Ilan (00:00)
One day we're going to be like completely strapped for content for a week and we're just, I'm just going to go and I'm going to grab, you know, random bloopers from every episode.

David (00:08)
Nothing but bloopers. All

the bloopers from seasons 1 and 2.

Ilan (00:13)
That's right.

David (00:13)
Hey everybody, welcome to prompt and circumstance. My name is David. And today we're gonna learn about creating an AI co-pilot.

Ilan (00:17)
and I'm Ilan.

David (00:35)
All right, Ilan, so you put together an AI copilot in a brief period of time in less than a day. So ⁓ let's learn more about that. What was that all about?

Ilan (00:42)
That's right.

Yeah, absolutely. So I do want to reference my sources here. Tal Raviv, who some of you might know, he's a popular product management content creator on LinkedIn He had a whole post in Lenny's newsletter about creating an AI copilot.

And I tweaked that to be focused on new product. But I hope that you'll be able to see here how

this could very easily be nudged in one direction or the other. And so how you could use it for your existing role, or if you're trying to create a new product from scratch,

David (01:22)
All right. Looking forward to learning more about this.

Ilan (01:25)
All right. Let me give a little background. So you might've heard in a previous episode.

One of the product ideas that we are working on is the next great AI tool for public school teachers. It's called Teachr. And as we've done initial discovery, what I found is I was often going to Claude to check my ideas. then I found this post from Tal, and it really got me thinking, how do I make sure that I have a structured environment?

with all of the context needed so that I can ask questions and it will give me advice on where to go. So let me show you how I did that in just one evening.

David (02:03)
And this is one of the evenings where it didn't ruin your date night.

Ilan (02:08)
not ruin my date night. were working next to each other on the couch. So ⁓ this was an evening that was planned for work.

David (02:17)
Okay, it's a working session.

Ilan (02:19)
That's right.

Ilan (02:20)
Now is the perfect time for me to tell you about our sponsor for this week, Querio. Have you ever found yourself blocked from answering a product question because you had to wait for a data team? Querio's AI powered data analyst sits on top of your data stack and allows you to ask natural language questions to get answers immediately. Try today by going to querio.ai and let them know that David and Ilan sent you for two months free.

That's a thousand dollar value for our audience.

Ilan (02:49)
Let me bring you into my Claude environment. And in Claude, I already had a project called Teachr V2. So this is a place where I had already uploaded the code for our application

David (03:05)
And this is the paid version of Claude.

Ilan (03:07)
Yeah, I should say I did pay for the pro version of Claude. I paid annually. I believe it's around $250 a

So what I did is within this project, the most important thing to have is instructions. Your instructions, which live here, these are called slightly different things if you're using Gemini or ChatGPT, but basically you can give the project background information on what it's supposed to do for you. And so exactly, the instructions are your system prompt.

David (03:36)
It's like a system prompt.

Ilan (03:41)
The instructions here basically tell it you're the co-founder of a company ⁓ and this system is an expert business and product coach, advisor, mentor. you can go through and basically tell it what you need it to do. So you're going to provide it context and detailed information.

and you need it you to think more deeply and see around corners. And then you also give it a persona, right? I want it to be encouraging, but also note when I'm getting stuck or I'm in a loop that I need to break out of. I also want it to make sure that I'm not being sloppy with the work that I do.

David (04:25)
You know, I think about this and I think about ⁓ human interactions. Wouldn't that be interesting if this is how we get introduced to one another? It's like, hey, nice, nice to meet you, David. Listen, I need you to be encouraging and I need you to do this and that. I expect this of you.

Ilan (04:43)
imagine if I'd been able to tell you, okay, David, I need you to be encouraging and ⁓ tell me when I'm getting stuck in a loop and be nice and a little bit funny, ⁓ but also sarcastic.

David (04:55)
Yeah.

You know, it's,

it's a funny scenario, but I, I, I feel like it could actually help with a lot of relationships, a lot of, and by relationships, I don't mean necessarily personal ones, although go ahead and try that one out. But, you know, professional ones, right? ⁓ I think the closest thing that I've seen to this is, is a read me like a manager read me. if any of you have heard of that, it's basically, ⁓ it's like an onboarding document that a manager would provide to, their team.

Whenever like say if the manager was new to the company or to the team and to say, look, here's how I like to operate. here's my strengths and weaknesses and here's, ⁓ you know, my expectations on, on how, you know, we can work together, right? Right. Like it's, okay to be direct to me, right? You don't got to sugar coat anything or maybe, Hey, look, ⁓ I need some context first before you give me the bottom line, stuff like that.

Ilan (05:48)
Yeah, I think that that's good foreshadowing for where I think something like this copilot could go. So once I uploaded the instructions into my copilot, the next thing I did was upload just the basic documents.

that I already had. So I had some interview transcripts with, with discovery interviews. I also had some notes from older interviews that weren't

So the next thing I did was I actually had it lay out a product strategy. I put in just my unfiltered thoughts on what it is we're building, who we're building it for, why we're building it for them, why I think that this market makes sense.

et cetera. So I have probably words or a thousand words here of just my unvarnished thoughts it helped me to create a

structured product strategy document. And I tweaked it a little bit.

David (06:48)
Hmm.

Ilan (06:52)
then simply uploaded it back into the project. So this is an important feature of these copilots is you gotta have a feedback loop, right? It helps you create something. Then you add that back to the context of the copilot. And now it understands that much more about you and you're able to sort of move forward iteratively.

David (06:58)
Mm-hmm.

Yeah, yeah, the recursion element of it. Yeah. It can be very powerful like that. where, you, you, take one step of like this, ⁓ huge web of, of steps. I'm not, I'm deliberately saying that it's not linear, right? ⁓ and, you, you iterate through that to, ⁓ all right. Little mermaid. All right. So, so.

Ilan (07:22)
Exactly.

Mm-hmm.

tentacles.

David (07:47)
right. So, so there's this whole web of, of things that you got to do. And then you iterate through one of those steps and then you, you, you land on like, okay, this step is good. Like we've, have some artifact and then now have that be, ⁓ a brick in, in the building that you're.

Ilan (07:50)
hahahaha

David (08:08)
that you're creating, right? Let that be part of the foundation of the rest of your work, yeah.

Ilan (08:14)
100 %

David (08:16)
And you can even do this with the system prompt, right? mean, the system prompt that you had as you showed us, it was fairly lengthy. And that could be something that you iterate on as your very first step to say, hey, look, ⁓ act as a prompt engineer ⁓ and also have these other traits ⁓ and help me create a good system prompt. And then that's your very, very first step.

Ilan (08:42)
That's right. And I think that actually gets to an important aspect of these tools, which is that when you're first setting your co-pilot up, you actually don't need to be perfect. I think perfection is the enemy of the good here. Create a system prompt. Think of something that you would like a coach or mentor or advisor to do.

And as you said, David, you can use the tool to itself generate a better system prompt. Or over time, if you find that there are aspects that are annoying or frustrating or unnecessary about how the copilot interacts with you, you can just tweak them directly there in the system prompt.

David (09:25)
Right. Which you cannot do with a person necessarily.

Ilan (09:30)
So, okay. So, so far we had the system prompt, the instructions on the project. We loaded whatever context we had created a product strategy document that it will work off of. The next thing that I wanted it to do was.

build a competitive research document.

So I use deep research, which we've talked about many times, to go ahead and look into a couple of key competitors as well as find any others out there in the market have some momentum who we might want to consider. And while I was off doing that,

I asked it to think about what knowledge is it missing in the context that I've provided so far and just have a conversation with me, ask me questions about important details that it should know in order to be the best possible copilot. So I spent maybe 10 minutes here where I was going back and forth.

Then at the very end, I had it create a document with the information that it gathered, only include new information, don't outline outstanding gaps, and optimize that document for being part of the project knowledge. So it just created an artifact for me that I then recursively added back into the project knowledge

that the prompts I used here, these came directly out Tal's newsletter post, guest newsletter post on Lenny's. So we'll link that in the show notes.

All right, so after that step was done, I basically had my copilot ready, right? It had some understanding of the strategy. had some context of users we've spoken to. It understood some competitive intelligence on who else is out there in the market.

and it had access to my code base. So I put it to the test. I hired my copilot for its first task. what I had it do was to help me create an initial requirements document for

David (11:25)
Mm-hmm.

Ilan (11:39)
our next initiative for this product, is to integrate with Google Drive.

And in here, I basically gave it just some high level notes about what I was thinking. I structured it. And then I asked it to add to my initial notes, maintaining the same structure.

David (11:57)
Now, as we go through all this, the way, I think one thing that's worth highlighting is the use of Markdown. ⁓ We have mentioned this before, but I think it's worth mentioning again. Markdown is a pretty good format for getting LLMs to understand a document. It's better than Word format, better than PDF. So for those who are looking to do this as well,

Ilan (11:57)
and

David (12:22)
We do recommend using Markdown.

Ilan (12:25)
That is very true. Thank you, David, for the highlight there. And in the tools that you use, whether you decide to use Claude or ChatGPT or Gemini, you should prompt them to use Markdown format for creating any artifacts that you're going to use. And even in Claude, an interesting

feature they have is up here on the top right when you copy a document that is created you can actually download it as a markdown file directly.

David (12:58)
Mm-hmm.

Yep. Yeah, I believe both GPT and Gemini have that functionality.

Ilan (13:03)
So what I got out of this was a reasonably competent and complete requirements document that I'm able to review and understand and make sure that I'm not missing some information. In fact, it gave me some critical questions to ask one of our key users who this feature is targeted towards.

David (13:28)
also gave you some, some pretty funny It says, Hey, your success metric is cute. But I mean, how good is it actually? There's that little sarcasm in, in the, in the prompt showing up.

Ilan (13:42)
That's right. And again, it did check me right? If I'm saying a success metric is ⁓ accuracy of the tool, but all the user talked about was time savings, then is accuracy really that important if the user just wants to save time?

David (13:56)
Mm-hmm.

Yeah.

Ilan (14:00)
And so I think this really shows the power of having a co-pilot like this, right? This was maybe an hour and a half of my time in one evening. And by the end of that hour and a I had ability to co-create one of these documents, which you can then continuously iterate with the tool.

David (14:24)
Yeah, this has been great. And I think there were some really important takeaways here. So first is that idea of recursion, the idea of working with the model to generate what ultimately would be context, additional context for the model to operate on. So starting with the perhaps system prompt and then,

context about the market and the thing that you're working on. For some others, ⁓ depending on what they are working on, it could also be the voice. If they're generating some marketing copy, it could be, look, here's examples of our voice, but not only are there examples, but here's the principles that guide our voice. So think that was really important. The other one is, well, use of Markdown. ⁓

So knowing that that is the ideal format, not saying that the others won't work, know, Markdown does tend to optimize for use in these scenarios. And, you know, being willing to ⁓ iterate throughout all of this, right? The expectation that it would one shot is not that realistic, but few shots certainly is.

Ilan (15:40)
Absolutely.

So one of the ideas that I've seen out there for these AI copilots is having it suggest automations that you should put in place to help you out. So I'm going to put in a pre-built prompt.

And I'm to have it suggest which automations that could be built in a tool like N8n, which we have access to, that would help save us time on boring or repetitive or Sisyphean tasks.

David (16:11)
wow, there's a word. Were you feeling like Sisyphus?

Ilan (16:15)
⁓ I just keep pushing that boulder up the hill.

David (16:19)
keep pushing that vibe code up the hill and the vibe code just rolls down with all these bugs.

Ilan (16:24)
That's right. for vibe coding.

⁓ So this is interesting.

talking about a customer interview intelligence engine. So when a customer call ends, transcribe the recording, find the key pain points, feature requests, pricing insights, and then create a structured summary and send it to us. This is Slack channel, assume that we have Slack, but ⁓ send it to us. That seems pretty useful.

And then competitive intelligence alerts. So when competitors, key competitors like Flint publish a blog post or pricing update, then analyze that content for implications for us and then alert us when, with a summary of the potential threats and opportunities.

know, I won't go through all of them here, but these just first couple are clearly useful. And, know, the next step would be to say, Hey, I want to do number one and number two. So give me step-by-step instructions in N8n on how to accomplish this. Use only the N8n documentation ⁓ and ⁓ user community to...

create these instructions.

David (17:40)
cool.

Ilan (17:41)
I think it's pretty cool. I think it's pretty powerful.

There's a lot that you can draw to being a product manager in a company, right? A lot of the same things that you would want as a founder of a new company or creating a startup, you'd also want as a product manager. You also want that competitive intelligence to be pushed to you. and the ability for a tool like this to kind of understand the context of your company and what would be helpful to you and be able to put in a general.

prompt or general thought and have it spit out really specific recommendations that are actually applicable to you, I think is hugely powerful for speeding up your work.

David (18:18)
Yeah. I, ⁓ I think that again, you know, regardless of, of whether, ⁓ you know, our listeners are in product management or any other, line of work, ⁓ this is something that could be very helpful, ⁓ as a general tool, for today, you know, in the same way that we don't use type writers that much anymore, or if at all we use word processing, it has just that much more power, right? That's just kind of the standard.

⁓ so, ⁓ in a similar way, having I think is going to be, a thing for, for people just to, to be at the right level of productivity.

Ilan (18:56)
for sure. And I think that you can see other pretty clear

advantages of having ⁓ a tool like this that's already built. Like for example, if you're onboarding somebody into your team, imagine being able to share a company knowledge project with them so they could get up to speed with, hey, what's going on at the company? Who does what? Who knows who? What are our customers saying? What are our main initiatives right now?

What are the questions that we have that we haven't been able to answer?

David (19:29)
Mm-hmm.

Ilan (19:29)
Or even as a manager, maybe being able understand and what your blind spots are. You know, what, what should you be thinking about in terms of coaching and mentoring and company strategy and pushing forward the right types of initiatives or how to discuss or who to discuss with. of that

with a tool like this.

David (19:57)
That's great. Well, thanks for walking us through all of that. It was quite straightforward, I would say. So I think any one of our audience members would be able to pick that up right now.

Ilan (20:09)
Have you built an AI copilot? Let us know. Leave us a comment. And in general, you can like and subscribe, leave a review, and follow us on all the socials at @pandcpodcast We'll see you next week.

David (20:23)
See you next time.