Cursor for Knowledge Work

0:02 Awesome. Well, thanks for sharing. Um, I think we'll just go ahead and jump right in. My name is Emily Gavrilenko. I'm one of the growth engineers on our go-to-market team, and my job is to make sure that every company with software engineers knows about Cursor and is using us. And I've been using Cursor since February of 2024, so it's been a little over 2 years now, and I love the product.

0:24 George, on to you. Hey everyone. I'm George, I'm on the growth team uh here at Cursor, and uh really excited to share a little bit more about uh some of the stuff we've been building internally and how Cursor uses Cursor. Awesome. Well, with that, let me do some slides on a brief introduction on Cursor for those folks that are new in the room, and we can jump right into the demos.

0:50 Awesome. So, this is our Cursor for knowledge work workshop, where we'll be doing a high-level introduction to Cursor, going through a demo of some top use cases we recommend for you folks to try out, and then George will be talking about how Cursor uses Cursor for knowledge work, um and then finishing off with some Q&A. Um, and the Cursor ecosystem, for those that don't know, is a platform for building with AI agents. So, you have access to all the frontier AI models, uh the best model harness to actually take

1:17 your human language and turn it into something that the agent can work with and give you outputs. It started off as a platform for software engineers, but as the models and the tools have evolved, we've seen a lot of folks um across the different other areas of knowledge work start to use the tool um in their workflows. And where can you use it? You can actually use it in a lot of different surfaces. Our most common platform is the Cursor desktop, which is the IDE we all know and love, but we also have

1:43 Cursor agents, which you can run on the web, through Slack, and now we recently launched our API and back um SDK yesterday. Uh you can also set up always-on automations, which I'll be demoing a bit later for those repetitive tasks you have going on. And whether you are a CLI fan or are on JetBrains, you can also use Cursor directly from there as well.

2:04 And some key resources, if after this call you're super excited and want to get dive in deeper, we have our documentation, we have more workshops, and we have our Cursor Academy on cursor.learn with some more agent best practices. And if you don't know, you can go to cursor.com and download to get started right away.

2:22 And there's a few different ways working with Cursor that has evolved um into more and more agent-centric. There's always the code-centric view that we all know and love, where you can go ahead and directly start editing files. There's the agent-centric way, which we recently launched with Cursor 3, um that is really makes agent front and center, where you're prompting and building the tools that you want. We also have the multi-agent terminals,

2:45 where you can spin up a bunch of agents in parallel, um and prompt them all in parallel as well. And then we're focusing more and more on design elements and UI elements, so you can imagine you can directly interact with websites. We recently launched a canvas mode, where you can build a React components directly in the IDE, um and you can also actually go ahead and ask Cursor to play around with design systems, and integrate directly with Figma as well. And finally, we have long-running async

3:11 agents. So, you can imagine if you're doing research on different accounts, if you are thinking about maybe wanting to break into a new area or segment, um you can spin off agents to think for a long time and come back with those kind of research that has completed. And we also have automations. And so, some that we have running is reviewing our codes whenever there's an issue in production, it launches off an agent to investigate and solve and fix, um and

3:38 then you can also imagine for my daily workflows, I can have a recap of every meeting with an email summary written, um and so forth. So, without further ado, today I'll be showing three core workflows that I use every day. Um, one is data analysis. Um, how do you make sense of a ton and ton of data? The second one is meeting recap. Um, so I'll be showing you how to integrate with your Granola notes, and you can imagine extrapolating that to your Google Calendar as well. Um, and email drafting. So, that way most of us

4:05 lives on in our inboxes, and now how can you pull in all the context to give you actionable results? And so, without further ado, let's get started. Um, can you see my IDE, George? Yes, yes. Sweet. So, this is the classic Cursor interface, um that is file-centric and um code-centric. But today I'll actually be in the agents window, so you can access this in the top right with the toggle, or with file new agents window. Um, and this will be a new view that

4:36 makes uh we'll be in for most of the day. And so, I have a few kind of snippets prepared. I have a file with usage data. And here I have um pulled together a list of Cursor activity that I've exported from the Cursor dashboard and then anonymized for this demo. And as you can imagine, uh maybe I have a request I want to figure out who is the top use Cursor user on my team.

4:59 Um, and the agent will go ahead and start analyzing the data, um analyzing and synthesizing, and returning the results I'm looking for. While it's thinking, um I did want to touch on that right now. I'm using Composer 2 fast model. Um, this is a combination of speed and intelligence that I love using for demos since it's kind of it's pretty fast, but we have all the frontier models across all the labs, so you can imagine using GPT from OpenAI, Opus from Anthropic, and so forth.

5:27 And you can see the agent returned, and it found that Richard Hendricks was the top user, um with this amount of agent usage. I can also ask and follow up. Hm. I see some folks are getting the reference. Love Pied Piper. Becomes more and more real every day.

5:47 So then maybe the next follow-up question is I see this one top user, but now I want to know the top model used. And you can see it actually goes ahead and does a sum of the usage. Looks like Claude Sonnet 4.5 is the top top model used. Um, this is some dummy data, so maybe this is not actually what you'll see. Um, I know internally a lot of our team is using Composer and Opus.

6:07 Um, but let's say now it's great, I have some data. Now I want to visualize it. Um, and so Canvas is actually a new feature we launched a few days ago, actually last week, and it's able to use React to modify the IDE and display dashboards. So, this is a great way um if you're working with data and you want to better visualize it or show it to a team, this is great off maybe for more one-off tasks. If you were wanting to create a dashboard that's always-on, um

6:37 you can actually do that with Cursor to create a website, which I'll be showing next. Um, but right now it's going through, thinking, making changes, and created the canvas file. And so, you can see here it popped up the dashboard directly in the IDE, um and you can kind of see the user usage, the different models that are used, the daily team metrics.

6:58 It's a little ugly, I don't really love it, so let me follow up with agent and try to ask to make it better. Oop. And so, I have a bunch of screenshots prepared. And then let me use voice mode for this. It looks a little ugly. Can you please update to match more of the Cursor dashboard?

7:20 And now um Cursor >> me? Oh, I had Alexa pop off in the background. Um, and so right now um you can see Cursor can accept um image as input, it can accept CSV files, um and it can also look through your code base. So, right now it's actually going ahead and modifying changes based on my request. And while agent is thinking, I'm going to kick off another agent in parallel um to do a second long-running task.

7:47 Um, so right now as you saw, the agent was working mainly with CSV data. So, it's a one-time export, but maybe you have a huge database like we do at Cursor, and you want to look at data through there. Um, so I have a sample prompt. My name is Emily, um and I want to figure out when I first started using Cursor. I actually had a second job before I joined um when I was at Roboflow, so I want to figure out on that account as well. And so, I'm going to a kick off an

8:11 agent, and right now I provided very little information on the database schema, on the structure of our DB, and so it's probably going to think long and hard and use up a good amount of tokens. And so, once it actually learns the structure and figures out how to work, um then I'll actually ask it to create a skill um to do this. So, it looks like it's a little confused.

8:34 I'm going to tell it to use the Databricks MCP. Um, so MCP servers, if are a great way to actually connect Cursor to your internal knowledge bases. So, right here I have my personal ones connected to Figma and Granola. And then I have to my team ones connected to our Databrick um our database. We have for Statsig, that's able to read through all our experiments, um and you can imagine a bunch of our backend services as well.

8:57 And our Fjord, so I haven't used them in a while, but can confirm the Databricks one works. And now you can see agent is off and running, looking through a bunch of our database schemas trying to figure out where in the world the data I want lives. Um, so we'll go back and take a look at it later once it's done.

9:12 Let me go back to see this analysis. Um, so cool. It's been working hard, and it updated a lot of the graph to look more like our dashboard. Um, it used the color scheme that I liked along with the cloud and IDE information. And then if I wanted to keep iterating, let me actually ask this for a tooltip so I can look at the numbers. Exactly.

9:41 Awesome. And then once that's thinking, I can actually launch another agent in parallel to convert the entire canvas to a dashboard. Create a React Next.js app to visualize the canvas on Cursor usage data from the CSV file. And then, let me actually tag it so it doesn't go on a long grind looking through the entire codebase.

10:08 Emily, uh some folks are asking if you can zoom in on your on your cursor. Of course. Is that better? Maybe a little bit more in case folks are like screen sharing somewhere. Cool. Good call out. Any other questions in chat I can answer? I think that's good for now. Awesome. Cool. So, looks like my data usage has finished looking through it. And so, looks like my first account was created on February 19th. And then, my Nysphere account was created on March 10th. So,

10:47 this is awesome. Now, to make it faster, I'm going to create a skill on looking at usage data. Create a skill for how to fetch user data and creation date. Cool. So, it's going to go off to the races and start thinking about it. Um now I have this app up and running. And I have my other agent looking through the canvas. And let's see what this is doing.

11:25 Cool. Let me ask Agent. Um and while all of those agents are running, uh I love having a ton of them in parallel on all my different tasks. Cool. Now it works. Let's see. And the hover state seems to be working. Um so, you can imagine I have a ton of ideas of things you might want to work on. And you can just prompt in natural language to get those analysis and the results that you want. Um I showed a few lots of different things in the IDE. So, I

11:48 wanted to take a step back and dive a little bit in more detail. So, we looked at the model picker. Um but models typically have three different flavors. Um one is intelligence. And so, you can imagine I can use the 1 million context window, which is the maximum. I can have high reasoning effort. So, you can see you can change reasoning from high, medium, low. Um and then, you can also control speed. And so, speed is typically more expensive, but um uses

12:15 the same model, just different GPUs. And so, if I have a quick task where I'm asking about when did I sign up, that's a pretty low effort um low reasoning task, especially if I've provided access to the tables. And I typically want the answer fast. So, I'll use Composer 2, which is a really fast and um medium um intelligence model. And then, let's say I have an error in my database, production is down, or I want to fix an early gnarly bug, um I might throw like Opus 4.7 at it, which is a really smart, longer running model.

12:44 Um and so, I'd recommend playing around with the different ones. They all have different flavors. Um we've seen that Opus is a bit more creative. Um the GPT models are a bit more um problem solving. Um and choose your favorites from there. One other thing I did want to touch on is the different modes. So, I've been in agent mode for most of this, but plan mode is really great if you're actually just getting started with a task and you want to figure out the best way to

13:07 implement it. And so, let's say I was working on a PowerPoint generation tool. And I might have a plan and I'll talk to Agent, what is the best way to do this? And it'll actually look through the codebase, it will research docs, and create like a markdown file of a step-by-step implementation plan. And you can review it, figure out if that makes sense. Um we also have multitask mode. So, it you can imagine a meta agent that's babysitting above each of

13:30 agents in parallel to do the task. I actually tried it out for the first time yesterday. This is a recent launch and it created two PRs in parallel. Um and then, we have ask mode, which is just a read-only view on your codebase, which is great um if you're just getting started with analysis, with onboarding to code, with a task, and you want to make sure Agent doesn't have access to write anything. Um as I showed, you can drop in images.

13:52 You can um try different models. And then, you have different skills. So, I used the create skill skill um to create the skill that I used, but um I have a bunch of different ones. You can also view them in your settings under skills, rule skills, and sub-agents. Um I have a big rule that says do not leave AI swap comments since I saw it first that these models tend to be very verbose and want to add a lot of comments everywhere. Um I added some

14:16 third-party Figma skills when working with design files um and some sub-agents for maybe working with different linear issues, for investigating logs, and for drafting customer responses. Um and then, one I use a lot is to resolve merge conflicts. And so, when I'm ready to push to production, we have a lot of developers working on the codebase and there's a lot of conflicts.

14:36 And so, maybe I want to address all the PR comments. And so, that way I can have Agent work on what other folks are asking it to fix. Um so, I'll be going back and let's look at the skill that my Agent created for me. So, it has the Databricks user lookup skill. Let me try it on George. And then, I want to use skill.

15:02 Cool. And so, now it should run a ton faster. I guess now I'm using the GPT 5.5 model, so it will balance Composer speed with this model. Um and instead of actually looking through the entire database, it's going to use the skill and it'll know exactly where to look. Um and then, while this is running, let's go back and see if my dashboard has finished building.

15:28 So, you can see here it decided not to work. So, let me take a look and see what changes I made. Uh it wants me to run this. So, sometimes Agent is a little lazy and wants you to do a command. So, oftentimes I'll just prompt it to start the server. So, as you can see here, now it's going to run start the server. Um and I can see here now that it's going to make my website for me. Um this one decided to open in the browser. And you can see the same thing that you kind of created in the canvas.

16:02 I have a website up and running that I can then publish and share with my team that surfaces the model mix, the user activity, and so forth. You can actually also open it directly in the cursor IDE. So, I can type in the localhost 3000 and it will open it um directly in the browser mode. Um depending how many monitors I have up, I'll decide if I want to code with it directly in the editor or not. Um but cool. We created a pretty cool website that shows cursor usage.

16:29 And now you can see here, for George, this one only took 30 seconds and that he was signed up on May 13th. Took a while to start coding, but then by July, he decided to try out Cursor. So, awesome, George. Um and then, one other thing I wanted to touch on, the last one, is actually using MCP servers. Let me switch back to the Composer 2 model. Um and I want to go through my past 7 days of Granola notes and get a weekly recap of what I did in all my meetings. And then, I also want an

16:57 action item list of things that I promised to do in all these calls. So, I'll make sure I don't get forget and get tagged in pings. Um you can add Granola directly here, but I'll go through the MCP servers. I can see Granola is already added and go ahead and kick off an agent to look through my meeting notes. Um and we have a variety of different tools and plugins. You can also check out our marketplace. Um we support connecting directly to Slack um for both fetching context and posting messages.

17:23 Um Notion, since that's what we use that internally as our internal knowledge base. Um Datadog for all our alerts. A lot of these are coding related um since we started off there, but you can imagine we have different product management workflows, design workflows, writing workflows, and so forth. Um so, I recommend checking these out and going from there to see what works with your um usage.

17:45 And now you can see here it's listing all my meetings. It's querying the meeting notes. And then, in a few seconds it's going to return a summary of the things I've worked on. Um while this is running, I also wanted to quickly touch on automations. Um so, you can imagine I could open Cursor every day, paste in my meeting summary request, and then get a response. But I can actually also build an automation to do this. Um automations are always running agents. You can set them up on a schedule. So, let's say I

18:14 want to run them every week. And then, we have first-party integrations for GitHub for any like code actions, such as a pull request was opened. Um you can also have it whenever like a channel is created and or a new message is sent in a channel that you care about. We have one internally for um let's say Composer 3 or Cursor 3 got launched recently. And every time there's a bug report, um an agent is kicked off, adds it to Linear, figures out if it's a duplicate or one

18:38 or not. If a lot of people flag it, it'll actually upvote it automatically. And then, if a user actually didn't provide enough information, if they're like, "Oh, something's broken." The agent will auto follow up asking it for screenshots of the issue, for replication steps. And so, it's funny to see those agents often running. Um someone actually created an automation um at every day, someone they Slack the user and says, "What do you want for dinner?" They specify and then auto creates a DoorDash group order and posts

19:06 it in our food channel. So, people have been very creative with the automations they're building. Um so, or you can also actually have a webhook, which I'll be showing later, so you can integrate it with any of your custom workflows. But for this one, I'll have it run once a week. I'll have it run in just any general repo. Um one thing to note is that right now for automations, you have to have it connected to GitHub. Um you actually, if you're running it to look at Granola notes, I don't actually need

19:31 any of this um code data, but you can imagine for most um cases, you could. And so, you by default have memory, so it looks at what you've done previously. Here, I'll connect to Granola, so it has access to my meeting notes. And I can directly authenticate it here. I will sign in.

19:52 And then, that's as easy as that to connect. And then I'll also add Slack since I wanted to Slack myself about my meeting notes. And it's as easy as this to have it run an automation every single day. And then you can also test it. So I can kick it off right now. It's going to run by default every Monday, but I can just trigger it right now and then you can see it often running to the races.

20:17 But maybe you can see this one actually finished. So by recap, I had a couple external calls, a growth sync, thought talking about the GTM product, and some action items. I'm supposed working on a community program, so I should create an application form and so forth. So you can imagine getting Monday morning recap of everything I promised to do.

20:40 And one final flow I wanted to touch on is our email drafting flow. And so I currently own enterprise at along with Ethan who's on the call. And so every time someone fills out our sales form, we want to make sure they get a quick response with all the context that they need. And so I can't show you too much of our actual email data for user privacy, but let me actually show you high-level how it's implemented. And so I have this web hook running that's always listening to any calls to

21:10 enterprise at. And I'm asking it to give a tailored response based on info from our docs, all our user table from our database. So that way we know whether they're an existing customer looking to expand or a net new might might be looking for a demo. I'll have the conversation ID that it's pulling. And by default we use Front for our email inbox handling. And then I have the API key automatically injected into the cloud agent, so it's able to run this automatically. And as you can see I connected our

21:36 database through Databricks, so it can actually query all the user data. I have Slack since a lot sometimes folks are asking about newest release features, and a lot of that conversation is happening in Slack. And then Notion where we have a lot of our internal documentation living as well. And you can see I have a zap that is called through Front, which is the best integration possible. It gets requests with the conversation and then triggers our agent to run. So that's high-level

22:03 how we set up these email automations. So wanted to share few use cases that we use internally, but I'll pass it off to George to talk more about how we use Cursor on our growth team. Awesome, awesome stuff. It is funny. I was reading one of the comments in the chat and they were like, "Whoa, you just blew by and did a billion things.

22:23 How to retain all this info?" But that is Emily. Emily builds incredibly incredibly quickly, but hopefully my section will be a little bit more slow-paced in terms of explaining at a high-level when we want to kind of solve some of the more common repetitive kind of the knowledge work, how can one use Cursor to help tackle some of that?

22:45 So you've kind of seen how Emily has shown how Cursor works, how you can connect them to CP's and build workflows. I want to at a high-level show you what happens when you take that same philosophy and apply to specific business problem we had internally. So giving the agent context, giving it tools, and iterating. And I'm going to go ahead and share my screen. So it's a bit easier to follow along.

23:11 Okay. Let me know if you're able to see my screen. Okay, awesome. So um our sales team had a problem that every go-to-market team has, which is prospecting and outbounding. So specifically finding the right people to reach out to, figuring out who to prioritize, doing research, and writing personalized outreach. And an SDR might spend anywhere from two to three hours doing this for a single account. So what we wanted to see was what if we brought AI into their workflow as a true pairing partner.

23:53 And here's what's interesting, and I think this is really the story of how knowledge workers can leverage tools like Cursor. We actually didn't start by trying to build an app. I believe someone in the chat was asking about this earlier like, "Hey, are you guys how can this be, you know, applied to non-coding tasks?" And initially we were we kind of started by just solving one small problem at a time right inside Cursor.

24:18 So for you know, sales reps who are trying to generate a pipeline, the first thing that they wanted to solve was like, "How we have data living in many different places? How can we bring the data together?" That was the first step. Our SDRs were pulling from five different tabs, Salesforce, Gong, usage dashboards, LinkedIn, enrichment tools. So the first thing we did was sit down in Cursor and describe what data exists, where it lives, and have the

24:44 agent write SQL queries to consolidate. I am not a SQL expert. But what I described was I described using plain English what we were looking for. The agent wrote the query, and I along with the SDR eyeballed the output and said, "This doesn't look right. The numbers are too low. This looks good." And we just iterated. And the second step here was, you know, within an SDR's workflow, how do you probably then prioritize? You have all the data in one place now, great. But not all leads are created equal.

25:14 So we described the business logic. Seniority matters. Usage um recency matters, when the lead come in, all in natural language. And the agent then turned that into a scoring logic. Wrote a Python file to help generate and score. And we tested it against what experienced reps would do intuitively.

25:32 And whenever the output didn't match their gut, we just adjusted the weights back and forth and iterated. And once we were able to get that in a good spot, the next thing that reps were asking for was like, "Okay, great. You have been able to help score and rank everyone. Can you now help figure out how we can reach out to these folks?" So the next thing then was like, "Okay, hey, like what actually comes into drafting a really good message? Understanding what are what are their AI

25:58 initiatives? What do the company do? What is right angle?" We encoded kind of what makes good outreach, so guidelines, not rigid templates. And the agent learned to do the research and draft contextual messages. I think the main thing I want to kind of highlight through this all is when we were going going back and forth and iterating, a lot of it was very much using plan mode. At any given point in time you describe the problem, and I think Emily did a great job of this of like, you know, one of the

26:25 things I've found to work really really well is like we have a lot of context in our brains as knowledge workers. We understand what good looks like from just having many many years of experience doing the same thing over and over and over. And we kind of are able to pattern match and see what the main things are. So a lot of it is how can you describe to an agent like, "Hey, this is actually the task I want to do. This is roughly what success looks like." And then the agent then takes the first

26:50 stab, plans. You might say, "Hey, this doesn't exactly look right." You give it feedback. It updates. And then all of a sudden now you're running. The agent executes. And then you then test it on a few examples and say like, "Hey, this actually looks good. This doesn't look good." And it's a constant iterative process until you've tackled a full end-to-end workflow. But it all starts with one problem. And you go one by one until you feel good about the output and

27:12 you're able to then extend it. So the main thing I want to highlight here is at every given step, I along with the SDR and the team building this was we were the domain experts. We knew what a good output looked like. The agent was very much the execution partner. And you know, the tool through it all was I I absolutely love plan mode because it's like you want to give it enough context so it's able to do really really good work because agents are

27:39 really powerful. The underlying models are so so smart, but often times they just don't understand the way we do our work. So dumping context, iterating on the approach before building it, a lot of that just came from just brain dumping. And often times my my my fat fingers are not able to type as quickly, but a lot of it is enabled when you just turn on voice mode, give it all the context, let it ramble, give everything that's stored in your brain to the

28:02 agent. It then does a really really good job. So what I kind of want to highlight is the process here. The problem, the context dumping, the iterating, the testing, and then the refining. And that loop, I would say in a nutshell kind of works and can extend to lots and lots of knowledge work.

28:19 So that that iterative loop kept going, and the requests kept coming. What initially started as, "Can you pull the best people to reach out to?" soon became, "Can you pull the best people? Can you tell me who to reach out to?" And then suddenly, "Can you also draft the messaging that goes out to these people?" But soon folks realized the power of this given it was linked to a lot of our underlying data and systems was, "Hey, can you also now help me do some high-level account research, build an account plan?"

28:46 So there are a lot of other use cases that came through through the pipe. Can you help me prep for renewal? So what we realized is that we should just give everyone the same powerful tools that we had wrapped in something that they can just use. So we built something internally called ChatGTM. It connects to all our data sources, our CRM, our usage data, enrichment providers, call recordings, etc. and gives the agent the same tools that we were using in Cursor.

29:13 So let me show you at a high-level kind of what that looks like using some obviously this is some of it's going to be fake data, but hopefully it gives you a sense of what is what is possible when you just start with a single problem. You're not thinking about building an app, but you're looking to solve that problem and over and over and over. Over time as you tackle more and more workflows, you're able to then string all that together and do something that's a lot more powerful.

29:36 So let me go ahead and show you what ChatGTM looks like. So we have built this internally to give reps um leverage. And the goal here is like the rep is the subject matter expert and we are learning from them every single day. They have a lot of context in their brain, but we want to give them the powerful tools to enable them. So a very good example of this is like hey um you know with prospecting with outbounding you need to build an outbound campaign. It

30:05 takes time to pull the folks then figure out who you should reach out to then um within who you should reach out to research the account etc. And it's like and then drafting the LinkedIn message. And so our our task here was like hey with the right tools can the agent tackle this with a good first pass with no human in the loop now that we've wired in the context of how our best SDRs go about their outbounding. So prompt here was like hey go here the

30:35 find the highest likelihood contacts that we should be reaching out to research the account draft copy for me do it in one shot. And the agent actually do able to do a pretty good job. So at a high level I'll just call out a few things that are happening here. The first thing is it's going out about looking and kind of exploring the tools that we've given it pulling the relevant data. And then what it'll then do is basically put together a pretty comprehensive outbound brief. So

31:04 gets you up to speed in terms of the account what had happened how much usage is there. It also goes out and about and looks at the at the internet does web search to figure out who are are there you know strong signals from recent events that have happened leadership transitions job hiring job hosts etc.

31:24 And then from there it will also go into well one it'll try to find any warm contacts of folks that we previously engaged with or folks that are already existing users and the goal here for an SDR is like hey I want to be able to create an opportunity to have a conversation about exploring whether or not an enterprise plan would make sense for them and it'll go out and about do this for warm contacts and separately it'll also be able to go on LinkedIn and find

31:52 people that are relevant for you. So earlier in the prompt we specified like hey we want to find folks that are director plus of engineering it goes out accesses LinkedIn and is able to pull these folks and also suggest like hey based off of their background they might be relevant for XYZ angle and this is the recommended kind of way to approach the outreach. And the one thing that's been pretty mind-boggling for our sales reps is it is also smart enough now with the right

32:21 data sources and tools to potentially into it and guess and infer what the org chart might look like. Obviously this data is never going to be 100% accurate. But work that the people you know days and weeks to be able to put together in the past now yeah I can take a first stab at and triangulate across many different data sources to suggest what the org chart might look like. And with all of this kind of research being grounded and giving kind of the user perspective on like hey who might

32:50 we potentially reach out to is now also then drafting hey for Raj for example director of software engineering power user and a leader this is the potential angle across LinkedIn across email for Chris who is someone who's not using the platform today but reports into the VP what the angle might look like here. So that is a good example of I think I just wanted to kind of give you guys a a a a high level glimpse into what might be possible when you start with a very very specific

33:21 problem and try to tackle one bit and piece of the workflow at a time of the problem and a lot of it kind of just goes down to can we describe the problem can you describe the context to the agent um so that they understand how to potentially build around that and then after that a lot of it is testing testing and iterating like what is what is the output of the agent being able to provide and we as knowledge workers the experts who have been doing this job over and over and over are able to very

33:50 quickly validate or invalidate and iterate and pair with the agent to do this work. So it's a very very magical experience being able to go from problem to solution so quickly. I feel like as especially as folks that don't have a coding and programming background it feels incredibly empowering. So yeah I just wanted to give a very high level overview in terms of kind of what what we've done and you know what you just saw was you know honestly built in a number of weeks

34:19 not months and the underlying process the iterative loop of here's a problem here's context let's figure it out together that's available to all of you right now in cursor. So whatever your domain is whether it's sales product data marketing support you have deep expertise about how your work gets done what data matters what a good output looks like what the edge cases are that knowledge has always lived in your head and what's changed is that now you can express that expertise

34:50 in natural language connected to real data sources and let the agent do the repetitive execution. So you go from I know how this should work to I have a tool that does it potentially in you know an afternoon. So if you're curious about any of this how we built it how you might apply it to your teams workflows please fire away in the Q&A or reach out after happy to go deeper. Awesome. Thanks George. Well with that we finished our plan content happy to answer any questions and stick around

35:21 for a few more minutes if you wanted to drop them in chat or Q&A. George is that an app that was created by cursor chat GPTM can anyone do this? Yes this was entirely created by cursor. When I first started full disclaimer I am not an engineer I am far from an engineer and Emily will be the first one to validate that. >> [laughter] >> But yes this was entirely built by cursor over the past few weeks. This lives so this is an this is an internal tool that we built for our for our team.

36:04 Mhm. And then I see there's a question best use cases for product managers to use cursor daily. I've run several cursor for product manager sessions happy to also check out the recording on our workshops page but one big one is actually estimating timelines. And so you can imagine you have a feature in mind you can ask cursor help me scope it is this going to take days weeks months for a team of X people to implement. The second one is actually creating PRDs. It can work with like plan mode really well

36:32 on looking at the code base scoping the project but internally we've also let like cursor enables teams to like show not tell what they're looking for. So Jacob one of our PMs if he's thinking about adding a new teams plan a new checkout flow. He will actually go ahead and ask cursor to create the modal overview and will like kind of show a live demo and the design team the ends team will take a look click around and if it works well directly ship. And then separately I think a lot of

37:01 people can directly start shipping stuff. And so a lot of like one very easy entry point is docs. And so a lot of times people will come and like hey like we changed the pricing oh like we added a new feature can engineer be pulled into update the docs for us. Now product managers will be like oh look I see something's off at cursor please fix here's the right info that should be there and I'll automatically do it. And some of them have actually been

37:27 shipping features as well whether it's smaller ones or bigger ones as well. Then how do we get early access to cursor latest and greatest? We actually have early access within the IDE. I can quickly share what that looks like haven't seen it in a while but you can go here under beta and then update access to be within our doc fooding team and getting night nightly is released once a day with our latest features so I recommend checking it out.

38:00 One question that I see Jeff submitted this is a very good question and not an easy question to answer but he said at cursor have you thought about how to manage the explosion of useful tools that people are or will be creating like where to host access control security etc. Um that is a very legitimate question and you know this is now empowering many folks to be able to build.

38:22 I think at a high level internally the way we've some teams have have kind of thought about this is it's easy to build a tool but a tool is not always uh going to be valuable immediately at the gates and adopted but if it there is strong adoption then very quickly the question is okay is this is this tool pulling on and using the right data sources and have the right access controls etc. But I think the culture here has very much been giving people and empowering people to be able to use all the tools at their

38:55 disposal to be able to solve their problems and and I think that has very much been the culture here at cursor but I think to your point with proliferation of these apps it then becomes very difficult to centralize and maintain but I think the kind of crawl walk run approach that our team has is let's see if someone builds something the validation piece is is it being strongly adopted and if it is being strongly adopted then very quickly it's okay do we have the right access

39:22 controls do we have the right data guardrails if people are going to soon rely on it. But I think that's honestly a question that is one that we're trying to learn how to solve and we're kind of flying the plane as we're like building it. I'm not sure that this is a question that has been like super super well solved in in orgs just given these tools are are are so nascent. Mhm.

39:45 And then I see a kind of a lot of questions about working within a large enterprise. Um right now kind of all this most of the rules and skills are created on my local machine, but instantly if something works well, if you wanted to share access, you actually want to promote it across the entire team. Or if you're an admin, you want to kind of distill it across the org. So on our Teams plan, you actually um by default can create team-level uh project-level rules. So you can save them within the

40:08 codebase, so someone who works within that repo gets access. And then you can also um on the enterprise plan set them up team-wide. And you can set set set it up. Right now we're working on on organizations, which are launching in the next few weeks. So you so you can specify design team has these um configurations, product managers has this one, security team has that one. Um but by right now you can also like specify org-wide. For example, some of

40:30 the ones we have is security rules. And so maybe when you're writing SQL, don't allow injections. Um and those protec- um protections are added org-wide. Um I see some questions on what's the best way to have Cursor deploy a locally hosted dashboard. I would recommend checking out Vercel.

40:50 Um Vercel um we use it for actually hosting cursor.com and all of our web surfaces. Um but it's also good you can create a personal account and if you install the GitHub CLI and the Vercel CLI, it makes it really easy. It'll actually push your website to be hosted on the cloud and deploy um via Vercel as well, so you can share with your team.

41:10 And hopefully that answers Timothy's questions, too. I think there were a number of people that were asking, "Oh, if we were to build a version of uh chat GTM or some, you know, internal tool, where would it live?" Um and hopefully Emily's answer there answers the question.

41:30 Um are there Right now, are there any plans for building out the browser capabilities in the agents window? Um that currently is supported. Um and I know I worked went through it a bit fast with a lot of different stuff, but you can actually can quickly share my screen and show you. Um but right here directly, you can go um open the sidebar panel. And then click the I mean, let me close it so you can see going from it from scratch. You can click the globe. And then you can go to the website you want to use directly.

41:57 Um and let's say I wanted to change some stuff. I can actually go ahead, click here, ask changes. Uh maybe I want this to be a bold. It I'll actually go ahead and make the changes directly in the IDE. Um you can also imagine wanting to take a screenshot, can take a screenshot of the section, and pay add it into chat's context. Um so makes it really easy to work directly with your website um in the IDE.

42:36 Awesome. Any other last questions that we should uh be fielding? Um let me Oh, go ahead, George. No, there's one other question uh that someone someone shared, which is is a great question. Great question, Yaseen. I I want to build a startup, so how can Cursor help me in this regard and what will it be?

42:58 That is a really, really good question. I think hopefully the uh the latter part of the presentation um was able to show like, "Hey, like if you have a specific problem, um now we're now in the age of where the models are, we are so empowered regardless of whether you have a technical background. And especially so if you don't have a technical background, to be able to go from problem, describing your problem, describing how to solve it, um to being

43:24 able to potentially turn that into a product. So initially when I was trying to solve this, I was not trying to build a product. I was just trying to uh just basically create a table, write the SQL query that someone could then use. Um and then over and over, um it soon became you someone had a problem, you solved it, then they had another problem, then you solved it. And you just continue to kind of iterate and iterate iterate and and you're you're

43:45 able to provide value from just being able to go from uh someone's knowledge, um expertise, um and codifying that into something um that is like purpose-built for them. So um hopefully um Yaseen, that gives you a high-level answer. Um yeah. Awesome. And then if you had any questions afterwards, I wanted to show one of my favorite features um in the Cursor docs. You can actually chat with them directly and ask questions like, "Is GitHub Enterprise supported?"

44:20 And then it will actually go through, look through our docs, and give you an answer um using the same underlying models that our agent uses. Um so yeah, looks like GitHub Enterprise is supported. Recommend you paste questions there if you didn't get to them during this call.

44:38 Awesome. Um and then I saw a few people asking about Microsoft SQL. Um ask agent about that as well, but I recommend checking out We have a SQL Server extension that um is a VS Code extension that we support as well. And you can actually also connect via MCP. Um you can probably ask Cursor how to do it and it'll give you instructions better than I could live on this call. Um and you can do queries directly there as well.

45:02 Well, with that, thank you so much for joining and have a great rest of your Thursday. Bye, everyone. Bye.