Introduction

And I'm going to try and give you a little bit of a different take on this.

From Q&A to Action: Extending AI with Tools and Playbooks

So APCPs are basically a way to give more skills to your AI system. And you've seen the example before where an AI system was able to plug into different tools and start to create a plan and push it out to another tool and do things like that. That was a fairly development-oriented example.

But actually, everyone in the room that's not an engineer, MCPs also are really interesting.

And the problem is it's very technical language that we talk about, but really it's like adding a capability to your AI system so that they can go outside of just a question and answer type paradigm. They don't stick inside your chat. They're actually able to act on anything else that you could do through a browser.

A Personal Take and Live Demo Setup

And I wanna give you a little bit of an overview of how I'm starting to think about this. Now, I need to get to my actual phone so that I have my, there we go. I have a little bit of a plan in front of me.

But the, obviously it crashes.

Building a Daily AI Assistant

The way that I have started to look at this, and all of this is extremely kind of on the edge, so if something fails, It fails, but I'm going to show you what I actually use on a daily basis. And at this point, it's really an assistant that has started to learn a whole bunch of my own processes.

The Digital Garden: Your Personal Operating System

And the way that I do that is I've created what I started to call a digital garden. And I think I'm going to try and make this more of a thing.

Why Documented Processes Matter

Whether the concept changes, the name changes, the idea is you need to have an environment where you have documented how you do things so that when you ask the AI to go and do things, it knows what you mean with those things.

Because when all of us here say, hey, I need a marketing plan for product X to go into new country, does that mean?

Or I want a meeting prepared. I'm about to walk into a meeting. Do some meeting prep for me. What does that mean?

Different people here in this room will have different definition of what does good look like? What do I mean with prepping a meeting? If you haven't documented that, your AI system is going to have a hard time understanding it.

It's going to try something, which is basically the average of whatever it finds on the internet it thinks is good. And that might just not be what you think is good. But when you have it documented, it can work.

Tools Used: Cloud Code and Voice-to-Text

Now, there's a little trick that I have within that. So I'm going to be using Cloud Code as well. But hopefully, you'll see how easy it is to use.

And I'm going to be using voice to text quite a bit. So when you see this pop up, it's just that my computer is starting to listen to my voice.

Cloud.md: Loading Operating Instructions

So, Cloud Code has a file at its root which is called Cloud.md, which is a set of rules that it loads the moment you start Cloud Code. It's nothing other than, it's kind of like a custom GPT. You have predefined instructions when you have a custom GPT and it kind of loads those as a starting point so that it acts differently to a standard chat.

Chief of Staff Persona and Context

Now, in my case, I've called them my chief of staff because that is how I now use my AI. And it has all the context of how I operate. So I tell it very little bit about me because it actually has a whole other folder about me on context, which it can go and grab if it needs.

Defining the Personal Operating System

More importantly, it has a definition of my digital garden or my personal operating system. Maybe that's a better phrase I should be using. My personal operating system has a bunch of things.

Playbooks: How I Like Things Done

It has playbooks, which is descriptions of how I like things done. I'll give you an example here, a whole list of playbooks.

How do I like external meeting prep to be done?

Example Playbooks: Meeting Prep and Internal CRM

Well, you need to search for my calendar for the meeting that I'm talking about. Only prep meetings that are with non-Mindstone employees, because I don't want you to prep meetings that are internal, like those not required.

Run my internal CRM researcher, which is another playbook that it also has. What does it mean to use my internal CRM? Let me actually look at where I have that here.

My internal CRM playbook says it is helpful. Search for any meetings with the person that I am talking about. Search for any emails that I've had with either the person or the company before.

I'm documenting what do I mean when I say specific things.

Execution Flow: From Request to Action

And then in the operating system, I say, well, whenever I ask you something, have a look at the playbook, have a look at the guides that I created, have a look at the context about me, take whatever elements you need, and then start executing on the request.

And I'll give you an example. of what that looks like.

Core Playbooks in Practice

Actually, just before I do that, I'll kind of run you through the types of, the ones that I use a lot.

Research Automations

The web researcher and the internal CRM researcher, I never use them directly, but I have a whole bunch of playbooks that make use of those playbooks in order to go and do what they need. If I go and want to do research on a particular area, it'll automatically use the web researcher to go and do that.

Or if I have meeting prep, it'll use both internal and web research to go and do that.

Meeting Follow-up Drafting

The ones that I use mostly, though, are meeting follow-up drafting, which automatically accesses all of my transcripts of calls. It then uses the internal CRM researcher and the web researcher to basically get all the context it needs on what was the meeting about, what were the previous emails I had, previous meetings I had with the same people, what has happened publicly with that company. It then goes and looks up examples of how I previously have written

follow-up emails, and it takes all of that into account and then drafts a follow-up email and stores it into my draft emails for me. That is how I do 95% of my email follow-ups at this point.

Live Walkthrough: Research to Proposal Updates

I'm going to run you through one of these. Let's see. Actually, do I have... I'm going to have a little bit of a... Actually, I'm going to do two things.

First thing, Have a look at all the meetings I've had with Dimitris from StarTech over the last few weeks as well as the partnership proposal that I put together in the content section and give me an overview of how after today's discussions we might want to update that.

Actually, I'll try and... There you go, a little bit bigger.

Scaling Tools with Super MCP

So the first thing I have, and by the way, this was interesting to see previously, because the problem was what happens if you have too many MCPs in a list.

Why Lazy-Loading Tools Matters

I built another tool, which I called Super MCP. I open sourced it, by the way, so you're happy to use it. It does, for the engineers, it kind of lazy loads the MCPs that you're using.

The problem I had was I have about 20 MCP servers, which are different tools. and the context window gets full almost instantly. Basically, I couldn't really ask it any questions because the AI was getting overloaded just by the definition of the tools that I was giving it.

I had to load this. I create this extra tool that would automatically go and figure it out. It has a few tools, which is search tool, use tool, and it only loads the ones that it actually needs at the time that it needs them.

Tracing the Workflow

You can see what is happening here. So here, I'll help you review the meetings with Dimitris from StarTech. Well, it didn't get StarTech. The second T, it never gets right when I say it.

But you can see, first thing it does, from the toolbox, it lists its tools. It finds that it has granola transcripts. It's gonna search for that.

Let me create a detailed to-do list, okay? Search for meetings with Dimitris, there we go. Now let me get today's meeting.

And then here it's going to go, let me look for the partnership proposal document. So it's looking into context first. I think it might will it find it Oh now it didn't find it initially with a simple search tool and I actually have a tool which does semantic indexing on my own File so it can locally search by meaning.

Hopefully it'll have found that okay. It's actually still using that so now it's searching more deeply my files across my machine to figure out what the partnership proposal looked like and Okay, now it's read the full partnership proposal. It's found it.

And then, here you go. This was actually the last one, because there were two. One that I had drafted and I had not sent you, actually. Okay, let's see.

From Context Gathering to Actionable Updates

Now let me update my to-do list. So now it's got all the context. So it has the meetings of today. It has the strategic proposals that we had from previously.

And now it should, it might actually decide to go and do more research, or it's going to now say, hey, based on the conversations of today, these are the updates that I think should happen. So let's see what it comes back with.

I have one of these problems. By the way, when you do these demos later in the day, they take longer than when you do them early in the day because all the computers are now getting divided between Europe and the US. If you do it at 8 a.m., it's the best possible time because everything goes faster.

What do we have here? It's got through.

Proposed Updates Summary

Overview the partnership proposal updates based on today's discussions.

Bottom-up approach, which we talked about today, how do you actually figure out how this first goes to different people in the ecosystem who then, once they have gone through the training that we provide, can then become advocates that in their companies are able to talk about, hey, this is actually transformational. Let's figure out if there's something we can go and do much more across the organization.

Enhanced post-program support structure, browser extension, dedicated AI monthly office hours, all stuff that we talked about today.

Document Annotation with a Strategic Advisor Persona

Okay, so now let me go and take this a little bit further.

Now have a look at your document annotator playbook and let's have a look at the draft proposal itself and annotate it for me based on the changes we need to make.

This is SuperWhisperer? That's SuperWhisperer, yeah.

Okay, so now I'm going to go here. I'm actually going to go find the content strategic proposal. Okay.

Change Suggestions, Not Direct Edits

So let's figure out what it's going to go to. So I have a playbook, which is the document editor.

And the document editor, I specifically created the playbook because I don't like it editing what I'm doing. I like it creating change suggestions, just like someone would in a Google Doc or Word. Rather than changing everything, it actually provides change suggestions.

Let's see if it actually does this.

Role-Play Editing for Better Outcomes

There's another thing in the document editor, which is I ask it to do the editing in role play. As part of the playbook, I ask it, take on the persona of someone that could help me edit this the best possible way, then simulate a discussion between me and that persona, and imagine the editor being a third person that's observing the discussion and editing, changing the document based on the quality of the conversation. So here it's saying, I'll act as a strategic partnership advisor, someone with deep experience.

Whoop, here we go. First things that are starting to get edited. And let's see what comes through here.

Based on these discussions, see what else it's suggesting. Here it's actually suggesting a section. Section needs to be clear about how much we go and pay for all this.

What didn't work this month? Open forum teams bring their specific AI challenges we talked about specifically. What is the thing that didn't work this month going through?

Iterative Annotations You Can Accept or Reject

You can see how it's now going back and forth. It's going through the entire document and annotating it for me to accept or decline whatever I think needs to go in or not.

I'm going to let that go through, but whilst it's doing, I'm going to put in the next bit.

Bringing the Advisor to Life with Voice

Now, for the continuation of the discussion to figure out what else we want to edit, can you please use your human voice to bring alive the strategic advisor? Let's use your human voice to make the strategic advisor really come to life.

I wonder if I need to put my sound up. Hopefully, the sound is going to get mirrored through that. Otherwise, I'll... I'll do that on my machine and you'll have to deal with that.

Right now it's still going through the editing, so I think it's almost at the end. So that probably was the last bit. Okay, so.

It actually didn't take the last command, so let me just check. Probably, probably it was getting used there. Okay, so good.

It's identifying what needs to happen here. Speak11 Labs is indeed the tool that it should be using. Let's see where it goes.

So again, if I do this at 10 a.m. in the morning, this goes much, much faster.

A Candid, Challenging Advisor Perspective

I think the audio is not working, so I'm going to... Right, Joshua. Wow, there we go. Transparency, transparency.

Let's cut through the noise here. I've been looking at this proposal, and while the vision is compelling, I'm concerned about the commercial structure.

You had a really interesting conversation with Dimitris today, and what struck me was his immediate enthusiasm for two specific things, the epignosis bottom-up play and the AI coach co-investment. He literally said, I got excited when you explained the habit formation approach.

But here's my challenge to you. You're asking for a which Dimitris acknowledged is significant.

Yet you've just added five major new components to the proposal without adjusting the investment level. That's either incredibly generous of you or you haven't thought through Should we flip this entire proposal on its head? Instead of leading with quarterly strategic days and monthly office hours, which feel quite traditional, why not lead with the disruptive stuff?

Imagine restructuring it like this. We're going to help you monetize your 20 million Epignosis users while building Greece's AI talent pipeline. Bang, that's the hook.

Then you explain how the bottom-up approach creates advocates, the executive dinners create top-down pressure, and the AI coach ensures it all sticks. The strategic days and office hours? They become the supporting infrastructure, not the headline act.

What do you think? Are you brave enough to completely restructure this, or are you too attached to the current format? I mean, we're joking and it's going through, but those were real questions.

Designing for Constructive Dissent

These are the types of questions you want your AI to come back with. One of my lines in the Chief of Staff is, I don't need you to agree with what I'm saying. I need you to be real and challenge me on the stuff that I'm asking you about. And that's what I want from the AI assistant.

Handling Failures and Timeouts

All right, so now we'd go back and forth. For those of you that would have looked at the terminal a little bit, this was a much longer than usual chat that would have come back to me. So the MCP actually timed out.

Configurable MCP Timeouts

You can change this. I forgot to do this here. You can change the MCP timeout so that it just stays for as long as it needs to.

When Integrations Break: Contact Book Issues

But just to show you something else, and there was something that went wrong earlier today, which is a shame that I have this clay MCP which the ClayMCP is supposed to be my contact book. So it has information on everyone that I have in my phone contacts, but somehow it played up. They had a new version earlier this week, just messed everything up. I can no longer use it with the Assistant.

Simulated Calls as a Workaround

So I have one step that makes my live demo just a little bit less good, but it's still going to work, which is... Can you give Dimitris a call and just ask him what his view of today's meeting is and if there's anything really particular that he thought was important to take into account? I'm going to say number. So this is the bit of the demo which is a shame. I actually have to give it the number instead of normally, like literally three, four days ago, it would have been able to just look up Dimitris in my contact book and then go and do the call.

But let's see what comes up here. You might want to put your volume as high as you possibly can as well. So it's got another tool, which is basically, I use Retail AI, which allows you to simulate phone calls. You have to link it to Twilio, which then kind of does that link. Okay, so it's now configuring that.

Hello? Ah! Okay, so it's not gotten the proper instructions. I'm gonna call that. If you can hang up the phone, actually, that'd be good.

I'm gonna try one more time just to see, hey, live demo being live demo. So it looks like you didn't update the actual instructions, because Eva just said hello instead of everything I asked you about. Can you just double check that?

Diagnosing and Fixing Live Issues

This gives me an opportunity to actually show you how I would go about fixing this problem now. So one, so I do this and hopefully this will do this again. Let's see if... Okay, there we go. So let's see.

So basically what it did is it launched the call without it having reconfigured the actual context of the call, which is what this was about. So it was just a general call. Now, here you go. You actually see it's updating that specific call. And now it should. This should be much better. Okay.

Hello. Joshua has asked me to give you a quick call to get your thoughts.

It went very well. It's a very exciting opportunity for us. Wonderful to hear. What aspect of the partnership proposal excited you the most? Was it the two-pronged ethognosis approach, the AI approach concept, or something else that really stood out to you?

First of all, the possibility to be able to offer something meaningful to something like 20 million people, it's really fantastic from a number of different point of views. But I also liked a lot, you know, this possibility for us to really develop the Greek tech ecosystem around AI and the business ecosystem in general.

You might want to hang up because it's going to be 20 minutes. It's crazy that I had no idea what they would say. Hopefully you understood.

It literally took the context from everything we just went through. Obviously this was not staged. You saw me fail the first one. It literally just takes the entire context of the conversation and goes through.

Learning Loops: Evolving Your Playbooks

Now...

Postmortems that Update Procedures

all every every there is an opportunity to show how this actually works for me right so this is now a thing okay i'm gonna say so the second one clearly worked have a look at your playbook for phone calls and update it in a way to avoid you ever making this mistake again So now I'm getting it to just look through what did you get wrong? Why did it work the second time, not the first time?

Update your playbook to not repeat this mistake again. 1This is how I now start to work with my playbooks.

Auto-Generating New Playbooks from Sessions

I have enough of these playbooks that I sometimes I just go in, I tell it to use tools very specifically, and then at the end of an interaction, it's like, actually, this is the type of interaction that I would have quite a lot. Go and look at all the playbooks so that you get an idea of how I write playbooks. Now, look at the interaction we just had over the last 10 minutes and write me a new playbook, call it XYZ, that allows you to go over these steps the next time that we go over it.

Concrete Improvement: Preventing Misconfigured Calls

Okay, so here, just to show you what it's added, warning the LLM retains its previous configuration between calls, you must update it before a new call. That's now part of my playbook going forward. I'm not confident like 100% that it won't ever make this mistake again, but it'll definitely reduce the number of times that it does.

Conclusion

Okay, I think that was the end of the demo. Hopefully it was interesting.