that task.
So say, for example, you have a raw output of your meta campaign ad data or your Google ads data and you have a project in cloud that says like it's a market analyzer.
I want you to, your instructions are to analyze this data and give me insights, right?
The LLM is determining how to like the model, is determining how to actually look at the data and perform insights.
And it's non-deterministic in a way, right?
Like it's, you know, it can look at it differently every single time.
And you're not giving the right guidelines on how to actually take the data and analyze it.
And I've seen this firsthand actually working with a lot of clients where a director of RevOps is looking at churn data, new subscription data and they'll put the file into the cloud project, and it's not giving the right output of insights they're looking for.
How this gets interesting is you can actually create scripts that are very specific.
So say, for example, if you wanted to have a very set of strict guidelines on how it should actually run and analyze the data, then you can create that within the scale itself.
To say, I want you to look at column X, Y and Z.
Multiply by this, divided by that, to the power of this to give me this insight.
That way it's actual functional code that's running this and it's not deterministic, non-deterministic by the model itself.
Yep.
So yeah, that's kind of the beauty of skills itself, where you're able to really bound or create the boundaries of what I should actually work towards for yeah, for like building out these skills.
So yeah, you can essentially have metadata with it, resources and code.
You can load it as needed.
And it kind of breaks down exactly how you should write these skills.
Now let's jump into some examples.
Let's do it.
This is the fun part.
How do you actually apply this?
So the first one we're going to go through is an artifact builder.
So you can actually go to Claude and it's preloaded with some existing skills.
So we're going to go to capabilities and essentially you'll see there's some existing skills that are preloaded.
So I have created these two ones right here.
We'll go through them, but I want to show you the ones that are already in there.
So you can have an artifacts builder, an MCP builder, and a skill creator.
So it's very meta.
You can create skills with skills.
So we'll go through an artifacts builder one, and I'll show you an example of what that looks like.
So say, for example, you want to create a tool that is relevant to marketers.
Marketers.
You know, when they run campaigns they always have to have UTM links to do proper attribution back to their data to see okay, which campaign was driving the most and when they're seeing the analytics.
So here, you know, I have added the artifacts builder skill.
Please create a UTM link generator for my marketing team.
What's happening here is that cloud is now going to reference that skill specifically that we have defined.
I'll show you what that skill looks like.
And essentially it's reading the documentation to understand how to build components.
Artifacts are essentially these like live apps within cloud itself that you can create very functional web apps.
And you can also share with your team as well.
So what it's doing is it's actually referencing that skill here and now creating an artifact slash web app of a UTM link generator that marketing teams can use.
And you can actually just share this with the rest of your team as well, or your entire team can use this as well.
I mean, it's literally a web app.
It's literally a web app.
But what's interesting is we're now creating a set of specific instructions and skills.
We're adding a skill to this LLM now that knows it has to follow this versus before, you were saying hey, code this web app.
And it's kind of, you're not really defining the guardrails or the parameters of what it should do.
And what happens is people get frustrated that they're not getting the right result.
Exactly.
And then they're like, oh.
AI doesn't work for me.
Exactly, exactly, exactly.
So this is where it gets interesting, right?
Like you as an individual, you have the opportunity to work with Claude and skills to build exactly the skill you're looking for to do.
That's a repeatable task.
So if you, as a marketer, are doing weekly tasks of reporting, create a skill that can actually help you with that.
And just explain to it in terms of what you're looking for, what you need.
Be very detailed.
If you were to assume that you're hiring someone else to do it for you.
Yeah, I mean it really is thinking about AI as a teammate, especially a junior teammate, that you have to really give it guardrails and really give it context.
Cause that's how you would you know, if you hired someone junior.
You would be like okay, these are the tools that you're going to use because they don't know.
The tools they're going to use because they're new.
This is the context that you need to know about our business and how we operate.
And then you kind of drip feed them.
You don't want to overload them, right?
Because then they're not going to remember everything or it just might be overwhelming.
So you drip feed them the context over time.
Exactly.
But it's also interesting is, if you start now, as these models get better and as the toolkit expands, you now have this like history of, like training and reference and metadata and memories that you've created over time.
Then.
So, like someone like me who uses cloud a lot, I now have a lot of pre-contacts of like memories and experience with building these projects.
Now I know exactly how to use skills and apply it here.
So I think that's where it gets really interesting.
I generally think skills is probably a huge problem solver for a lot of problems.
I've seen firsthand working with people.
Like I've worked with a lot of teams right now that have actually like a lot of go-to-market teams that have used cloud as part of the workflow.
And the number one feedback I get is the output was not on what I expected or it's incorrect.
There's two reasons for that.
One is The prompting is not good, right?
The prompter.
The prompter, the problem is them.
But the latter of it is.
It's also the you can prompt.
You know I worked on with them on setting the right guardrails, the prompts, the access to tools, the right retrieval of context, and it still doesn't get it just right.
And I think this is where skills come in and solves that problem, where it's just that task.
So you now have an artifact that's fully functional and working, you can actually share with your team members if you wanted to, and you can essentially like, provide a url like humboldtcom or ideabowsercom and it'll append um um, the rights like google and you know, cb.
Was it Black Friday, Cyber Monday?
Yeah, something like that.
And it'll actually create those.
I will append it for you.
I don't know why there's a clear button.
It should be a submit button.
You can also sell this to other people as a product, right?
Yeah.
So I think Claude, in collaboration with someone else, created this repository of skills.
I don't want to butcher the names.
I'm not going to say it.
But basically there is a directory of some sort with skills and plugins because they recently came up with plugins as well which is like a collection of context and tools and skills and prompts all in one that you can install for your cloud workflow.
So there is a huge opportunity for people to sell skills.
Okay, sorry.
What's the difference between plugins and skills?
Yeah, plugins just came out last week, which is like a plugin.
You plug it in and it has ancillary features MCP access, context system instructions and, I think, skills now.
Don't quote me on this.
Cloudsman Shipping.
I'm having trouble keeping up.
You know, when I was I was building this out.
When I was writing this out, I was like I'm trying to understand what skills is.
And as I was actually building with it, I was like, it's clear to me.
Cause initially my gut reaction was, this is overcomplicating it.
How is it different from projects?
Right.
And now I understand why.
Okay.
Yeah.
So, yeah.
So we essentially, you know, HTTPS, ideabowser.com.
We can go Google, CPC, Black Friday, Cyber Monday, and then generate the URL and we have the URL.
Boom.
So that's one use case.
Let's make it more interesting.
I am interested in finding A-B testing ideas for my website.
So I have a skill that essentially looks at A-B test generator.
And what it does is that you provide a URL and it will come up with or experiments for you to run for your website to increase conversions.
And it actually, the skill, I created the skill using the skill creator.
I said, I'm going to give you a URL and you're going to run a framework on actually how to run good AB tests for me.
So we're going to test this and see what it looks like.
And then I'll show you an example of how to create your own skill as well.
So, hey, Claude, I have just added, I have added the A-B test generate skill.
Can you run, can you provide me with A-B experiment ideas for humblelytics.com?
And what this will do is here because I have access to an MCB called fire crawl.
It would actually use fire crawl to scrape the URL, the page and the contents and then come back with a very clear framework on experiments to run.
So while that's running, maybe I'll just show you an example of what that actually looks like.
And essentially it looks something like this.
So it gives you an experiment pipeline impact, confidence ease, ICE score and performance.
You know it was actually a really good one.
I actually did it right before this call, before the session today.
Was it asked me to?
It told me to actually test um, shifting the case study that I have above one section above.
So it was like, here's a section and then case study, go to Amelia social proof.
And I was like, damn, that's a good idea.
Why am I showing the features when I should do the social proof?
So I'm running an A-B test right now to see which one is likely to drive more conversions in signups.
So that's interesting.
Like it really breaks down exactly the control, the variant, the headlines you should be testing.
So experiment number one, experiment number two.
And if I really wanted to, you know, just take this, put it into my app and then run through an experiment.
You know it'd be really cool if you can automate this so that you know every month send me a report.
Yeah.
What to change.
Exactly.
Exactly.
So if you really want to yeah, you can.
You can probably write a skill that I wonder if you can.
I wonder if I wonder if you can already do that today, where you write a skill that writes a script that automatically sends you a report every single week or every month.
Why not?
I am going to plug in my app and say we do that automatically in our app.
So every week we have like four sets of sub agents that go through your website and give you insights from like a copy, conversion marketing, like a designer as well.
So every week, and then we give you like an optimization score.
So similar kind of similar approach here, but I think doing it within cloud is actually really interesting as well.
Now, what I really really, really want to show you is a problem I've been trying to solve for the past couple of months with these companies I've been working with, which is take data and give me the insights that I actually want to look at.
It's such a repeatable task, and it's so important that...
I think I can't confirm yet.
Skills has probably solved that problem for me now in a way.
So I uploaded a file called Traffic Analytics.
It's just basically like a CSV of just a bunch of campaigns and revenue data and whatever.
And I was like, I need some insights on this.
And that's what really matters to a lot of people in terms of just the cost go down, the CBC go up, down.
What does a trial conversion look like?
X, Y, Z.
I provided a file and it referenced the skill and it has a set of scripts within that skill to then do a comprehensive analysis of the data the traffic data.
So, overall performance your total spend was 400K.
Your revenue was 854K.
Net profit conversions which channel did better than the other?
So you have a clear idea.
And I'll be honest.
I am going to be honest, but I would say that if I had done this through a project and I just uploaded a file with a set of instructions without running scripts, it would have probably hallucinated some of the data.
That's what I was going to say.
When I look at this, this feels...
If this wasn't in Claude and I had a product manager send me this, I'd be like, yeah.
You know, this feels like that level of fidelity.
It just, you know, it's...
It looks right.
It looks right.
I can't confirm because I didn't look at the Excel.
I'm just looking at it.
It looks right.
Yeah, it does.
Yeah, exactly.
So, and I'll show you what it looks like essentially within the breakdown of the skill itself.
So skills, what you do is you have the actual skill.md file itself.
So this again is a breakdown of what the scale is, the scripts it should run, and then generate data for 90 days, generate data for seven days, generate data for 10 campaigns, and then what the structure should look like.
So you can actually use this to define it.
And if I want to take a step further, I can use cursor to then update the skill itself.
And I'll show you an example of how to create your own skill.
And then you can also reference files.
So you can see here.
It says see references metrics, MD for detailed metric definition and typical ranges.
So if we want to go back into references, we can see what metrics MD has, which is all the you know definitions or glossary.
So, as a marketer, you get to define what these are, and you should be doing that so that you know when you run these scripts and skills, it gives you exactly what you need, instead of getting the LLM to actually do it for you.
And then the scripts are made by Cloud itself, where it's running a Python script on calculating all this for you.
So it's accurate in some way or another.
Yep.
Cool.
Let's see where we are.
Yeah.
So let's now create.
We've gone through A-B testing ideas.
We created an artifact.
We got some working insights.
I think we should now just create our own skill.
Do you have anything in mind?
Tell me if this is possible.
I tweet every day and I also have a newsletter and every single week I Basically I use my tweets like if it rips on Twitter, I'll kind of expand on it on my newsletter.
I have a specific type of style how I write on my newsletter.
So what would be really cool?
Like this is something I would hire for, potentially like almost like a ghostwriter.
So is it possible to have a skill that basically like, looks at my tweets and turns it into long form content that I can review and be the editor of?
Okay, let's try.
Let's find out.
We're doing a lot.
You're like, maybe.
Yeah, no, absolutely.
I think we can figure it.
We can try.
So, hey, Claude, I just added the skill creator skill.
So we're using the skill creator to do that.
Can you make me a skill that takes an existing tweet provided by the user and turns it into long-form content for LinkedIn?
For newsletter.
I think what I would do in this scenario is making sure that we have a reference file of your existing newsletter.
Would you need an export of all my tweets?
Exactly.
Maybe we can try to do one example one right now and then...
And then see.
But ideally what I would do is actually I actually have this.
I have an automated bot that looks, has all my tweets and then finds the most viral ones and tries to like expand on it.
So ideally we would export all of your tweets.
And then.
And then, you know, we would, in order to keep it updated, we would need to make sure that we're, you know, constantly updating.
Yeah.
Yeah, exactly.
Exactly.
So, while that's running behind the scenes, I'm going to scrape, I'm going to get some examples of of your posts on Twitter.
Yeah.
That's me.
Okay.
So what's a good tweet?
Yeah.
You find a tweet that like speaks to you and then we'll, I like this one actually.
It speaks to me.
Pricing?
Yeah.
I mean, this is a perfect one because it was too long for Twitter, but I posted it anyways.
But even if I did this on a newsletter, I would totally expand on this.
I gotcha.
And then... for the newsletter, which is where's the best.
I mean, I, I, I don't even know if you can.
Yeah.
How do I find it?
Go to, I think Greg Eisenberg dot CK dot Greg Eisenberg dot CK dot com.
All right, cool.
So let's take this actually.
And We're going to copy this and then create notes.
Export this example newsletter.
Export it as a markdown, right?
Yeah, as a markdown.
Exactly.
Exactly.
So we're going to go back to Claude now.
So now that we have an example tweet post and example newsletter, we're going to upload that as a reference file in there as well.
So what we'll do is let me just drop this in here. here, add these files as references.
You know, what's interesting is they actually give you a zip of the scale for you to upload.
And we're saying now like add this as a reference into that scale.
Okay, stuff is happening.
Things are happening.
We're seeing instructions.
Yeah, so we're adding instructions to, similar to how I showed before, where there's instructions to reference the file in the folder.
We're going to see a zip now with a folder called references.
And if you open it up, you'll see these two examples in there.
Boom.
So now we can essentially download this and re-upload it back into cloud.
There you go.
Sweet.
Okay.
Tweet to newsletter.
And we're now going to go to settings capabilities.
Upload a skill and we're going to upload tweet to newsletter.
And now we're going to write the skill.
Try and chat.
I just had a skill.
Can you turn this tweet into a newsletter format?
And you, you're going to be the judge of this.
Tell me if you think is good.
Yeah.
So we'll go back.
Oh, you know, we already had it, but I'll just copy paste this to see how it works.
Okay.
And we're going to go cloud.
Let's see what happens.
I'd be surprised if this crushes it on the first try.
Because think about how...
I don't know, I'd be surprised.
I hope it does.
So I'm very curious.
We're doing this live and I'm curious to see how it actually comes out.
And I want to get your honest take on it, right?
Because for me, at least from a data standpoint, to get the insights, I think it's interesting.
Cause like there was challenges with projects before to get the right insights.
So honestly, this is fire.
I mean, so tone of voice.
So we just did this in one shot, but if we really wanted to, we would take all of your existing tweets, all of your newsletters, and then use that to generate like a style guide or a tone of voice and then kind of refine this point.
It's not bad.
As a starting point it's not bad, and i'll even take it a step further.
Like this is especially not bad, really good, especially not bad at all.
Until next time, keep building and keep raising, not keep raising.
Yeah, we don't, you know, but keep building.
I like uh for this to a founder who's been sitting on the same price point for too long.
Like, I like that.
I think that's really smart, right?
We talking about part of market fit.
We should talk more about, Pricing market fit.
That's a banger.
That's actually a banger.
If I tweeted that as its own one-liner, that probably would do really well.
Until next week.
Yeah, I think so.
100%.
Now it would have been interesting if we could technically say okay, now I'll go scrape tweets and embed it in there.
That would be really cool.
And the cool thing is you can actually create skills now.
Dude.
This is crazy.
I actually think... Here's where it gets interesting though, right?
Because you can now create skills that generate visual graphics because that's... a thing now.
So you don't have to do MCP calls like Canva or anything like that.
You can programmatically create these visuals so we can update the skill to say hey, now add images as well in there.
Yeah.
This is pretty good.
All right.
Yeah.
So, you know, we covered why projects matter, how it's different.
I think...
And I think we got a good idea that skills are probably more deterministic in terms of what you want to do, in terms of how you want to define the skills.
And you can now do programmatic code in there with MCPs and tools.
And how impactful context is.
And, you know, essentially how it differs from everything else.
So when we cover some use cases as well, the last thing I want to talk about is kind of, just like you know, I saw this report.
I don't know if you, if you saw from ramp where they were tracking subscriptions for like different AI tools and they saw that there's a dip happening.
And they're saying there's getting stickier in enterprise, but the AI is off ramping and it's not as sticky as we want it to be because cost is coming down.
I actually want to say now, with what we're seeing, with skills and all the education and awareness around prompting, we should be able to solve that gap.
Because the reality is a lot of companies are investing in AI.
And there's reports now saying that it's not actually being as productive as we thought.
The issue is, I think, prompting and context.
The issue is people, not AI.
Yeah, that's the reality, right?
There isn't enough AI fluency and education around how to actually do prompting.
People write, build me a SaaS, 1 million AR, don't make mistakes.
And the reality is no.
No, you got to give the right amount of context and do some prompt structure.
And I think, when it comes to Anthropic, they do a really good job of not only building with intention, but creating the resources, the education to help people actually become more AI fluent and giving them tools to do that.
And they're very deliberate with what they create.
And it's actually real problem solvers.
Like I've never seen a company so dialed into customer feedback and just creating something around it.
It's almost like they heard me in conversations about what issues I'm having and why skills matter.
So the net of this takeaway is AI adoption may be falling and the adoption rates may be down for this month or this past quarter.
I think part of it is just because companies don't have the right resources and the people to build education on AI enablement and AI fluency.
And then, once we see that come into play, adoption is going to come back up, and there's a tool now to support that.
Beautiful.
Well, thanks for explaining it to me, honestly, and everyone else.
For more of Amir, I'll include links in the show notes where you can go ahead and follow him.
X is the best place.
Amir, M-X-T-A-M-I-R-M-X-T.
I appreciate you coming on.
Thanks for having me.
