Hey, it's me here and today I'm going to

talk about cloud skills and how they're

different from sub agents and projects

and why this matters and how you can

apply it for work and what they actually

are. So, let's dive into it. With

projects, um, you're essentially using

them as workspaces within claude or even

chat GPT as well. And what it comes is

that it's essentially a set of system

instructions that are repeatable

instructions you want to follow. It has

context and files you've uploaded. It

has memories and access to tools.

projects are specifically great for

collaboration like kind of like a school

project and you're able to use it with

other team members. So say for example

you're part of a marketing team and you

have a set of ongoing repetitive tasks

or instructions you want to follow. For

example, you are a marketing expert in

analyzing in marketing data and creating

reports. It has a persistent knowledge

base and memories for continuous tasks.

But the thing is you have to um

constantly update the instructions over

time based on the tasks you want to

complete and also make sure that you

have uh contextual files you've uploaded

that are being regularly updated based

on your new data that are coming in or

whatever your needs are. Um I would say

it's great from a sense that you can

always have repeatable instructions in

there. But um I'm going to talk about

why skills could be an add-on to

projects within projects and as an

individual chat. So if you don't want to

have any uh consistent context or

instructions in there and I'll get I'll

get back to that. So with sub agents

this is another subset of you know

specialized skill set um but this is

more applicable for claude and what this

is essentially is you're able to spin up

individual agents um to break down

complex tasks or workflows into

individual ones with specialized agents.

So, say for example, you're building a

very complex feature and you want to

delegate the the front end to one agent

and the back end to another and they're

going to be focusing on specifically

that. So, for example, you know, you

have an agent that is building out the

UI and the agent that's building out the

backend API endpoints and the context is

isolated to that conversation window.

So, whatever context you provided and

you know, you gather in that

conversation. Now, skills are very

interesting. What is it? With skills,

you're automating a lot of the workflows

and tasks that can be applied both at an

individual chat level and at a global

project level as well. And it's great

for uh tasks that you want to set the

constraints for. The guidelines and the

steps are built for for by you. So, say

for example, you know, Kaparthi says how

you have AI as your coworker. Now you're

giving them an extra skill set for them

to follow step-by-step instructions on.

And you can create like very detailed

instructions for it to follow and it can

be repeatable, laser focused on the task

and you can give context as needed and

have embedded scripts as part of it for

it to follow in the instructions. So you

want to run actual code. Why does this

matter? I think part of it is because

LMS are so nondeterministic. you can

essentially have some sort of

hallucination or the wrong output or too

much context rot that degrades

performance. And there's a whole paper

about context and what is the right

amount and how much context should you

give and how you should prompt that. So

I want to kind of show you some use

cases and how this would compare to for

example um a project instruction or

system instructions and what skills

compared against that. But um you know

there's some documentation online on how

to actually create custom skills and

we'll go through an example together of

what that looks like. Um but you can

also read the docs uh yourself. Um but

also here what's interesting is that

cloud itself talks about why skills

matter for work. The example they

provide here is you essentially have a

set of tasks that you've already figured

out what works best and you want to

structure it better. So say for example

um it enables you to package that

expertise. So before skills, you had a

project that helped you with quarterly

business reviews. So claude can already

build presentations, but to get it to

match your standards and or stay up to

date with your standards and your

preferences, you would have to document

those guidelines in a project. And it

doesn't always automatically apply that.

With skills, you're able to use that

within a project context or individual

chat context and then ask it to create

that presentation for you by referencing

that skill and it would automatically

apply those guidelines and constraints.

So, it's very much again very refined

focused set of instructions that you can

use and you know if we're it fits really

well within projects. So, again the the

links are here take a look at it. They

really talk about kind of um how the

this fits into the workflow ecosystem of

claude and how you can essentially

refine your approach with the skills.

Now I have a set of um I need to analyze

some data and I've essentially given it

a marketing data, sales data, contact

data like some closed contacts and some

web traffic as well. And I've said,

"Hey, give me insights uh from these

documents in this format." And I've

essentially, you know, given the data

and the format I'm expecting and this

LLM is going ahead and and making these

calculations for me. It's ran some code.

Now, what's interesting here is two

parts. This is likely it is a weekly

task that I have to do every single

week, but also at the same time, um not

only is it repeatable, but there is a

possibility of hallucination or a wrong

calculation happening by the LLM. Now,

this is where skills shine where you can

essentially create a skill directly uh

within Claude and have it follow the

steps exactly created by you. So, I went

ahead and created a skill myself. I

said, "Hey, um I want to create a skill

that can analyze a set of Facebook data

through web traffic and give me the

insights I need on funnel conversions."

And I provided a sample file as

reference for it to follow. Now, you can

create skills directly in settings. So

should go to capabilities in settings

capabilities skills and then you would

do the skill creator and you can click

uh click on try and chat it would open

up a chat and then you say what skill

you want to create. So I've already

created one called the funnel analyzer

and what it looks like is something like

this where it's essentially the name of

the skill with a front matter on you

know a specific set of instructions it

has to follow. um and then uh script to

actually analyze the data and metrics

which is like a reference uh like a

glossery for it to reference uh as it's

calculating this. So I went ahead and

created the skill in the skill creator

and then um I loaded that skill into the

you know into the settings through here

as a zip file called funnel analyzer and

now we're going to actually try it and

see what that looks like. So, I'm going

to click on try and chat or you can go

and chat directly and just like write

use my funnel analyzer skill. We're

going to drop in the files here and

let's see what happens.

Ideally, it uses the skill. It follows

the scripts directly and gives me the

insights that I need without any room

for error or hallucination and the fact

that I don't have to repeat exactly what

output I'm looking for from the

insights. So, it's going through all the

files. It's reading the scale. It's

looking at all the files. It essentially

says, "Okay, cool. I have all the six

files I need." Now, it's going to

actually write the script. And it should

be Yeah. analyze funnel. Yep. Perfect.

Now, let's see what happens. And there

we go. So, we got a executive summary on

what our spend is, our revenue

generated, cash collected, and what our

raw assass is, and essentially what our

current CAC is. So, I took a repeatable

step, created a skill around it, and I'm

using add-on scripts to actually

calculate exactly um what metrics I

would need and insights I would need

from this. So, you can see it has a

little bit of a difference between um

what we had captured here um and then um

what that script is essentially

displayed uh earlier. Now, keep in mind

I had to just change the data a little

bit from this original chat and then

this new one. Um, just because I I

realized I need to like switch up the

numbers and take out a field cuz there's

an error and I was analyzing it. But the

point is what the key takeaway here is

we've taken a repeatable step, created a

scale around it and we're decreasing the

likelihood of any hallucination or

errors as part of this process. So I

think this is really really helpful in

the sense that we're not only getting

we're repeat we're automating a

repeatable step but also have more

confidence in the output and the

insights that we're getting from this.

Um anyways I hope I've covered kind of

how this is different from from projects

and sub aents how you can apply it for

your work as well and what you can do

like for example I did like a marketing

insights analyzer that was looking at

more topofunnel than revenue. I built

one for AB testing for Humbolty where it

essentially analyzed a site and gave me

um experiment ideas and um it gave me

some like suggestions on what to run

experiments for and even artifacts as

well. Hope that was helpful. Thanks.