Today I'm going to be walking through the different layers of AI capability for your business.

And you'll use this as a path to building your first agent.

G'.

Day, My name's Mike from Lone Wolf Unleashed, and welcome to another week.

So last week we went through how to build your first agent.

This is a layer on top of that, so if you haven't listened to that one, please go back and listen to that one first.

This one I'm going to be walking through a little bit more about the different layers that go into building AI agents that will give you the maximum benefit and, you know, trying to work towards making things faster in your business.

And I have fancy little slides here that I, yes, I did use an AI agent to build these, so you can see those today if you are online.

This will also be published to YouTube, so you can check it over there.

Make sure to go and follow my channel.

But let's kick right in.

There's three different layers that I'm going to be talking about today when it comes to designing an agent.

And they are going to be, in some way, they're going to be able to be embedded into whatever tool you're using on whatever plan you're using.

I've tried to be a little bit general here because obviously different people are using different types of tools, which is fine.

I personally, I'm on Claude.

I'm paying for Claude Max.

I'm getting excellent efficiency out of that and I hope to hopefully soon be able to put out some content about how specifically I'm using that connected with Obsidian and connected to my other different.

Now MCP servers go high level and Asana, which is pretty cool.

Three layers.

First one is connections.

This is an AI that can reach into your other tools.

Okay?

They're the task managers, your calendars, your documents, they connect to other tools.

The second one is the Playbook.

So these are your workflows, your instructions for how the AI agent is to operate.

And then thirdly is identity.

So who is the AI?

How are they supposed to think?

When are they supposed to push back on instructions?

I, in my custom instructions have that, you know, I want to be challenged on some of my ideas so I can get really good outcomes.

I don't want it to be a yes man.

Oh, yes, Mike, that's an excellent idea.

No, not all my ideas that I put into an AI engine are good ideas.

And, you know, because I'm using a lot for brainstorming and understanding client situations, I really do like having that little bit of pushback when I'm trying to brainstorm about how to deal with certain situations.

So that's layer three, identity.

There's different spectrums of what to do for using AI.

A lot of you have users as a chatbot.

Okay, you sat in ChatGPT and you have put in queries.

Maybe it's a sentence prompt, Maybe you've started doing your searches in there that you used to put into Google.

So that's a fairly rudimentary type use for it.

Okay.

It's capable of a lot more.

The second one is assistant.

So an assistant is something that helps you do your work.

It's not just search, it's not just a chatbot.

It's something that produces outputs for you to use.

The example there is to draft a meeting agenda.

So you have an agenda, it's templated, and then you're going to draft the meeting agenda for the quarterly review, right?

So you have a template which is standard.

And then the quarterly review content is the, the custom stuff that you're.

You're getting it to generate into there.

Then you have the agent.

So this does the work and it follows your methodology that you've defined.

It checks your task manager for overdue items, drafts the agenda, adds action items, and then asks, should I send this?

Okay, so it's more autonomous in terms of it has a broader scope of the content is putting in, but it now has more context about what to put into that quarterly review and what to do afterwards.

So it's got an expanded scope.

So the agent has three things that a chatbot doesn't.

It has hands, so it can reach into your other tools.

It has the playbooks, which follows your specific workflow step by steps.

And it has a point of view.

Okay, what type of specialist is it?

What is it supposed to be doing?

What perspective is it supposed to have when it's looking at your different prompts that you're feeding it, the different inputs that you're feeding it.

When we're looking at connections, what we're trying to do is we're trying to give the AI hands.

An agent is going to be a little bit like an octopus.

It's going to be taking information from different places and it's going to be feeding that in and then it's going to be pushing it out again.

Without the connections, the AI can only talk about your work.

It can't actually do any of the work with connections.

It can do your work in different AI tools.

I'm going to work through a little bit here.

So ChatGPT we're looking at different plugins, we're looking at GPT actions.

Okay.

So we're gonna be connecting to external sources and that's gonna be through your GPT builder.

So if you're paying for the Premium subscription of GPT of ChatGPT, then you'll be able to have access to those functions.

With Claude, it's the mcp, it's the model context protocol.

So you're basically installing the connections through the plugin marketplace or you're configuring that manually.

You can literally ask it how to do that.

So today I was doing this with GoHighLevel and it produced me an instruction list about where to find it in the settings.

Set that up so I can have Claude just attached to the mcp.

What does that mean?

It means that I can have a workflow where I can go, I'm going to add a task and assign it for someone to do something.

And I'm also going to add in the opportunity into go high level so we can track our pipeline.

You know, now we can execute multiple different tasks just from one prompt because it is able to reach into those different tools to do that.

With Microsoft, you're looking at extensions and plugins and then with Gemini you have the extension, so it's built in for Workplace, for the workspace and you can have your third party via other extensions, your common high value connections.

You know, I've already mentioned Asana, you might have your other project management type stuff in there.

There might be communications.

It might be, hey, once this is done and you can see that this is happening, can you send a message off to so and so in this via this channel?

You might have it update documents.

I have that doing that right now.

I have it connected into my Obsidian workspace.

I don't draft notes anymore.

Claude does that for me.

I review them.

Once I'm happy with them, they stay there.

If I need to go back and reference them, Chord can do that.

It can provide me summary of what those notes are and what the connections to those notes are and things like that.

It is epic.

Honestly, it really does change the game in terms of how you put all that information together, especially in a space like mine where you're having to deal with a lot of information, then we have permission.

So we need to be thinking about what is the model allowed to do.

Okay, so this just like any other user that you would put into your ecosystem, we have to be thinking about that for an AI agent as well.

So what are they allowed to do?

Are they allowed to read, are they allowed to create an update or are they allowed to delete?

So there's going to be specific instructions here about what you allow it to do and when.

And you need to define what those are.

Don't miss that, because otherwise files will go missing and you wonder why.

So here's an exercise you can do.

You can map your connection opportunities, right?

So list three systems you use every day.

Might be Asana, Outlook, it could be your Google Docs, it could be Google Calendar.

And then for each one, what do you do the most often in there?

So do you check what's overdue?

Do you set calendar meetings?

Do you do xyz?

And then which one would save you the most time if an AI could do it for you?

So I'll give you an example.

I have previously set up a telegram bot that will go to an NAN workflow to punch meetings and blackout calendars in certain calendars.

I do that because I have to manage multiple calendars across different clients to update my availability and that makes it easy for me.

I can just do it from one message prompt.

So you might try to do that.

You might say, I want to be able to set in meetings or things into my calendar.

I do that a lot.

I'm going to connect into whatever calendar system I'm using and I'm going to have the AI agent start to interact with that.

So have a go at mapping out your connections like that into the common tools that you use.

If you've gone through the process work that I've done before in a previous episode, and you've got some process models there, you'll have already listed your systems about which tasks are dealing with what systems and you'll be able to basically go to there for inspiration about how an AI might be able to help you do stuff there.

Then we have Playbooks.

So teaching your AI your workflows, describe to it, feed it one of your procedures.

I'm not going to dwell on this anymore.

Give it some of your knowledge about how that task is specifically supposed to be done.

The worked example is the Daily Standup Playbook.

Okay, so we have a standup that we do with the team.

Give me my incomplete tasks that are due today or overdue.

Give me some tasks I completed yesterday and then the tasks I'm currently blocked on, I can share those immediately and I can punch them through a channel.

This basically automates a stand up between say a manager and their team, or with a Scrum Master and a development team, things like that.

It Streamlines it incredibly quickly because all that information is already available in the tools.

Okay.

We're not having to have someone stand there and scroll through their tasks and just basically feedback what is already there.

Those people can now focus more on that.

And I know there are many developers out there who would be happy about doing that.

I know they don't like the stand ups normally.

So what makes a good playbook?

So we need to be specific about what the output looks like.

Right?

So if we're filling a document, what's the document template?

What information goes under what area?

For example, okay, you might show a table with columns for a task overdue due date, what data are we using?

Then we have the safety rails.

So you know before deleting anything you need to tell me, ask me to confirm, tell it what to ask, not guess.

Okay.

So if there's a particular framework you're going through, you might have it ask you specific questions so you can get good output.

And then you define the step order.

So previously I, in the last episode I went through an example of what it was to build an agent that has a workflow.

Okay, this one was a communications disk agent that would go out, do some research, come back, draft up a communication based on a template.

Okay, so that's a multi step workflow with multiple steps.

So you need to determine those.

And remember, start simple here.

We're not trying to conquer the world.

We are just trying to make some of the boring, laborious things that we do more efficient, more streamlined.

Identity.

So we're going to turn the AI into a specialist now.

So you've probably from the early days of ChatGPT, I remember a lot of the things was around context setting.

You know, I want you to take on this role so you can give me context back.

That still remains true today.

And we're going to still give the AI an identity so they know the context of which to respond to you.

It's kind of like when you're at a barbecue, you don't ask the gardener for finance tips, you ask the guy who's in finance, finance tips.

It's the same as here is we're giving the AI agent an identity that is appropriate for the response that you're wanting to receive.

So then we have the four components, okay?

We have the role, we have the directives, we have the knowledge domain and the routing.

The role defines who the AI is.

Okay?

The directives give the rules the AI must follow and will never break.

Make sure that these rules, the really important stuff, stays at the Top of the prompt and then domain knowledge.

What are the keys?

Concepts, terminology.

You might feed it a framework that you're trying to work to, you know, feed it some of my transcripts.

If you want to have a systemization agent, head over to my website, grab some transcripts down and see if it can start populating in some procedures and knowledge from you based on the information that I share.

And then you have your routing.

So when someone asks X, use Y.

Okay.

So you can start to dictate about when it uses certain other parts of your ecosystem based on keywords or actions.

So here's the full stack.

We have identity, playbooks, connections and permissions.

So who am I?

What are my rules?

What workflows do I follow?

What systems do I need to reach into and what can I do without asking?

So you might go away and build your own agent now, which could be a communications assistant.

So you know, you can build on working assistant for your business.

Right?

So we have a social post, client email, content plan, review response, elevator pitch.

You might choose to build an agent around trying to do some of those things, you know, and then we're going to write those knowledge files that the agent could refer to as their knowledge base to go.

When I've, I've been asked for this, I can reach out into this file and get what I need to do for that.

Okay.

It saves having to put so many words in just one really large instruction list because you'll consume all your tokens just trying to feed it the context that you need.

If you refer out to other files, it will only refer to those files.

This is building up in say, an ecosystem like Claude skills, where it will only consume the context for the tasks that it's been given to do and then you'll write your system instructions.

So I have an example there, which I'll share on my website.

You can find lonewolfunleashed.com There is a resources page there you'll be able to head over to.

And then you can set it up, you test it and you refine it.

So the first version is not gonna be perfect.

It's gonna have some bugs in might behave some ways, or it might have terminology that you don't like.

You can always go back in and you can refine what those instructions are.

You can even just get it to update its own instructions.

Okay.

That's what I've been doing with Claude is I will get it to edit its own file based on the feedback that I have for it as we go.

There's some common pitfalls there.

I'll also share them on my website.

Guys, I want to thank you so much for joining me today.

Learning how to better configure the different layers of what an of what goes into an AI agent.

Thank you so much for joining me today.

You could been doing so many other things, but you decided to hang out with me and learn about that today.

Head over to my website, subscribe to my newsletter.

I'm going to be pumping some stuff out there out on there soon to do with more of this stuff in more detail.

And as always, I hope you have a great week and I'll see you next Tuesday.