Introduction

My name is Melissanthi. I'm a senior product designer at MindStone.

And today I wanted to kind of walk you through how I use AI to speed up my workflows. I'm a solo designer in a startup and therefore I need to do research, design and everything on my own. And I've been trying to create a lot of playbooks and use agents in order to speed up all of these processes along the way.

Tools and Setup

So what we're going to do today is I'm going to use an app called Cursor. I don't know if any of you have heard of it, but essentially it helps me use a lot of agents at the same time without using a lot of context. It helps me also use an internal operating system that we have created that is called Rebel, which is the thing that Josh is going to be showcasing later.

The reason I'm going to be using these two tools together is a problem that we have been encountering as a team over the past few months, which is the following. We have been really encouraged to use different AIs and different tools like Plot, GPT, Gemini, try everything that is out there. And we have been.

1However, everything that we have been using, we have been using individually and we haven't had a serious source of truth or context for that. So every time we tend to use something, we have to input the same context, we have to give information about the platform that we're building and things like that.

Creating a Shared Context with Rebel

So, we have this internal operating system now that has all of this shared context, that is connected to different tools like our Slack, our emails, our Notion, our Linear and our tickets and it pulls information from there so we don't have to repeat and repeat throughout the different parts of the teams.

The UX Problem to Solve

And therefore, what I'm going to do today is take you through the process of someone bringing a problem to me as a UX designer, me trying to figure out how this problem has has been created by doing some quantitative and qualitative analysis with my agents.

So to begin, this is, I don't know if you can hear me if I'm facing that way because I can't see anything over here. So I may turn slightly.

Okay.

And then trying to also implement the solution even though I don't code on the front end.

Running the UX Auditor

So this is the interface of Cursor. So for the people that are not familiar,

I have just opened here in the middle a browser tab. This is a normal, it can be Chrome, Comet, anything that you are using.

On the left, I have here a folder structure, which is my internal operating system. And essentially what I've done in this folder structure

that we have different parts of the product, I have created a folder called product design and I've gone in and created my playbooks.

Playbooks: Structured Prompts for Repeatable Workflows

Now you may be wondering what is a playbook. 1It's essentially a very well -structured prompt. So it's a long prompt that has the right context, the right information, and the right rules of how to act.

And I've tried to think, what are the things that I tend to repeat in my workflows?

Research. I always do research. I sometimes do audits. I do the copy myself.

So I need to create these repeatable things that I can keep using throughout my work.

Case Study: Low Prompt Practice in MindStone

Now, we're going to start today with a problem that the Customer Success Team brought to me, which was that in our platform, MindStone which I don't know if everyone has context of what we do but we basically help people learn about AI

especially professionals. We have these courses and people have been learning and they have been progressing, they've been gaining points

and in order to unlock the assessment at the end of the course they have to have some specific requirements filled and one of these requirements is that they need to practice a specific amount of prompts.

prompts. So the customer success team realized that they're not practicing enough prompts. So my role is to go and figure out why this is happening.

Kicking Off the Audit

So the first thing that we can do together here is to run my UX auditor.

This is the long prompt that I've mentioned to you. You can see I have given it a persona. It's supposed to be an expert with lots of

experience. I've given it a goal on what it's supposed to complete. And then it has context.

So the context is the different background of the app, and all of these are different folders and files inside the structure over here. So as they keep getting updated by different colleagues, then we can keep this always up

to date, and we don't have to keep repeating and giving new context. I've given it required steps and questions that it might need to ask me before I begin, but also if one of my colleagues might need to use it.

So if a front -end developer wants to go and make a change and wants to do an audit themselves, they can also use this. And you can see the prompt has quite a few details.

Maybe it can be a bit shorter, but that's something we can keep refining as we're working. So I'm just going to close this file and I'm going to ask it to run over here.

So I'm going to say run at UX auditor. This is just a chat interface, normal, like chat GPT if you're using it, but it's basically an agent that can continue running things, it can create a to -do list and it can keep keep executing instead of just stopping after one step, which is quite good.

Agent Autonomy and Test Scope

Hopefully now it will ask me what it is that I do want to audit, because that is part of the prompt that I have created.

And once it starts auditing, it will be able to go on its own, basically, into the browser tab that you see on the left, and start clicking around and testing the app on its own. So I don't have to actually do anything. It will go and kind of act as a user.

I've also loaded my user personas in this folder structure so that it can know and try and act like this persona sometimes when I'm testing stuff.

So if I make this a little bit bigger, okay, we can see here that it's made a to -do list, it says confirm the scope is the first step, then it's going to go and capture interactions, check the forms, things like that, things that I've defined in my audit, for example,

and it's asking me about the scope, which pages do I want to test, things like that.

up.

So I have already pre -written a little prompt so that we can save a little bit of time. And I'm going to paste it here.

So I want to audit the AI competency program.

This is one of our main programs.

I want to take the program page where we have the requirements and I want to take the sandbox. This is the environment where we practice prompts.

And I want you to use already the account that I have there.

So, if we give it a couple of seconds, hopefully it will start working instead of just thinking, and it will do that.

In the meanwhile, because it might take a couple of minutes, depends on how fast it feels like being in the moment, we can run a different agent.

So, on the top here, oh, you can see, it's going to start moving on its own on the side.

Quantitative Research with Internal Data

So, another agent that I can run over here is something called UX UX research helper that I've made. So what does this helper do?

We have an internal admin platform that has the data of all of our users. So it has their progress, it has how many points they've gained, if they've earned their certificate, their badge, all of this kind of information.

And we have created an internal MCP that connects through cursor and goes and finds all of this data. So what I can do here is ask it to run my helper.

That should also ask me for a scope of which cohorts because we have different groups of people that we want to test.

If I go and I ask you to test my whole platform in one go, that's going to use a lot of context, it's going to eventually slow down a lot, potentially crash. So it's kind of good if you select smaller pieces of data to go through at a time perhaps. Perhaps.

So while this is loading, I'm going to copy the next prompt that hopefully will ask me about what it is that I want to investigate.

Complementary Analytics: Hotjar and Mixpanel

Another thing that it does in the way that I've structured my prompt is that in my research I also tend to use Hotjar to look at heat maps, recordings, and sometimes Mixpanel for different types of metrics.

So I've also instructed to help me think about different boards that I might need to create to look at data outside of here as well.

So you can see again, it's created a to do, and it's asking me now, oh, it didn't ask me. So that's something that we're going to have to change.

So no big deal, because we can go back to our original prompt and straight away paste this and say I want to investigate actually the users that have prompt level zero, time frame, are these last few community cohorts.

This is something that happens. If I test this five times at my house, in my work then the next time that I tested it might not ask me the question

so you need to keep refining the problems that you've made until they actually work more consistently. So I will just resend it and hopefully this time it will start doing that a little bit better.

In the meanwhile we can

check how the auditor is doing.

So you can see if I scroll through my chat here it's been taking screenshots through the pages that it's been going through and And over here, it's updating the to -do list.

It's still thinking. Because it's doing kind of a comprehensive analysis, it might take a little bit of time.

And at the end, it's supposed to produce a document with the whole audit findings and save it in a folder on the left -hand side with the rest of my product design stuff.

So every time someone runs an audit, we have a history of what has been done.

So it's always easier for the AI to know what was the problem, How was it solved and it's always good to keep a record basically of what we're doing

Now, let's go back to this one Okay, so I asked it to look at the last three community groups. The community groups are public information So I think it's okay if we try and find users from here

And you can see that is using our internal MCP and it's going and finding the leader boards from the last three groups to find Who had progress but didn't have enough prompts and things like that.

So

It's going to try and give me some statistics at the end of this.

Qualitative Research: Recruiting Users

Now, while these two are running in parallel, we can go to our third playbook. So we're going to gather, actually, our quantitative data.

What about qualitative data? Now we need to do some interviews.

We of course like to have some personas on file and user agents, but it's also nice to connect with our users.

So what's an easy way for me to go and do some interviews? use.

I have created another playbook that helps me, again, with the internal MCP that we have, go and find these users that we need.

If it was a real recruitment for users right now, I would be able to probably use the information from this one and the people that it finds here.

But because I don't want to publish any private emails, I'm going to use some of our MindStone internal emails.

Automating Recruitment and Email Drafts

emails, and what we're going to do is this playbook that I've made, which is called, if we open a new chat, so I have named this, recruit user interviews from past cohorts.

So basically what it does, I will tell it, find me people from the last three months that have this amount of points, this progress, and it will go and look at all of the information that we have on the users, create a list of users for me.

It's going to save it on file. So every time I do this, it saves it. And then every time I create a new list, it checks for duplication.

So if I've already been contacting people, it's not going to put them again on the list because I don't want to harass them every time I want to do a survey or something for an interview.

It's also going to go and check my Gmail because it's connected to my Gmail and see from the interview request that I've sent in the past which ones were successful and which ones weren't.

So it's gonna try and take some statistics around there in order to do better copy next time and it's also gonna run an A -B test on its own before it drafts an email for me.

So it's gonna do an A -B test, test it with my personas and then draft an email for me and keep it in my Gmail for me to send.

I'm not confident yet to send it straight away so I always like to double check before I send it that it's there and ready.

So if we ask it to run, again, it should ask me kind of what's the purpose? What do you want to achieve? Who do you want to contact? Hopefully, we'll do that this time.

Let's see. You can see that it's thinking what is this playbook, basically?

We can go back to the other ones and check how they're doing until this one decides to work.

Okay.

Okay, so we have our audit here, and it has some findings, so we can see assessment, it's locked until the weeks are complete, sandbox, recommendations. So we have a few different things, however, it hasn't created a document with a proper

analysis like I've asked it, so that's again part of it, so I could say, why did you not create the report? report. And hopefully that will prompt it to start doing a little bit of a better job.

Now if we go back here, we can see that a document has been made, which is the analysis of the platform and the users that we have on the platform.

I find it quite hard to read so far away, but we can see over here that it's gone and pulled from community cohorts and we have some numbers of people that don't have a prompt level.

Prompt level zero, but that they have progress, for example. So it's starting to give me some different statistics. It's starting to create a hypothesis

as to why this might be happening. And then it's suggesting to me, okay, go and do this Mixpanel boards. And these are some questions that we still have. So it's created a nice file for me to go and interact with now.

Let's go back to the interview one. Okay, here we go. So now, luckily it's asking me, when do you want to interview people from? What type of program?

program, what type of filters, do you want them to have specific points, or do you want them to have finished the program, so I can go and say I want them to be from this mastery July test cohort, which is just me and my colleagues, basically, and I don't want them to have a prompt level, but I want them to have done something. So hopefully we'll go and find the users that are in there without me having to go and check anything.

While this is working, and hopefully the audit is also working, okay, here we go, we can can see it apologizes and it's made the audit report as well so let me just accept it so we can read it a little bit better so we have key findings it's taking out over here let's see

issues and opportunities missing to add the requirement disconnected practice assessment guidance so it's gone through and it's actually figured out where the assessment is and when the the prompt is and everything like that. It's making some recommendations.

Findings and Reports

Now, the last playbook that I wanted to show you is this.

Since we can't obviously contact people for interviews right now, they're not gonna book for me in the next five minutes,

I've created some fake interviews on my calendar for Thursday.

Planning User Interviews

And the idea is that every time I have interviews, I need to create a research plan and I need to make sure that I have a specific goal,

what I'm gonna talk about, what are the questions that I need to ask and what it is that I wanna get out of it.

The Interview Helper Playbook

so I've created the last playbook for today which is called the interview helper which makes a plan for me so the interview helper I tell it to look at my calendar it connects to my Gmail and you can go and find wherever I have an interview it looks at the user their email and then it goes and pulls data from them so it will tell me what the role is what their progress was so I can go in my meeting prepared and then it's also creating a plan it will it will ask

me hopefully so let's test it so run UX research no UX interview helper let's see and while we're waiting we can check how is it going over here okay so you

can see it says it has found the mastery test you like awkward it has found eight users with progress above zero that have zero prompt level and you can see the names of the people here which are actually me and my colleagues making test accounts so it's gonna start creating a lot of information about them

now if we go back to our plan okay there we go so it's asking me is this a single interview or is it a multi participant study because we need to know if the questions should be directed towards everyone for example and is it feature feature testing, what type of study do you want to do.

So I'm going to paste again my prompt. It's a multi -participant study with a focus on understanding why people have progressed but have not practiced enough prompts to create the right prompt level. You can find the participants in my email and the rule books for Thursday. And we can call it prompt level requirements.

So we can see how it does now. Okay. So that is going to start thinking.

And over here, we can see all these files that you see here that have changed are, as I mentioned in the beginning, like it always saves a list of the people that I have contacted. So it's always going and updating and adding the new emails of the people that I have contacted.

And it has also created, hopefully, let's see, is this our new file? Okay. It hasn't actually made the Gmail draft yet.

So sometimes it asks me, sometimes it does it automatically. So it says approve to create the drafts, yes, create the drafts, please. And we'll see how that goes.

This one is also running. We can open one of them as well so we can have a look at the playbooks.

So if we open interview helper, for example, you can see again, same structure, experience UX researcher, these are the goals, this is the context. context, we have the specific process with the questions that you saw in the beginning,

checking the calendar, researching the participants, so it's always trying to very clear structure everything over there.

Now, let's give it one more minute, okay, so now it's creating the Gmail draft in my email, so hopefully if I do open my email, this is our platform, we'll see you very briefly later.

Hasn't happened yet. Still running. So I need a minute.

Let's see if the other one has any questions for me. No.

Working in Parallel with Multiple Agents

And you notice that this has allowed me to work in parallel on four different things.

So a type of research that could have taken me a week or weeks to plan we have been able to do in hopefully 15

15 minutes maybe it's a bit longer but let's see so created the Gmail there we go so suddenly I have an email in my draft which is called quick 10 minutes to improve your AI mastery it has my title it's asking the people like a few

questions it has a link to book some time in my calendar and even put like little image at the bottom that I had instructed to in the past so the only The only thing I can do is send it, and it's put everyone in BCC as well, which is something that I had instructed it to.

So in the future, if I end up improving this, I can just ask it to run this whole thing and send it on its own, and I don't have to do anything. And hopefully people book some calls with me.

Now for the last thing that hopefully is almost finished, here we go. So we can see it has created the study for me now.

So, if we scroll down, it has gone in my email, figured out all the people that I'm interviewing, it's created a little information about them, so Stefan, CTO and co -founder, this is how much progress he has, and so on, then moving down, it has some, like, best practices about

reminding the people that there are no wrong answers, it has questions around the prompt levels, certificate requirements, and so on.

So these are the four playbooks.

From Insight to Implementation

Now, I have gone through this in the past, figured out what the problem was, and I wanted to make some design changes to the program page.

Because the tech team has enabled me to start using cursor inside our repository, I was able to, again, within this interface, go and ask it to make changes.

Sometimes I use Figma to design and give it small, like, frames in there.

there, but I just wanted to show you very, very quickly that without knowing how to code

at all, this was, for example, our programme page before and this was our assessment and certificate and this is our new programme page which I've just changed by chatting with Cursor and these are our new assessment requirements, basically.

So I don't know how to code and, again, by using Cursor, I was able to do my research, think of the solution and implement it on my own.

Conclusion

And that is everything about my UX workflows.