Claude Code for Product Managers with Sachin Rekhi

All right, welcome folks. Welcome,

welcome. We're going to uh give folks a couple minutes to uh join us here and then we will get started. Great. We are

going to get started here. Um as we have a ton of fun stuff to cover in today's

session. Um so uh thanks for joining me

session. Um so uh thanks for joining me and thanks for being here on this topic of clawed code for product managers. So

let me um kind of talk about this for a second here. Um and u start by just

second here. Um and u start by just saying who am I? Uh I recognize some of your names uh in the list but uh for those who may not know me I am a startup

founder that I've started three different companies uh anywhere. am

connected and my most recent startup is called Note Joy. It's a note-taking app for both individuals and teams. I've also been a product executive at uh

companies like Microsoft and LinkedIn.

Uh most recently at LinkedIn where I got to incubate the LinkedIn Sales Navigator product and business um which has now grown to over a billion dollars in sales. These days I spend most of my

sales. These days I spend most of my time actually teaching about product management in particular um helping product managers master their craft in

this age of AI that we're all living in.

Now all the content that we're going to cover today about claude code for product managers is actually from my course AI productivity. Uh so we have an

upcoming cohort starting April 7th that not goes into not only goes into this topic of um clawed code but also talks about things like AI prototyping and how

we could be using these tools effectively to prototype the solutions and then test those solutions with customers. We also go a lot into

customers. We also go a lot into customer and data insights um as well as product strategy and execution. So, if

you've been feeling overwhelmed by all the things going on in AI and are curious about what are the most impactful tools and capabilities that you need to know as a product manager,

this course is for you. So, I encourage you to check it out. But let's jump right in and talk about what I want to cover today. In today's session, I want

cover today. In today's session, I want to start by making the case why Cloud Code, why I think every product manager should be downloading, installing, and

using cloud code as one of their main uh AI tools today. I then want to show you what's possible with these tools. In

particular, I want to show you what can you actually do. I'll be demoing a lot of stuff um to show you what's now possible with them. I then want to spend the rest of the sess session showing you

how you can get started with this and how you can start building AI workflows using cloud code immediately after this session. I'll do that by helping you get

session. I'll do that by helping you get started with installing the tools you need to do so. But more importantly, we'll spend most of our time really getting into the weeds of how do you build the AI workflows that I just

demoed. I'll then talk about where to go

demoed. I'll then talk about where to go from from here and finally try to leave about 10 minutes for your questions.

Okay, let's dive into making the case for why Cloud Code. Now, I'm sure all of you are using AI tools today, using them for everything that you guys are doing

on a regular basis. Um, whether it's ChatgPT Claw Gemini Microsoft Copilot, I'm sure you're using these tools on a regular basis. But here's the

thing. I've now realized that chatbot

thing. I've now realized that chatbot style tool where you ask it a question and you get a response is useful. But

what's now possible with agentic platforms like claude code is this idea that we can start building AI powered

systems. These are systems and workflows that really automate the entire flow.

And in fact we can also build them in such a way that they automatically get better over time. And that's really what's unique about Cloud Code. Now, I

want to explain how Claude Code accomplishes this with some of the key capabilities it has that are actually not available in some of these other uh

tools. First, it's focused exclusively

tools. First, it's focused exclusively on artifact generation. Now, where Cloud Code comes from is it was a tool built for developers. And so, what it's doing

for developers. And so, what it's doing is manipulating code. But it turns out that same manipulation is useful for documents, for product specs, for report

generation. And so it's really good at

generation. And so it's really good at creating an artifact, improving an artifact, and modifying an artifact. And

so that's really useful as we're thinking about using this for real product work. Now, the other thing

product work. Now, the other thing that's great about these tools is that they allow you to leverage rich local context. And what I mean by that is you

context. And what I mean by that is you can store all the data information you have in local markdown files which is just sort of a formatted text file and

it can read those really fast really easily. In fact faster than doing things

easily. In fact faster than doing things like calling a third party server or using MPC MCP servers and so you can give it a ton of context and it can consume that context for your product

workflows really quickly. Now the other key thing is it automates workflows with key capabilities. It's designed to go

key capabilities. It's designed to go work auton autonomously. Whether you're

building agents or skills or commands, all of these things are key capabilities that ensure it can automate those workflows automatically for you. Now,

not only does it has these automation capabilities, but it lets you run a ton of different tools. command line tools give it access to pretty much any capability that you want it to to allow

you to do things on your behalf. It can

even write code on your behalf to go accomplish tasks and I'll show you how it does that. And finally, the beauty of it because everything is stored locally in markdown files um it avoids vendor

lock in. Now, there's lots of vendors

lock in. Now, there's lots of vendors building tools that are really agentic in nature. Take Notion as an example.

in nature. Take Notion as an example.

Very powerful. But one of the things I love about Cloud Code is I have no vendor lock in here because it's really focused on local markdown files that I can take with me to the next tool. Now,

I don't want to uh make it sound like Claude Code is the only game in town. It

turns out there's a bunch of emerging agentic platforms. Claude has another one called Claude Co-work which makes it a little bit more userfriendly to use.

Right now, Claude Code is more powerful.

So, that's part of the reason I prefer it over co-work. You've probably heard about Open Claw, which again very powerful, but its challenge is tons of security issues at this point. So, not

something I would recommend for actual um uh production use. And the reason I'm betting on cloud code is really this chart. When you look at anthropic and

chart. When you look at anthropic and its continued adoption in the organization and enterprises in particular, it is crushing it. And so

learning this tool, Claude Code, is not only going to give you the most powerful capabilities, but you're also very likely to use it in your current or future jobs because the adoption has

been so great. Now, I will say there might some new be some new game in town in in a couple of months. Um, but the skills you're going to learn are equally

applicable to all these other tools um that I mentioned.

All right, so that's enough about me talking about why Cloud Code is so great. I want to show you what I

great. I want to show you what I actually now use Cloud Code for on a daily basis for doing my product work.

Now, I have built out about a dozen different PM workflows that I am actually using Claude Code for on a regular basis. Now, those of you who are

regular basis. Now, those of you who are familiar with my frameworks might know that I like to think about the product role as dividing the role into the real

tasks of vision, strategy, design, and execution. Now, I've already been able

execution. Now, I've already been able to build a bunch of capabilities in particular around strategy, design work,

and execution. So, if you look at a

and execution. So, if you look at a couple of these here, uh strategy, I now use cloud code to critique my product strategies. You know, I'll tell you, I

strategies. You know, I'll tell you, I have not found AI tools to be very good at generating a product strategy, but they are incredibly good at critiquing that product strategy. And so, that ends

up being a really useful capability um that I'm now automated in a full workflow. Now, I also like to keep tabs

workflow. Now, I also like to keep tabs on my competitors. And I'll show you how I use these tools to make sure we're still price competitive at my startup, NoChoy, by updating competitor pricing.

um using these tools automatically. I

can also do things like generate competitive tearowns um and automate that. Now I use it a lot in the design

that. Now I use it a lot in the design process as well. Starting with my user research, generating interview scripts, summarizing customer interviews, conducting NPS analyses. It's really

powerful for its ability to synthesize information and distill it uh really easily. I also use it for all my data

easily. I also use it for all my data queries, answering ad hoc data questions as well as creating my weekly dashboards. And finally, I use it a lot

dashboards. And finally, I use it a lot for automating execution things like drafting meeting agendas and generating release notes. So that's kind of a lay

release notes. So that's kind of a lay of the land of all the capabilities that I'm now using cloud code for. I wanted

to focus on four of those and get into real life demos of how these actually work. So let's do it. Okay. All right.

work. So let's do it. Okay. All right.

So now I am going to open this up here. And so

what we're looking at here is my Claude code setup. I have claude code running

code setup. I have claude code running in the terminal over here on the right hand side. I'll explain in a little bit

hand side. I'll explain in a little bit how to go set this up for yourself. Um

but for right now here it is running. I

also have on the left hand side I'm using Visual Studio Code which is just a text editor. now normally used for

text editor. now normally used for developers uh but it helps you actually view the markdown files um or the skill files that you're going to be using. Now

you can see I've created a project called product hub. Uh this is just a folder on my computer where I have all of my actual um product related work.

Now inside there is a clawed folder called skills. And so I've built all my

called skills. And so I've built all my workflow outflows out as various skills that I can click on and engage with.

Now, um you can see all of the different skills I have here. You can also see that I have projects here where I keep all the information related to my projects, which are various markdown

files, which are just text files um around each of these projects. I even

have a resource directory of things I want Claude to know, including no choice strategy, value propositions, pricing, and things like that. Okay, so let's

start using this um as I'd actually use it. So, let's say I wanted to do a uh

it. So, let's say I wanted to do a uh critique of product strategy. Now, I've

been working on this um product strategy idea for a product I might build. Um

let me open that up. So, I've been thinking, you know, hey, I'm a personal finance nut. Wouldn't it be cool if I

finance nut. Wouldn't it be cool if I can build this personal finance product?

I call it the radical financial planner.

And I spent some time thinking about what it might do, what are the problem we're solving, what's the target audience for it, who's the value proposition, business model, all the

typical stuff you throw into a product strategy. Um, and uh so,

strategy. Um, and uh so, okay, great. Uh so here's what that

okay, great. Uh so here's what that product strategy looks like. Now I want to use cloud code to critique that product strategy. So I can start running

product strategy. So I can start running my skill critique product strategy. I

can then give it access to this file. I

can type at and start typing the name of the file and now it has um can autocomplete to radical financial planner. Okay, let me kick that off.

planner. Okay, let me kick that off.

All right. So now it's running and so I'm running the skill that I created critique product strategy and it's going and reading all the information in that school in that skill all the best

practices I've told it about my product strategy. Now it's going to start

strategy. Now it's going to start composing the critique. So it's gone off and running this autonomously based on the skill definition I've created. So

let me go ahead and show you that skill as we're waiting for it to come back here. So in this folder over here

here. So in this folder over here critique product strategy I have the actual skill in it there's the skill.md

file which actually defines the skill and you can say the you can see here it says the goal of this is to provide a product strategy critique play devil's advocate and point out flaws or

limitations in the provided strategy.

Don't be nice. Point out in detail why the product strategy may not work. I

then detailed the specific steps of the workflow I wanted to follow. Here I'm

saying verify that the product strategy addresses each of the following six strategic questions. So make sure it

strategic questions. So make sure it covers all the things that are important in strategy. And then I say leverage the

in strategy. And then I say leverage the following knowledge files to critique each of the dimensions of the product strategy. This is where I've really

strategy. This is where I've really given it the brains of how it should critique the product strategy. Now, what

I've done is taken all my course content on product strategy and downloaded it locally. And so, I have here um this is

locally. And so, I have here um this is from my course uh what great strategy looks like and it has all of the text of that including the dimensions of a great

strategy, some of the key attributes of a great strategy. I then have all the specific lessons about how to come up with a great value proposition, a great target audience, a great strategic

differentiation. And so what Claude code

differentiation. And so what Claude code is doing is reading all these best practices I've fed into it and then

using that to actually critique um uh the actual uh product strategy. And so

that's what it's doing here. And so this is one of the key things when you're actually developing these skills. You

want to show it what great looks like.

In this case, um what great looks like in my mind is everything I've already written on product strategy. And so you can feed it that knowledge, that documentation, and then it ends up using

that to actually go and do um this critique.

Okay, great. So this is taking a little bit longer than expected. Um, so let me actually um open up the previous

generation that I'd done here. Okay, so

let me uh open this up in preview. And

so let's take a look at the critique that I put together. So it first did a strategic completeness check. Shows me

that all the dimensions are covered. But

now it's starting to go in and critique each dimension. Okay, so actually it

each dimension. Okay, so actually it just uh wrote that file. So you saw that it updated. So here's the critique. It's

it updated. So here's the critique. It's

telling me that my target audience definition is too broad and lacks the rigor of a true bullseye narrowing. And

so it's leveraging all the knowledge where I talk about you need a bullseye audience. Um, and here it's saying, hey,

audience. Um, and here it's saying, hey, the audience that you provided was too broad. It wasn't specific at all. Now

broad. It wasn't specific at all. Now

it's complaining that my problem to solve is surface level and fails to apply the outcome motivation gap framework, which is the specific framework I speak to in the course. Now

you can see we can keep going here. But

what's fantastic about this is it is actually critiquing my product strategy based on a very specific set of best practices. Best practices that I

practices. Best practices that I developed myself or you could use someone else's if you're excited about um some of the other great thought leaders and strategy. Bring in their

context and then get critiques. And so

this is a great way to use um these tools um for product strategy. All

right. So, let me show you another skill here. Um, I'm going to type clear here.

here. Um, I'm going to type clear here.

And now I'm going to show you another skill which is um update competitor pricing.

All right. So, let me fire that off because that's going to take a little bit. But um what this is going to do is

bit. But um what this is going to do is start actually opening a browser and browsing on my behalf all of my

competitors pricing pages and then downloading all their information in their pricing matrices. So it's first starting with a baseline of my own product Nojoy. It's going to

product Nojoy. It's going to nochoy.com/pricing

nochoy.com/pricing looking at the various price plans that are available. looking at all the

are available. looking at all the features that are available there and it's starting to extract pricing data from it. And so what's great about this

from it. And so what's great about this is it is literally using my browser and navigating my browser on my behalf. I'm

not I'm not touching anything here and you can see it go. And so this is the great thing. You can start using it for

great thing. You can start using it for this kind of work. Now there are many different approaches I could have used to get this information but there's a particular reason I chose this one. In

the past, anytime I asked um Deep Research or ChatGpt or Claude, hey, can you get me the pricing for this product, it was woefully out of date. It would go

run a search, find some old pages with some old pricing information, and it was never accurate. And what I realized is

never accurate. And what I realized is the only way for it to be accurate is to tell it to use a web browser and to go to each of these pages and go compile the information. And so that's exactly

the information. And so that's exactly what it's doing here. And so you need to be thoughtful about what's the best way to get the data when you're actually

doing this kind of analysis. And in this case, even though this approach is kind of slow, it's the most accurate way for me to get the data that I need here. And

so it's going to keep going through each of these competitors, looking at the pricing page. Um, it even found the

pricing page. Um, it even found the pricing pages automatically for me. And

it's going ahead and executing this.

Okay. So, um, I'm not going to let that finish here. um because it's going to

finish here. um because it's going to take a little bit, but let me show you what it produced. So then if I go to uh competitor pricing, open this up.

Okay, so what you're seeing here is an executive summary of all the pricing. It

has the pricing for each of the products. Note, my own note-taking app,

products. Note, my own note-taking app, but then all my competitors. Evernote

has a free plan, a $14.99 plan. Oh,

still going here. Uh so let me close that out. Okay. and

that out. Okay. and

go back to what it generated. It's now

showing me every single uh plan and their pricing. And this is accurate.

their pricing. And this is accurate.

I've handverified this because it went and pulled it up from the pricing page.

It then even gives me commentary here.

So, um you know, it says No Choy offers a generous free plan making it one of the most accessible free options. Uh you

know, it'll tell me things like um Evernote's free tier is the most restrictive. So this is really use of

restrictive. So this is really use of competitive intelligence that's accurate and autogenerated. All right. So that

and autogenerated. All right. So that

was the second demo I wanted to show you. Let me uh hop over here to another

you. Let me uh hop over here to another demo on summarizing uh customer interviews. Okay. I'm going to clear

interviews. Okay. I'm going to clear context again. Summarize customer

context again. Summarize customer interviews and then give it my uh customer interview uh folder. Let me

fire that off. Okay. Let me tell you what this one does. So, what I have here is a folder called customer interviews.

In that interview, I have video recordings of different customer interviews that I've done. Just calling

them company A, company B, company C.

And these are just literally dumps from Zoom meetings of customer interviews that I've done. What this tool is going to do is take that video. It's then

going to transcribe that video to text using a command line tool called Whisper, which is a transcription tool offered by OpenAI for free. And then

it's going to take that transcription and then summarize each of those interviews. Now, I've guided it on how I

interviews. Now, I've guided it on how I want it to do those summaries. If I go here, I've actually even given it a template and said this is basically

um how I want you to conduct uh and put together that summary synthesis. I want

you to include some details about who we interviewed. I want you to summarize the

interviewed. I want you to summarize the key takeaways. I want you to actually

key takeaways. I want you to actually put together problems and pain points. I

want to understand their current workflow. And then I want you to give

workflow. And then I want you to give direct quotes uh that the customer said.

And so you can see here what's powerful about this is I'm not just saying summarize the interview. I'm giving it a very prescriptive template for it to go use to go produce this. And now it's

told me it's actually produced the summaries. Um so let's go take a look at

summaries. Um so let's go take a look at some of these summaries. Um so I'll open up company A. Uh the summary that we produced. And so it's telling me uh this

produced. And so it's telling me uh this was the interview we did with Carl. And

it has a bunch of the key takeaways from him. small team with dis disperate

him. small team with dis disperate proportionate feedback burden all the key problems and pain points we then uh get into feature requests and very

specific quotes from that customer interview now what it's also doing is I then asked it to go create a cross interview patterns document so take each

of these interview question interview summaries and find patterns across them I even gave it a template to do though do so. And so this is trying to

do so. And so this is trying to synthesize feedback from all 10 interviews into a single one. So again,

this is going to take a little while to run. Um, but I have the previous run

run. Um, but I have the previous run here. And so let me show you what that

here. And so let me show you what that looks like. And so it's giving me an

looks like. And so it's giving me an executive summary of all the companies and interesting insights from it.

telling me kind of a concise summary of all the different companies, their company sizes, and the patterns of pain points that recur based on prevalence.

So, this one feedback fragmentation occurred for 10 out of 10 interviews ease. Um, and then giving me quotes. Um,

ease. Um, and then giving me quotes. Um,

all right. So, yeah, it's telling me that it's done with the recent run. And

so, it's just going to update this file.

Okay, here's the most recent run here.

And um yeah, so this is just a great summary of information that it's giving me um with really kind of no effort on my part. All right, so I've kind of

my part. All right, so I've kind of shown you a bunch of different scenarios here of how you can actually use these

tools um to uh really uh supercharge your workflow. So let me go back to the

your workflow. So let me go back to the slides here.

Okay, so I hope I've inspired you on the kinds of things that you can now do with these tools. I now want to talk about

these tools. I now want to talk about how do you go do this? How do you go build each of these workflows that I just showed you? Let's talk about getting started. Now, probably the

getting started. Now, probably the comment I hear most frequently about cloud code is that it feels very technical. It feels like it's for

technical. It feels like it's for engineers. I have to use the terminal.

engineers. I have to use the terminal.

Um, I want to assure you that this is not nearly as complicated as you might think. And so let me show you the easy

think. And so let me show you the easy steps to get started here. So the way cloud code works is it works on projects. All a project is is a folder

projects. All a project is is a folder on your computer. So you can just go create a folder anywhere in your computer. I called mine product hub and

computer. I called mine product hub and then rightclick on it. Go to services and click on new terminal at folder. All

that's going to do is bring up this terminal window. Now, all a terminal is

terminal window. Now, all a terminal is is a textbased interface to your computer. And so, you can start typing

computer. And so, you can start typing commands in here. Now, if you go to Claude Code's installations instructions, it will tell you to type these three things. All you need to do is copy paste them. You don't have to

understand what they're doing. All it's

doing is a simple installation. So, you

do the curl command. You then type Claude and Claude is running. The first

time you do this, you have to log into Claude. So, you type /lo and boom,

Claude. So, you type /lo and boom, you're in Claude. Now, at that point, this is what you're seeing. You're in a terminal window with Claude code

running. Now, at this point, this really

running. Now, at this point, this really isn't all that different than using Claude in the web. You have a input box where you can type anything you want, just like you put in a chatbot. And

that's really how it works. And so, at this point, it looks very similar to the way you're using your existing tools.

So, that's how you get started. Now,

once you've set this up, there are a few other tool choices you're going to want to make. So, uh the first one is which

to make. So, uh the first one is which terminal you want to use. For the most part, the built-in Mac terminal or Windows PowerShell is good enough. Um

now, the cool kids are using tools like Ghosty. So, Ghosty happens to be the one

Ghosty. So, Ghosty happens to be the one I'm using. It has some nicities where

I'm using. It has some nicities where better color coding themes and whatnot.

Um but a nice to have, not a need to have. One thing you do need to decide is

have. One thing you do need to decide is what editor you actually want to use.

Now, as I showed you, when we're actually running our skills, we need to write the skills and view the documents that are produced by those skills. And

so, we need a text editor to do that.

The most common one people are using is Visual Studio Code. That's the one I showed you a minute ago, and it's great because it's completely free. Um,

another popular choice that people are using is Obsidian. Obsidian is a note-taking app that really specializes in viewing and manipulating markdown files, which is the default markdown uh

type of content that these tools make.

And so, Obsidian might be another choice that you might look at. Um, the third set of choices you might want to make is voice. One of the things you realize is

voice. One of the things you realize is that typing is kind of slow. It's

actually really fast to dictate what you want Cloud Code to do. And so, it's useful to install one of these voice plugins. I happen to use Whisper Flow,

plugins. I happen to use Whisper Flow, but there's lots of these out there. In

fact, Cloud Code itself just ramped to 5% its own voice feature. So, um you know, in a in a week or so, you're not even going to need a third party tool.

It's built in. Um but it's just a nice to have because then you don't actually have to type. You can simply talk to your um LLM.

So, as you saw, this was the setup that I have. I have Claude code running in

I have. I have Claude code running in the ghosty terminal um using Visual Studio as my text editor and I use Whisper Flow for voice. Now, this does

mean I'm managing two windows next to each other. If you want an all-in-one,

each other. If you want an all-in-one, it turns out there's a Visual Studio uh extension for Cloud Code that actually brings it directly into the window. I

use this sometimes as well. Um but these days I do like having the separation.

And let me just show you what it looks like if you happen to use Obsidian. And

so in this case, I'm using Obsidian to view the documents on the computer.

Again, all these tools are are text editors to give you a way to view what you're seeing um inside that. And so

that's sort of your setup choices. Um

key thing here is pretty much everything I've shown you is free. The terminals

are free, the editors are free. Uh these

voice tools are uh paid, but usually premium. The only thing you'd have to

premium. The only thing you'd have to buy is Claude Code itself, which is a $20 a month subscription. Um, honestly,

even if your employer weren't going to subsidize it, I'd personally pay for it.

It's made me so productive as a PM that I consider it essential to my workflow regardless of whether my company is sponsoring that payment or not. Okay.

All right. So, you've gotten this set up on your computer. You have the tools installed. Now, let's get into the real

installed. Now, let's get into the real work of building your own AI workflows, just like the workflows I showed you a minute ago. So, the first question you

minute ago. So, the first question you really want to ask yourself is when should you build an AI workflow, right?

Not everything is worth automating and not everything can be automated. And so,

when I am thinking about what I'm going to automate, I start with these two live two highle questions. Is it worth building? and is it possible to build

building? and is it possible to build it? So, let me break that down. When I

it? So, let me break that down. When I

talk about is it worth building, I'm trying to see is there a real advantage to doing this. The first reason I might build an AI workflow is because there's

a particular advantage for having AI do it. The most common advantages is that

it. The most common advantages is that AI can do it faster. You know, as you saw when it was synthesizing customer interview feedback, if I had to read through that and produce summaries, I

would have done it a lot slower and so I'm getting a huge time advantage um in that case. Also, I find that it can

that case. Also, I find that it can sometimes do things more comprehensively than I can. I can get it to look at every single uh competitor website, every single kind of third party

competitor out there. um and it can do that comprehensively where I would tire just being a human. And so sometimes AI has an advantage and that's why I'm choosing to build a workflow out of it.

Now other times it's not about that AI is better at it than I am. It's just

that it takes a lot of time for me to do it and I find myself doing it often and I'd love to offload that to AI so that I can go work on a higher level work. Now

in this case I might just be trying to get it to do it as good as me or maybe even slightly worse than me. Um, but

that still ends up being a huge value ad to my workflow. So, those are the two questions I ask myself about the workflow when I'm deciding whether it's worth automating. Now, the second

worth automating. Now, the second equally important question is, is it possible to build? And there are key things now that AI is incredibly good

at, but there are some limiting factors that limit you on what you can really build with AI. The first thing you want to ask yourself is can AI uh acquire the

appropriate context. A workflow is

appropriate context. A workflow is really all about taking some data, manipulating it, and shaping it into some output. But in order to do this

some output. But in order to do this well, the AI needs to be able to get access to those tools, access to the data, access to the context you need for that particular workflow. Now, it turns

out this is more challenging than you might uh think. And so I'm going to talk a lot about context strategies here, but you need to figure out can you acquire the data so that AI can then automate

the workflow. Now the second key thing

the workflow. Now the second key thing is you need to ask yourself does this workflow have discrete steps? Can I

break it down to a list of 10 steps? If

you can great. If you can't because there's a lot of if then conditionals or uh you know judgment then it's going to be difficult to create an AI automation

out of it. Finally, is there limited human judgment required? Can we build this workflow in such a way that it's fairly um okay for an AI to accomplish

it or do I need human judgment? If it

does need a lot of human judgment, it's going to be difficult to automate that.

But I will um push you to really think about this because AI is incredibly good these days at even things that we would previously call judgment. Okay, so these

are the questions I ask myself for any workflow that I'm uh thinking about automating. Frankly, my process these

automating. Frankly, my process these days is hey, if I'm about to begin to ask uh that I would have historically done manually, I first go through this, is it worth building an AI automation?

Is it and can I build it before I even begin the manual task? Because this will save me so much time in the future. So,

let's now talk about how do you actually go about building it? And I want to break this down into a concrete set of steps that I have now um found to be very reliable after building dozens of

these. The first step is to detail the

these. The first step is to detail the steps of this process. You want to break down the task into its discrete steps.

Um you can tell cloud code to execute.

Now the question I always get with this is like well how many steps do I need?

What level of granularity? That kind of stuff. So it's easiest for me to just

stuff. So it's easiest for me to just show you what this looks like. So let me show you a skill that I produced called answer data curiosity. This is a

workflow that will take any natural language question I have and write a SQL query, execute that SQL query against the database and then come back and give

me the result and visualize the result.

So it's really powerful now because I never have to write SQL anymore. I can

just fire off to this skill any natural language data question I have and it answers it. So as I mentioned the first

answers it. So as I mentioned the first step to build something like this is to detail the steps. Let me show you the steps that I actually wrote to give you a sense at the level of detail you want.

So if you look under the workflow section you can see I say start by analyzing the database you have access to understanding the tables and columns and summarizing the database schema in a

concise format. This is just important

concise format. This is just important so the AI knows the tables in its database. Um, so it knows how to write

database. Um, so it knows how to write the queries. Then I say let the user

the queries. Then I say let the user know that you are ready to answer their data questions. Prompt them to ask any

data questions. Prompt them to ask any question they have about the data. Step

number three, for each question the user asks, construct the appropriate MySQL query to retrieve the relevant data.

Then execute the query against the database. Um, step four, create an HTML

database. Um, step four, create an HTML report that includes the original question, the MySQL query used and neatly format a table of results and

then visualize uh that graph and save it to this specific folder. Then open up the HTML report and show it to the user.

So you can see here I did detail the skill in very specific steps, but I've done so in basically human language, right? I'm not writing any code here.

right? I'm not writing any code here.

I'm just describing in explicit human language what I want it to go do and then the LLM is incredibly good at executing that. And so when I say

executing that. And so when I say detailing the steps, this is really what I mean here.

Okay, so the next step here is decide the context strategy which is determine the most reliable way to get the required data. So, this is kind of where

required data. So, this is kind of where I honestly spend most of my time when I'm building a workflow because this is the most challenging bit where for every workflow, we need to get data from

somewhere. Now, I figured out there's

somewhere. Now, I figured out there's really these five different ways to get data for that workflow and you need to be thoughtful about which one you want to use and which one can you use for the

particular workflow. So, let me explain

particular workflow. So, let me explain each of these. The first one is local files. Now if we had the data we need

files. Now if we had the data we need saved as a local file then great it can just reference that local file that is the fastest most reliable thing. So I

spend a lot of time uh making local files. So for example I even created a

files. So for example I even created a skill called save meetings which downloads my meetings from granola which is a great meeting app and then saves them as local markdown files in my

project hub d f folder. That's fantastic

because now cloud code has direct access to all my meetings when I want it to draft meeting agendas for me when I want it to you know produce specs from uh those files. So when you can get it in a

those files. So when you can get it in a local file usually a markdown file that's the best context strategy. Now

the other thing that's cool is that you can install command line tools to go do work on your behalf. And so these are tools that are running in the command line to go do anything. So, for example,

I was telling you about whisper, which is a tool that transcribes audio and video into text. And so, that's a tool that I installed and gave Cloud Code access to. And so, that's another great

access to. And so, that's another great reliable way to go access information.

Now, another great way is these MCP servers. It turns out all of these AI

servers. It turns out all of these AI agents now support this protocol called MCP, which is just a way for connecting them to thirdparty services. So let's

say I wanted to connect it to my Google Docs or my notion or my Slack. I can

install an MCP server and then at that point it has access to the information.

Now let's say that doesn't even exist.

What I can do is actually tell Claude Code to use a third party API write code on my behalf to access that API and access the data. Now this is really

powerful because you don't need to know how to code. You just have to tell Cloud Code to go do that and it will go do it for you. And so that ends up being

for you. And so that ends up being really powerful as well. The final thing is the browser agent. I showed you that update competitor pricing skill. That

was using the browser agent. Now, this

is the least reliable, slowest way to go, but sometimes it's the only way to go. So, it's good to have that in your

go. So, it's good to have that in your toolbox as well. So, let's just talk about some examples of these context strategies for things like summarizing

customer interviews. Um I was using the

customer interviews. Um I was using the local files plus using that command line tool um for uh that uh you know transcription for things like um

conducting an NPS analysis. I downloaded

the NPS results in a CSV file. So that's

a local file but then I actually had cloud code write Python scripts to analyze that NPS file. And so that's really powerful as well. U I even have it use gamma as an API to create a

presentation. And so this gives you a

presentation. And so this gives you a bit of an idea of how you can be thoughtful about what context strategy to use based on where the data lives and how you need to access it to produce

your workflow.

Okay. Now the next thing we have to do is determine the appropriate workflow primitives. It turns out AI now supports

primitives. It turns out AI now supports all these different ways of automating workflows. Skills, agents, commands,

workflows. Skills, agents, commands, hooks, and plugins. Now I'm going to simplify this for you guys because right now I've looked at all of these. The

best uh uh primitive is skills. Skills

is kind of a super set of agents and commands. You can actually use as an

commands. You can actually use as an agent to spawn and parallelize work and you can also invoke it via command as you saw me any um skill you can invoke

via slashcomand. So at this point right

via slashcomand. So at this point right now you can just use skills as pretty much um the catchall workflow primitive

uh for building your actual workflows.

Now once we've decided that we can start shaping the output. The idea here is you want to determine how to best influ influence the output to meet your quality bar. Now I found a couple of key

quality bar. Now I found a couple of key strategies for this to doing this. The

first is templates. So you can create a detailed template with exactly how you'd like claude code to generate the output.

I showed you a couple of examples of that and how I guided it to create my customer in interview synthesis in exactly the way that I wanted it to. Now

I also find it really useful to give it best practices. As I showed you for my

best practices. As I showed you for my product strategy critique, I gave it all of my course content on what a great product strategy looked like. And now

it's leveraging that as it's critiquing each of those product strategies. Now,

the other key technique is inspiration.

I'll often just give the workflow examples of what great output looks like and tell it go produce the output looking something like that. The reality

is I can be just that vague and it will do exactly what I want it to do. Um, so

that's really powerful here when we're shaping the output. Okay. So what we're going to start doing here now is creating a skill where we define the key workflow in a skill.md file and then we

shape the output using things like templates and best practices and sometimes even writing scripts. And so

this is ultimately what a skill is. It's

just a folder with some text in this markdown format with skills and templates and best practices. And that's

kind of it.

The final key thing here is you don't have to build this workflow all yourself. You should rely on cloud code

yourself. You should rely on cloud code to help you build it. And so when I showed you some of those workflow uh markdown files, you can see they're getting kind of complicated. Let me tell you, I didn't write a single one. Claude

code did. And so that's the key when we're building these workflows. We want

to build them incrementally with cloud code.

All right. So, let me go back and start showing you how we can start actually building uh our own AI workflows.

Okay. So, I'm going to open uh up everything here. Let me do clear again.

everything here. Let me do clear again.

Okay. So, let me just uh get this prompt here. Fire it off so I can get it going.

here. Fire it off so I can get it going.

Okay.

Let me explain this. So let's say we want to create a generate release notes skill and I want that skill to basically read the code that changed and then

write a release note for me. It's one of the tasks I typically have to do. It'd

be great if I can offload it to AI. So

let me actually tell you what I wrote for my prompt. I said let's create a skill called generate release notes that takes a GitHub commit URL as an

argument. It then uses the GitHub

argument. It then uses the GitHub command line tool to download the details of the commit inspecting the title, the description.

Okay.

Um and the change code files to create a userfacing release note. This release

note should have a title case title and then one to five paragraph description of the change from the user perspective, the value of the change to the user and any pertinent details. The release notes

should then be saved to this directory.

That's all I told it to do. Now, let's

see what it produced. It created this uh generate release notes folder, wrote a skill.md file, and wrote out the entire

skill.md file, and wrote out the entire skill for me. So, when you create a skill, you have this uh stuff at the top called front matter, which is just the name, description. Um, it then describes

name, description. Um, it then describes what the scale does and then it's figured out, oh, you gave me a URL. I

need to parse that URL and get this um Shaw hash from the URL. It's figured out how to go do that. And then it has figured out this GitHub command line

tool, researched how to use it, realized that if it gave um a uh command line uh prompt like this, it'll get the response that it needs and then it gets the

information it needs and then it starts uh generating the release note describing sort of the format that I asked for. Now, this is kind of

asked for. Now, this is kind of incredible, right? I just gave it a

incredible, right? I just gave it a pretty natural language prompt. I could

have even dictated that if I wanted to and it then produced a very structured skill. So let's go run this skill. So

skill. So let's go run this skill. So

it's done here. I can say now um generate release notes and then let me find that um example commit URL.

Going to type that in here. And so

that's just one of my recent commits for Nojoy. And it's firing off. it's going

Nojoy. And it's firing off. it's going

to actually go run the commands and um actually produce that um release note.

So, it's going to take a second here.

But yeah, so this is the key thing I encourage you to do. Sometimes you're

probably going to download other people's skill.mmd files and see they

people's skill.mmd files and see they look very complicated, but they were probably written the exact same way I showed you having Claude code do the hard work of create a detailed scale.

All right, so it said it's ready. Um,

uh, it put it in this release notes. And

so let's take a look at this open preview.

And here's the amazing thing. Let me

just show you over here. This is the commit that it actually wrote this um release note for. You know, I'm kind of lazy with my commit messages. All I

wrote was web clipper added article summaries. I didn't give any more detail

summaries. I didn't give any more detail and it has access to the code and it came up with an actually pretty great uh release note. It's saying the noy web

release note. It's saying the noy web clipper AI summary feature now works with articles and LinkedIn posts, not just YouTube videos. when you clip an article using the AI summary option, the clipper will automatically generate a

structured summary. So, this is

structured summary. So, this is fantastic that it's able to do this um already on my behalf. And so, this kind of gives you a sense of how I might

actually kind of build um this. Now, we

have a skill that's working, but let me just describe how I'd make this skill better. Now, um I'm not going to run

better. Now, um I'm not going to run this because it's going to take a while, but let me just show you the prompt. I

might say something like let's add a subfolder to the generate release notes skill called examples. Then download the last 20 release notes as markdown files from this d uh URL into this directory.

Name them this. And then it's going to download all my existing examples. And

then finally I'd tell it something like this. Um, I'd say, uh,

this. Um, I'd say, uh, let's update the generate release notes skill to reference the example folder when generating the release note to follow the similar writing style. And so

that's an example of now I'm incrementally improving the skill by saying, actually, this writing's okay, but it's not really the voice and tone that we typically use. Let me just give

you 20 examples of our last set of files of release notes. Download them and use them as reference. the next time you come up with one, make sure it looks like those. And now it will just go do

like those. And now it will just go do that. Again, all I did was type this and

that. Again, all I did was type this and it would go do all the work to actually enhance that skill. And so this is the key when we're talking about how do we

actually make these skills even smarter.

Okay. All right. So I hope I've given you a sense now not only of how do you install cloud code get it going but how you can start building skills for your own use case and how easy it is to

collaborate with cloud code and where you don't have to do any of the technical work you have to answer those highle questions that I talked about what are the steps what's your context strategy how are you going to shape the

output and then work with cloud code to do the rest of the work for you and that's how I'm actually automating my workflows so let me talk about now that you've gotten this engine of building these workflow is going where to go from

here. So, as I mentioned, the cool thing

here. So, as I mentioned, the cool thing about cloud codes is it gets smarter over time and there's a couple of key capabilities that enable that. The first

is more reference context. As you use cloud code, you produce more markdown files that becomes more context that it can then use for the next workflow or the next question you ask it. So, you

want to be have a bent towards always taking your output, saving it as a local markdown file because that's how cloud codes can get smarter. It also just recently launched an automemory feature

which is very similar to the features that you have in chat GBT but you can actually say hey remember this for the future and it will automatically add that to its own memory. As you saw I use

templates extensively in my workflows and so as I get smarter about how I work I can just improve those templates.

Maybe I want to have it collate additional insights beyond what I've typically done in the past. One of the things I love about templates is let's say I had a question from my user

interviews that was unanswered. The cool

thing is I can update a template and then have it rerun the summary across all the previous interviews I've done to see if I can extract that insight. Now,

that is something we would only ever do with AI because we'd never go back and manually look for these insights in those customer interviews because it's too expensive. But now AI gives us that

too expensive. But now AI gives us that ability. Now, of course, you can also

ability. Now, of course, you can also keep improving your skills as well, that actual workflow steps to make it tighter and more useful.

All right, so today has been a whirlwind tour of getting started with building these workflows, but I've only scratched the surface of the capabilities of Cloud

Code. I'm showing you here a bunch of

Code. I'm showing you here a bunch of additional amazing capabilities that it offers. Things like remote control. I

offers. Things like remote control. I

showed you how to do this from the terminal, but it turns out you can actually set this up so you're actually prompting directly from your mobile device as well. Um, you can have it

actually schedule tasks for you. Um, you

can even do things like plugins so you can share those skills with other people in the organization. And so these are advanced techniques that I encourage you to check out once you get the basics

going. And as I said, I have a course

going. And as I said, I have a course that goes into all of this, helping you as a product manager to go even deeper on cloud code and more importantly

helping you to master all the key workflows that exist across u the work that you're doing as a PM. And so I really tried to design this course to

distill the key capabilities that every AI powered PM needs, whether it's using cloud code, prototyping, doing customer synthesis, coming up and critiquing product strategy, saving you time on

execution. Um, so I'd encourage you guys

execution. Um, so I'd encourage you guys to check this out um at the Reforge website.

All right. So, uh, as promised, I've actually left exactly 10 minutes here at the end for any questions that you have.

And so, I'm going to invite Arch up um to uh help us uh with those questions.

>> Yes, I'm very impressed with your time management skills. That was spot on.

management skills. That was spot on.

Bunch of great questions come in and I'm sort of grouping them together by theme.

So, first of all, lot of questions about how you think about using claude code versus claude versus co-work. A lot of folks

perceive overlapping use cases. So,

curious on your take about when you'd pick each one, what you think the strengths are.

>> Yeah, that's a great point. So, um I didn't have time for this, but I actually have a limitation slide of cloud code. Um, but basically I don't

cloud code. Um, but basically I don't use cloud code for everything. Anytime I

want to automate a workflow or have it run autonomously, as in I want to fire and forget, that's my default for cloud code. Now, when I have a simple

code. Now, when I have a simple question, I'm still using chat GPT and Claude. Sometimes, actually, Claude Code

Claude. Sometimes, actually, Claude Code has such a desire to like update the skill. I sometimes just go to Claude or

skill. I sometimes just go to Claude or ChatgPT and be like, "Hey, I have this random question that I'm going to go ask it so it doesn't influence it." So

that's the primary way I use the two tools. Now, one additional caveat is

tools. Now, one additional caveat is actually Claude Code sucks at deep research, which is kind of a shock, but Claude is great at deep research. Chat

GPT is great at deep research. So deep

research is one of these things I still do in these other tools. Um, but that's how I think about it. Now, co-work

may eventually be a userfriendly replacement for claude code. I've just

found right now I keep running into limitations that claude code gives me the full power and I don't think it's that hard to learn that I preferred cloud code over co-work.

>> Awesome. Um, next group of questions is around sharing and collaboration. So, a

lot of what you showed us is stored locally on your computer. How would you think about uh sharing this either what you build or the underlying skills with the rest of your team so that you can

all work better together?

>> Great question. So, there's let's talk about two parts of this. The skills and then the context. So for the skills, the key capability that they launched is

something called plugins, which is a way to package up skills, commands, agents into a sharable plug-in that anyone on your team can download. And I encourage

you to check that out and do that. What

it does is it takes basically that skill folder and uploads it to GitHub on your behalf. And now you can keep updating

behalf. And now you can keep updating the skills there. and then other people can type a single command to download all of those skills. And so I'd encourage you to have one for your team

and then you guys to share best practices and update them. Um, so that's a great way to make those skills shared.

Now the other big question is how do you make the context shared? This one is a little bit more difficult because you're going to start clobbering each other if you try to move all the context into the

cloud and it's very difficult to um do that. So often what I'm doing is I have

that. So often what I'm doing is I have local context and then I push that context to all of our collaboration tools, whether it's notion, whether it's Google Docs. Unfortunately, these tools

Google Docs. Unfortunately, these tools aren't great yet at directly editing those, but it can push content to them.

So I'll manipulate a local file and then I'll push it using an MP MCCP server to Google Docs or Notion for sharing.

>> Amazing. That's a huge unlock. Uh, next

question related in my mind is about security and you know it seems like Claude has access to a lot of the context important for it to do some of

this cool work. But with that, how are you thinking about securing your data?

What kind of access are you giving it to the files on your local machine and how do you think about that?

>> Yeah, so great question. So there's lots to talk about here. I'll just give you the kind of TLDDR. Um the key thing is you can scope uh claude to only look at

content within your project folder and at that point it's looking at anything that's in the project folder and so I'm just careful about what I put in the project folder. So that's one key

project folder. So that's one key capabilities here. Now the second key

capabilities here. Now the second key thing is um actually by default claude is constantly asking me for permission to do anything at all. Can I run this

command? Can I go access this MCP

command? Can I go access this MCP server? can I write this file? And so it

server? can I write this file? And so it is very explicit about what it's doing, frankly, to the point of being really annoying. And so they have a bunch of

annoying. And so they have a bunch of ways to like remove those things. But I

actually like that it's safe by default.

Now compare that to something like OpenClaw, which removes all the friction by removing all those safety checks, but oh by the way, it can go do anything on your computer and it's going to do it

without prompting. That's very scary.

without prompting. That's very scary.

So, I actually think some of the trade-offs that Enthropic has made um on trying to be safety first is key. Now,

also when I give it access to thirdparty data, call it an MCP server to access my Google Docs and whatnot, I'm careful to by default only give it read access and

only give certain skills write access where I really need it. And that really helps safeguard my data as well.

>> That's a really great point about friction sometimes being our friend.

Next question, we're going rapidfire on all of these is a lot of people pointed out, oh my gosh, the token usage must be astronomical. So with that, do you have

astronomical. So with that, do you have any tips for how to optimize around token usage limits and how you think about costs relative to the output that

you're getting? This is a great question

you're getting? This is a great question and you know I didn't go into this because it's more of an advanced topic but I think about this all the time and so it turns out as I think about context I think about what's the cheapest way

from a token perspective to get the appropriate context as well. So local

files are actually cheap as long as it doesn't have to search across all the files. You can scope it to a given

files. You can scope it to a given directory. So I'll tell it go only look

directory. So I'll tell it go only look at customer interviews, go only look at product strategy and then that will scope the local file context. It's both

fast and token efficient. Now,

unfortunately, the big thing that everyone was talking about about six months ago was MCP servers, which is a great way to connect these tools to um cloud code.

Unfortunately, MCP servers are very token expensive, extremely token expensive. And I'm now at the point

expensive. And I'm now at the point wherever I can convert an MCP server to a command line tool, way more token efficient. And so it's this weird

efficient. And so it's this weird subtlety where you're like, "Oh, I can get the data between an NCB server and a command line tool. MCP server is a little bit easier to install because

they make that easy to do, but actually it's going to cost you more tokens." And

so I'm like moving workflows to command line tools in order to do that. And so

um I do find there's cases where I need to be thoughtful about this stuff. Now,

the simple solve for me is that Claude Pro for 20 bucks a month was pretty useful, but I'm clearly on Claude Max.

100 bucks a month, so I don't even have to worry about these things and I could just go fire at will. Um, but yeah, if you want to be a little bit more costconscious, um, then you are going to have to worry about that. And those are

some of the techniques you can use.

>> Amazing. And then I think a great question to end on is all this is really cool. I'm excited to try it out. It's

cool. I'm excited to try it out. It's

the shiny thing. And can you talk a little bit about how you actually think about the quality of the output that you're getting? What kind of guard rails

you're getting? What kind of guard rails do you have in place? What kind of checks? What frameworks would you

checks? What frameworks would you recommend that the folks in this call use to ensure that what they're getting out matches what they need?

>> It's a fantastic question. You know, I showed you about a dozen skills. Let me

tell you, I've tried to build about two, three dozen skills and I've tossed them because the output was crap. Um, and so this is a real challenge. I see on social media people posting all the

time, look, I automatically wrote a PRD from some raw text. Every time I've t that tried that, I've just gotten terrible output. Um, and so it is this

terrible output. Um, and so it is this thing where I review the output and decide, is it good enough to put my name on it? Um, before I decide to use this

on it? Um, before I decide to use this regularly, now there's so many techniques I've now developed to improve the output. Um, things like we talked

the output. Um, things like we talked about, give it a template of exactly what you want it to do. Give it best practices. Like go have it effectively

practices. Like go have it effectively read the book on how to do that thing well. Summarize the book, put that into

well. Summarize the book, put that into the skill, and then use it. And it

actually works. It's surprisingly good at it. Finally giving it examples. Like

at it. Finally giving it examples. Like

I literally will give it 10 examples, 20 examples of what great output looks like. And cloud code is really good at

like. And cloud code is really good at following it. And so I am usually never

following it. And so I am usually never satisfied with the first version of the skill I develop, but then I've developed these things that I can do to refine the

output and then judge um if it's any good. You know, people talk about like

good. You know, people talk about like judgment and uh uh taste is really the thing that matters these days. This is

what they're talking about. Taking the

output and deciding if it's good enough and then iterating on the workflow until it is >> amazing. Thank you. And with that, I

>> amazing. Thank you. And with that, I know we're at times or we're all done with questions for now. Um but just want to extend a big thank you to Sachin for sharing all this. This is an awesome

session. and Sasha and I'll hand it back

session. and Sasha and I'll hand it back over to you for any closing words of advice you have.

>> Thanks folks. Um I've hope you found this valuable. Like my big challenge to

this valuable. Like my big challenge to you now is go install cloud code and build your first workflow. Um I think I hope I tried to make this as actionable

as possible for you to do so. And so I I really encourage you to make that your next step.

>> That's what I'm off to do now.

Thanks so much. Have a great day everyone.

Awesome. Thanks, folks.