Introduction

Just before my demo, I want to show you something that we are developing internally, which is the way we work with these tools right now and we put in practice a lot of the things that Henry has discussed, but we can do like for two or three minutes just that quick reverse prompting and you can see the screen, is it okay?

Setting up a live demo

So to do this, we need a problem. What kind of problem do we want to solve? Any ideas?

Reverse prompting for job search

so we are here trying to find a new job so I would like you to help me in this research think through our job search step by step and ask me one question at a time and these two crucials are these two questions are crucial here think through our job search step by step and ask me

me one question at a time so I'm flipping the script it's the reverse so rather than asking questions to the model I'm asking the models to ask me

Assigning roles and context to the AI

questions and I would probably add here imagine you are a very experienced recruiter with plenty of knowledge about how the job market works there's another

thing that Henry said which is the role the importance of the role giving the AI a role or a context and this is very we are doing this very simply by giving the AI a role a context on that is a persona that they should assume first question

Kicking off with clarifying questions

what type of role are you aiming for next examples what are we looking for

Speaking to AI vs. typing

leading role I always speak to the models when I use AI and that's for two reasons so these an external app that is called super whisper and it's for two reasons first it we tend to speak faster than we type and also it changes the

dynamics of interaction with a model we reveal a lot more detail when we speak and that detail is useful to inform what the model will do so we are not there with two or five fingers or ten fingers typing things but we actually speak to

On capturing tone and emotion

the model and share as much detail as we can capture question for you does it

capture the emotion like I need a job now and I'm looking for a job it does it change well it cannot doesn't listen to the voice but depending on the words that you use you would and in the frequency of the words and now we are

Guiding thinking with structured questions

framing things it will adopt the style which branch of engineering by the way I'm improvising here so I mean I know the technique but this is just to give you a sense of how you can structure the conversation and so it's it's guiding guiding our thinking it's saying all right which of the of the different scenarios we want to look at.

I think I will stop here.

Key takeaway: flipping the interaction

1What I want, my point was to show to you, I mean, tagging to what Mattia was saying, reverse prompting.

And so can AI say no?

How reverse prompting changes the flow

In this case, in a sense, we are bypassing that question because we are really flipping the terms of interaction. It's more about AI replying to the questions that, or AI asking me questions

guiding my thinking and I'm using the model too and I'm taking those answers and co -generating co -creating with a model no it's an additional technique

When to use reverse prompting

that you should think about your tool stack and can be very useful especially when you're trying a new tool and you want to see how it relates or when you are not quite sure when you basically have a problem to solve and you're not quite sure

about what kind of context you need to give to the model or how you should be interacting with the model it depends on the on the prompt that you need to write and how you're planning

Deep research prompts and tooling

to use that prompt if it's a deep research prompt definitely and who knows here about deep research and I think of deep research yes I'm really happy we've got

Evolving model behaviors

did the wife and we got the Wi -Fi working all right so and actually now if you notice they start putting this reverse or even adversarial prompting it

Models increasingly ask clarifying questions

happens sometimes you didn't ask I always ask to ask me question when it's important I always ask that ask me ten questions they're all smart but even when I don't ask now I said before we start can you clarify ABC my problem

quite long something so I see that now in the mother they start putting that before it was really as you say you're writing always yes they don't they go there you had to now they start asking so the mothers are that's why we say

Prompt techniques entering the mainstream

over the the prompt technique they are evolving because the mother will start integrating them in the mainstream can I can I show can I show the so like let's

From prompt to research: Perplexity example

say that by the way we don't have enough information here it's just for the sake of argument on the basis of the information can you put together a deep research prompt that I can run a perplexity to do some research on this

Generating a deep research brief

subject and so now this is giving me produce a comprehensive

structure analysis of the current job market for data science roles including applied data science machine learning engineering and it's sort of inferring

things that might we might remember we have told the model you are an expert at this so it's it's basically catching information from that side of expertise is and is creating a prompt on the basis of that information so we could we can take this

Why Perplexity for deep research

and then how many of you have heard about perplexity a few and because perplexity so perplexity came up can i say this well it is it is a lot better performing this particular deep research tasks so think about it as a mix of Google and an

LLM so it would perform a hundred depending on how much power you decide to give to it he would perform a hundred two hundred three hundred searches in parallel and he would then summarize them with an LLM Claude is a comparable function charge EPT is a comparable function but they are not as powerful

On paid plans and productivity tradeoffs

And you have the pro version. Yes. I've got the pro version because I've got... Paying. Because...

You have the question about Milan as well. Sorry? You have that question as well. I think you need to clarify your strategy about paying. Because this is what they ask you in Milan as well.

Yeah, I think we... Ultimately, we're a company of ten people. and so we we are effectively we have made a decision of expanding the IT budget as much as we need to be as productive as we need to be I mean these

in the grand scheme of things is 50 I think 50 years a month for 50 queries I would be saving half a day per query this is baking let's leave it there

Introducing Mindstone’s agentic system (Rebel)

because I actually want to show you how we would work at Mindstone these days so this is the system that we are using and so the way you could think about you

Think of it as a chief of staff

should think about is as a chief of staff so I can give quite vague instructions to the system and then the system would go and execute on the base of my instructions it has got access to my gmail it has got access to my slide generating tools

so you can just create files or documents that i can look at but let's ask the system itself

Demo: asking the system to describe itself

to describe what it can do hello rebel I'm here with what about okay a good crowd for our Paris meetup can you tell the audience here what you do what your

capabilities are in a few sentences so now what it's doing this is connected this system is connected to our company knowledge all the tools that they might be using note takers Gmail so it's just pulling the information and giving us an

Capabilities overview from the assistant

answer right to me the parties meetup crowd I'm rebel Stefanos AI assistant I help him with emailing communication drafting follow -ups outreach emails sales

and support I have access to Stefanos workspace with skills memory and the ability to learn from past work so you the way you should think about it is a series of prompts that are continuously updated automatically and they basically allow me to by just giving very brief instructions to execute on very complex

Running multiple tasks in parallel

tasks we don't have much time so I will I will start three in parallel because we are launching this next week so yeah i'm just giving you a peep so i will start three things in

Working like a manager with spinning plates

the ability of the the advantage of this with respect to other tools is that you can actually start working really every person can start working with like a business leader i remember

this this manager in italy once told me and this he was overseeing 5 000 people my role is to i see like my role as having all these spinning wheels and i have to go there and like poke them from time to time and ensure that they all keep spinning and this is how every each of us could

Access and rollout plans

potentially be able to work in the foreseeable future or today if you in a milestone case and like your students they have access to it while they're subscribed so we we are planning to do

What "launching next week" means

parallel sorry you said you were launching this next week yes so this is what what does it mean

the training also on rebel but we also have a training program that would cater generally available tools just we developed this because we realized that the the tools were coming the tools that we could see in the market were coming short in uh in different ways and hopefully you will see

Tech note: MCP and tool linking

why so the mcp is the way we the technical term that describes how we link to various tools

Live tasks: LinkedIn post, follow-up, slides

so it's basically a new standard it's like the api of ai pretty much can we write a linkedin post

about this event rebel i will do three things in parallel the post thank you i will do the post then i will do another thing in parallel is i will write a follow -up email based on a meeting I had in October with the MindStone organizers and then I will

do a third thing I will take the transcript of Harry's presentation and I will generate a PowerPoint presentation on it all in parallel we're the meeting

Follow-up email: prompting and checks

with the MindStone community organizers in October 2025 can you write a follow up email about that meeting and yes it's quite far away I know don't worry about it just write the email it's the first thing that the system will do is to ask

me this is so late why are you I mean this is if I were to give the instruction to a good chief of staff the first question would be have you not sent it before it's two months can you look in the downloads and there should

Summarizing a talk from a transcript

be a filed called homo promptus is the transcript of an intervention during the the Paris meetup.

Can you just take that transcript and produce a quick summary for me?

Just need to correct that.

Reviewing the first LinkedIn post

We can review the first LinkedIn post.

Last night, 60 people gathered at Kedge Business School for Paris' last AI meet -up of 2025.

More, more, more, more. It has an absolution. It's an Italian AI. It's exaggerated.

Not so yet. Well, it has access to the registrants. Ah, it's the registrants. There you go.

Reality vs. generated content

But this is a good lesson, though. It is. See?

The reality and what is written. And you have a percentage of people showing up.

Draft outputs are starting points

Three sessions, three angles on how teams are actually adopting AI.

I show the more than a stock for business leaders on a prompt to share lessons on organizational transformation and we were meant to have this fireside shot I think we might skip it but I think you're in that he knows that

you're co -hosting so it's pulling that information the point is not that this is perfect by the way the point is that I've got something that I can copy paste and sent within one minute and I'm not spending five or ten minutes writing

Summaries and source materials

this now we've got the summary of Harris presentation the one millennium is here

if we look at the I will show you what this is based on that was the original file I was I recorded with my phone which by the way this is how I do presentation I write a script and after I create with tools the slide from the

the script now I write I like what you say there is an advantage when you talk you to you give much more we talk about intentionality when you want to be precise you choose the words the structure the writing for me at least

Auto-generating slides with Gamma

it's more precise can you create a gamma presentation out of this and can you share the link with me so gamma how many of you have heard of gamma so I'm basically saying go on that app that is externally sitting on the web and create

Email draft review and iteration

a presentation for me in the meantime I've got the email ready in my inbox so let's have a look 7 -eleven let's have a look at email I realize this follow -up is coming quite late remember the meeting was in October I wanted to

circle back ah got the date wrong well community organizer meeting a share a summary of what we discussed and based on the transcript of the call because I had a note taker on the call it's creating that the email let's go my

signature again I would spend some time editing this but I didn't I didn't have to do any copy pasting like the basically that the time between the time gap between thinking and execution as Henry and Mattia were saying is really

Under the hood: Anthropic-based agent kit

getting compressed yes yeah it's a good it's a good question so we it's based on the cloud agent kit so we are using the club right now it's it's linking to the cloud models and we have basically built an old system of prompts around it that that and some guidance on how they should be updated yeah it's creating the

presentation any more questions is this making sense perhaps you can ask to

Prioritization and personal productivity

prioritize you know you have two hours left before leaving but you should be doing based on the email you receive how which task you should prioritize today Can you tell me what I still have to do today by looking at my calendars and the emails?

Integrations: calendars and to-do lists

And you could, for example, in principle, link it to a to -do list on Notion, on any other tool you might be using, so you might actually be able to access that to -do list and check which I'm not doing and optimize as well your calendar yeah I

think you dialogue you know on Sunday evening that's my calendar does it make sense the weekend any more questions or doubts yeah and it keeps the memory up

Perplexity results: breadth and sources

to date oh we need to check perplexity so these again as we were waiting you see yes so we got a detailed report 1025 data science job market yeah all the landscape definitions and strictly speaking this was based on a very basic

and I probably I'm not afraid to say on a bad prompt and the good thing about this paid version that Henry doesn't like is that you can actually you can have a you can check all the sources separately you also have the images that

were found and it's probably five ten pages and you see we were able to go through a hundred and forty sources let's see if we manage so in the

What got done while we talked

meantime what have we been able to do this I had to skip because of the meetup well these are the emails that are waiting for me once i go back home in the meantime

presentation was done i'm not sure it's going to be as good as your slides but i hope so i use gamma 2 so it's better it's pretty good it's not bad my friend it's always um sleek it's always

neat that's the messaging then you have to pay attention for more the images i mean the

70–80% outputs and human refinement

point is not I guess I think we have this underlying assumption that the outputs that technology produces should be perfect here the point is not that it's perfect the point is that it's 70 80 percent there and now as the manager

as the one was that the knowledge I can go here and because all of these is editable you asked me when was in Milan and for tell you how I do presentation and I've done that for 20 years.

Designing for intention and emotion

So not only I do the transcript, but when I brainstorm about the transcript, I say, can you create a gamma prompt for that slide?

And I'll say, what is the emotion we should? So I go at this level.

And not only after I go to gamma, not automatically like he's doing imagines, but in each slide, there is a brief of the emotion.

This I would have not done, personally.

Automation plus augmentation

only you know so this is the intentionality but obviously when you combine automation within with augmentation it's wow and you're right after you need to refine you play a little bit

with gamma you can refine the now they they will even go to kind of emotion space as well so i mean

Agentic editing of slide structures

let's let's let's show that can you send me this slide that's i think i think that's that's the ultimate test can you change this light from five bubbles to a pyramid and again

see the agentic interface you can talk to it you can change things and I do it here because it once I've got the thing here then it's I can operate on it

Henry, I think we have the new slide for you. What is this? The five cognitive senses. Nice pyramid.

Model limitations and representation choices

So this is a good example as well.

He doesn't understand the correlation between the topics He doesn't understand the context.

Pyramid vs. wheel: choosing the right metaphor

because this is like Maslow, like you think the one on the top is more important than the one on the bottom. So if I show you that, you think that, you know,

the ultimate one is analyzed and it's a wheel. For us it's a wheel. it's a wheel it's a process so it would be become much more precise yeah any more question yes this

Prompting with context and intent

Using reverse prompting to get unstuck

this this system or thinking and not using the air not using to produce I think it's that reverse prompting that I showed you at the beginning yeah sometimes you sometimes you start your it's that sensation that we all have I'm stuck you put a bunch of context you throw it to the model you say ask me one question at a time be my coach help me out i know when structuring the prompts

Balancing constraints and creativity

how do you avoid or control for if you have an overly prescriptive prompt does it prevent

creativity from the ai that could produce new ideas that you hadn't thought of so if you're overly restrictive or you're overly prescriptive in the prompt it might you know you say yeah

totally there's no order in this so it should be a circle uh nothing takes preston whereas actually it might be that actually you hadn't thought of it as a pyramid the pyramid was a

Let the AI interview you

good idea totally i think i go back to that flow of reverse prompting because i think okay i want to find the best shape if again if you decide to go the ai route you can go and take a walk and get some inspiration outside if you decide to go the AI route and you're

stuck just very quickly say okay let's play here I've got this problem I'm not sure about the shape ask me about the shape for this slide ask me a little bit ask me one question at a time to guide my thinking I personally I'm more of a

A quick Slack automation demo

fan of the walk myself but oh I will do one more thing because I promise the

team that i will be doing this can you send a message to the team on slack saying hi that we are live from the meetup in paris because it's also linked to our slack which is like our teams the teams that we use i didn't i didn't specify the channel any more questions are we waiting for

On roles, personas, and quality of answers

this i did a prompt yesterday night telling that you know you told to give it to give the llm a role right so i i told it to act as a university which is the oldest university have the best professors you know of business schools and give me like 30 points uh so that i don't have to do an

because you are I mean well what I will say is it just depends on the context that you give to the ai and the role that you give to the ai and the so in a sense it can just play any persona that you would want it to play so the answer is only relative to the information that you give i think we need to get away from thinking that the answer is absolutely true and just depends on how much information we put in

Posting updates with safeguards

we post on the general channel so the team is not worried and they know that because we are working on this live they know that the demo that went well that Wi -Fi was

working of course you have my permission my friend yes so you have to approve all of them one by one and we also have a security layer on top so you can basically disallow certain actions ex -ante so for example the system cannot send an email on my behalf that's disallowed just can't but you could have

Security, privacy, and governance Q&A

have used a transcript of mine without my authorization. I ask for it, then I. Well, now in the role play. But when you do automatize like that,

if I send you an email with some sensitive information among the 3 ,000, when you automatically pull that, do you think you're going to ask the authorization to all the pieces of information?

Well, it just depends on which systems as a company you decide to plug in the model. I have a question about that.

I have a question because whenever you plug your MLM with your company's contact, at one point your company's contact goes into the MLM as well. So there's no security for your personal data or for your company's data. But the...

Because I feel like with ChimiCad, you know, that you can create with your own stuff. Like, so your personal or company's data are not mixed with the whole element. So how do you, why do you take the decision to work with the code and how do you manage the security of your company's data?

Hosting choices: self-host vs. enterprise offerings

Two options for my answer, basically one is you host your own LLMs, which honestly is not the path I would recommend by default because you will realize that it's very expensive to grow your skills, set up, you don't have the economies of scale, etc. So don't do that unless you have very good reason to do it.

And by the way, the cost of your token will be very high.

and second one now more and more you have this enterprise offerings from the big players anthropic is going to there so basically they will provide you their lms but hosted in your virtual network yeah correct so uh yeah uh but uh prices are not disclosed on the landing page

Costs and tradeoffs

Yeah, it's expensive to work with. In a way, it's also the same expense. Bring your own cloud is as expensive as self -hosted.

This is for corporate something, definitely. We are in the same boat.

I live with it every day in my job.

Live status updates and team size

In the meantime, our message got posted on the Slack channel with everyone else.

So we, I mean, the team is... 60 of us.

Oh, didn't check that. Yes?

Legal considerations and disclosure risk

It's just kind of building on that, and I don't know if this is something only I've thought about because of what I do, and people that love technology and want to embrace it, they might be worried about cyber threats or security, whatever it is.

It's just something that came up at a presentation the other week, and I hadn't even thought about it in that much detail. but something to be aware of and think about that hopefully never happens but let's say you get

involved in a legal dispute and the output of the AI that you have on your computer is whatever the output was is relevant to that dispute and the model has taken things from your entire you know your whole company everything on your computer don't have to take in the orange choose the right words.

Is that all disclosable in disputes?

And the second thing that's very, very important is legally privileged materials advised from your lawyers if you are engaged in a dispute down the line. That should be protected from disclosure.

So your lawyer might have said to you, oh, you've got a crap case. You're never going to win. Don't do it. You're doing

it anyway. But if the AI, if you then use it for a commercial question, that the AI has used legally privileged information to help generate that response, is privilege now waived over those materials?

Again, in a dispute, are they now disclosable to the other side? I don't know how often that comes up in the It does.

Data readiness and target customers

I think you just need to be very good as an organization in getting your data in order. And I think many organizations are not at that stage.

For us, it's, and for this product we would be targeting people organizations of a size of 50 to 100 to start with because they think we think adoption is going to be faster because

that as you scale you're just going to have a lot of complexity in getting your systems in order so this tools like this can be used

Policy-enforced outputs and DLP approaches

things is like, when you say disposable, it's like an employee in a way, who knows everything, because he was, you have an accountant, and an accountant is not meant to disclose your confidential information in a dispute, same as an AI agent is not meant to disclose your information

because it's considered as an employee in a way, and so to do that given that it's hard to classify data as you said, because you know sensitive, confidential, I mean DLP projects, data loss prevention projects projects are so expensive and so complicated to implement and to maintain.

So the current solutions that are in the market in a way is to actually let the data go, do the agent the way you want, but then have another AI that is policy compliant that will check the reply every time and reverse from the result and do not let any reply to go out if it doesn't fit the compliance. And actually there are solutions in the market that are doing that.

Right-sizing controls for company scale

very interesting yes I've met a few startups that are working on on that specific issue and I think it's going to be a huge mark especially for large organizations because I mean for an organization of our size 15 people we sort of know what goes where and we are also

So actually, the ROI would be nonsense. You would spend 50 % of your budget in stuff that are it.

I have a question there.

Resetting memory and retention controls

Do you have any way of asking the AI to swipe off all past tangents and data they have collected?

Yes. In this case, I can basically just erase everything.

Open-source architecture and deletability

I mean, what this is, they are text files, ultimately. So, we have basically a collection of text files, system prompts, and I can describe this because the system is going to be open source.

There will be a premium version for organizations, but I can talk freely about it. It's a series of text files that have quite specific information, and they are orchestrated in a way as precisely to allow this sort of, okay, delete everything you know about me.

Wrap-up and next steps

and that common we mean that the memory side the memory we just go I would say we had a long discussion so I would say we can wrap up and then we go to that to

the frog like last time what do you think the Maslow pyramid can we eat with the agents