Introduction

So, hello everybody. I'm Chris. I'm a research software engineer at StarTech.

What this session covers

And today I'm going to show you a little thing that I made for a colleague of mine. And I think it can serve as a good template for many other uses. Possibly many things, almost anything you can imagine.

Who Cloud Code is for

And also, I especially want to introduce Cloud Code as a tool not only for developers, but pretty much anyone.

Why Cloud Code

And Cloud Code has some advantages that the browser of the desktop app lack. So yeah, much better.

Key advantages over browser apps

Yes, essentially, it has the advantage of being able to access your file system and write, read stuff you have there.

Also you can even utilize code, it can write code for you just by describing what you want and run it for you, even if you don't have a single idea of how a code works.

Reliability and reproducibility

And also it's hardly reproducible.

You can write a list of, let's say, a process and it will follow it to the letter.

It's not random like you give a bunch of instructions and expect it to give the same result each time, or a consistent result.

Feature parity with browser versions

And it essentially has all the pros of the browser versions. You can search the web, run code, as I said before.

It can

Getting Started Requirements

The only things that might be holding you back at first is that, essentially, you need someone to configure it, especially if you're a non -developer.

Initial setup for non-developers

Essentially, set up the CloudFly and then Python. And you need to familiarize a little bit yourself with what is a CLI and what is an ID.

What I will demonstrate

But other than that it's a really powerful tool and I'm going to demonstrate that.

Use Case: Social Media Post Automation

So essentially a colleague of mine wanted to, he essentially wanted to streamline the way he creates posts for various social media.

The manual workflow before automation

So the thing he had to do before the tool, before Cloud Code and the tool, the solution I provided him, was let's say he had a text like an article like this one.

Cross-platform constraints and citation needs

This is from an older event and he wanted to create let's say Twitter, now X, Facebook, LinkedIn and Instagram posts for platforms which are vastly different both in their style and in their restrictions and also you need

to search again in order to cite a person so a person has a different LinkedIn profile Instagram profile same goes for companies and so he wanted something that could take care of the whole thing and I'm gonna run it for you

Live demo plan

just to see how it works and then I'm going to rebuild it from scratch so that you can see the process.

How the Command Works

So essentially in the current condition you just have to write a command and give it the article and everything else just runs on its own with Cloud's agentic flow.

Entity recognition and web search

Now essentially it reads the text, it can discern and recognize entities such as companies or people inside there and then search the web to find related information such as their social media links.

So essentially it's searching now the speaker, it's searching Yotech and StarTech and wait for it.

Reusable commands in Cloud Code

So essentially Cloud Code has a thing called commands where you can even define your own just by describing what you want. You can even say, like, make me a command called whatever and describe what you want

and next time you enter Cloud Code, just like this time, you press the slash, you can have a reproducible process waiting for you.

The POST command

And that's what I did, I created a POST command, which is a social media, as you see, copywriting expert. Those are saved in here, but you don't even need to know that.

It's still searching the net. I taught it to be thorough.

After that, I'm going to show you a trick to make it even faster. Let's have it finish though. Oh, it's here.

Demo output overview

So, essentially, it outputted me a text with all the information someone would want to make those posts.

It has a Facebook section, as you see here, an Instagram one, a LinkedIn, and a Twitter. And if you go in the end, you can even see the links on how to cite each one properly.

It has searched the web thoroughly and presents you with the links you need to cite them and how to cite them. So everything is streamlined.

Extending with Local Storage

And the thing is, as I said before, one of the powerful aspects of this tool is that

you can essentially read and write files.

Persisting profiles to CSV

So, since we fetched all that information and companies like StarTech and Yodeca are almost always going to be on the posts, so why not store them somewhere? Like let's say a CSV file.

Helper scripts for read/write

So essentially the thing I did was I asked in the instructions the agent to create scripts scripts to read and write on the CSV file, which is like an Excel.

And you can have essentially the links here and inspect them. And it also has stuff like how confident it is that those are correct.

Speeding up with cached data

And also next time it searches, let's say I do it again.

This time it's programmed not to search again the web, it does search locally and since those profiles are here and essentially it won't have to search the web and use the accurate profiles that hopefully the person has moderated before.

probably they're moderated by the person, it can finish, it is finished already, so

Putting It All Together

And so this is a comprehensive tool that uses web search, it utilizes code without you needing to write write code and creates local stores of files that you can then edit, share or keep.

From prompt to working system

And so this essentially the whole thing was created with just a text and just a simple recipe let's say.

Recipe-driven implementation

I essentially prompted this file which is broken into steps, you don't have to be be so specific as I was and essentially it has some instructions about how to create the project, what it does, how to store, in which format to store, whatever it stores,

how to search, how to act, you can even add more instructions like our company policy on writing posts on this platform and this and append them here and I can reproduce the whole thing here by just writing implement and then give it the file and

When to use Cloud Code

essentially in five minutes from now it will have recreated the whole thing with an input so small as this and so if you have any process that requires web search file editing and maybe utilize code for some reason maybe calculate something.

Why it’s a powerful fit

I think Cloud Code can be a really good and powerful tool

Sharing the recipe

to help you and essentially I also have here QR code to serve the recipe for anyone interested and use it probably as a template to create something similar.

Rebuilding the Project Live

I'm gonna leave it here a little bit while our agent is is creating essentially the project from scratch let's get back to our agent it's still doing its thing see now you don't need to know what happens here but you can always refine the outputs

Customizing instructions

essentially all right it hasn't created it yet as you can see here i asked it to create a special document in Docs where you can write whatever instructions fit you best on how to do the process.

Controlling output format

So essentially in what form you want the output, what style maybe, if you want spacing, if you want the links below the posts for easy

Stay in chat—no file handling needed

copy paste and you can you don't even have to to uh you don't even have to act on the file you can just ask cloud here to give it the file and tell it i want next time to do this like that so you

Command creation status

don't even even have to touch to touch the file system so yeah it's almost ready it created the let's say instructions, it should also now create the command, the post command

Running the recreated command

which is prompted from here, so now it's pretty much ready and if I were to copy let's say the article from here and paste it on the new one and ask, okay it's

almost over, okay perfect I'm gonna interrupt it and I'm gonna close and reopen cloud to take to read the new command and you can essentially specify as many commands as

End-to-end reproduction

you want and now paste the text and now it's going to do the exact same thing i did but i reproduce the whole thing with just a simple recipe which anyone can adjust so again it's going to do the whole thing again and that's pretty much it any questions you said it very

Q&A: Output and Consistency

Thanks for this. Is it possible that we see the output of this?

Where outputs are saved

The output, sure. Right now it's created in this file. Here, this is one created from the previous run.

Sections in the generated output

It's over here. It has a Facebook section, an Instagram section, a LinkedIn one, a Twitter, and the links below.

Tweaking instructions for layout changes

and let's say I wanted to have the links appear next to each social media I can just write it I mean adjust your instructions so that next time so essentially adjust the instructions so the links appear next to the respective

Reproducibility vs. determinism

platform and if I run this it will essentially update its own instructions to do exactly that and this is completely reproducible.

Each time you

run it you can expect a consistent result unlike let's say custom GPT or simple chat process.

Self-updating instructions

So essentially now it's reading its current instructions which are stored here, and it's updating them. So it updated its instructions and also the file for us to see.

Working entirely in chat

And you essentially don't need to leave the chat interface. Any other questions? Sure.

Q&A: Commands and Determinism

If the MIPER . If I'm understanding this correctly, So you're executing a command, you like that command, you like that prompt, let's say,

Are commands stored for reuse?

and then the AI is storing that command so that we can reproduce it later so that it has the exact same output? Yeah. Is that right?

Where commands live

Essentially, it created here in Cloud Commands folder, it essentially created the definition of this command, this post .md. So essentially, it's time you invoke this command,

command, it will read through all these, which are your specific instructions on how to do stuff.

Will results be identical every time?

And it's 100 % the same? I find it hard to believe that every time you run this command, the output will be exactly the same.

I understand that it may have very specific steps and that the AI will generate approximations of basically the same output, but is it exactly the same? No, it's not exactly the same.

Practical consistency vs. exact sameness

there's some randomness in language models anyway so you could never have the same the same result ever but still you're gonna get something about the same line it's about once you're taking the human being if you have to do it thus it's gonna do it one way and you if you ask again for the same task it's gonna do the same thing but maybe a little differently

Is there a standard process?

Is that your process or is there a standardized process about doing this, about storing this? This specific one is my process.

Authoring and adjusting commands

I essentially ask it to create a command through chat, through prompting, and explain what I want this command to do. The more information you give, the more accurate the command will be and effective. And you can still adjust it. I can tell it to, you know, take this post command and I want this differently next time. Still, again, without leaving the chat, ever.

Yeah, I love your process, by the way. Thank you. Sure.

Tips and Related Tools

Pandoc for format conversions

Just to help out people who are going to try this, there is a very nice open source tool that's called Pandoc, P -A -N -D -O -C, and since you're using MD files, which is markdown it converts many documents to mark them.

You can actually write any stuff on, say, a marked word, and just put it on the folder and it will create an empty file for you, and vice versa. So then you can have output and doc files.

Awesome. Still, it doesn't have to be... No, it doesn't. No, of course not.

Other document and data workflows

But it's a very neat tool if you want to, say, put an Excel file in there with some of your own data as well, or a document that you got from someone.

Yeah, very nice. I'm keeping this one here open, for anyone interested.

Closing

Any other questions? Awesome.

After-session support

We will also have time afterwards to be chimed in here to connect and ask any questions that you may want.

Thank you

Thank you very much.