Name: Build Your First Claude Skill
Uploaded: 2026-06-30T18:16:19.349Z
Duration: 59 min 32 s
Description: Build Your First Claude Skill

Transcript for "Build Your First Claude Skill": Alrighty. Okay. Folks are trickling in here. Hello, everyone. Welcome. Welcome. On this Tuesday, we are calling in Siddhant and myself from the San Francisco. I hope everyone is excited about, if you are joining us from The US, this upcoming short week for Independence Day, woo, July 4. Also, lots of folks from all over from Scotland, Romania, Colorado. Romania, Colorado. Awesome. Ireland. Great to see you all joining. Amazing. And, again, thank you so much for taking this time on a Tuesday to join us. We're just gonna get right into it today. We've got a lot to cover, so welcome, everyone. I'm Cassandra. I work on product marketing at Assembled. Assembled is the first AI native workforce management platform built for modern customer support teams. I am joined here today by Siddhant Shetty from our solutions engineering team. We've also got Melissa Evans with us backstage. Mel is gonna be monitoring the chat a little bit just to ensure Siddhant and I can get to any and all questions. I can see some of you there in the in the chat or in the q and a. Feel free to, to pop in there. We wanna keep this as conversational as possible. Another little housekeeping note here, since we will be sharing our screens and just jumping into Claude, I want to alert you to the little Zoom buttons in the webinar window here. You can see those little magnifying glasses on the bottom right corner of your screen there. You can zoom in or out there if the window is a little too small for you. So let's just get going here. A little bit of agenda to orient us. So like I mentioned, we do wanna spend most of our time here in Claude today, but we'll give a really brief primer on a few core concepts, just to level set, make sure everyone, has some of those in their back pocket. Then we're just gonna get right into a demo of what's possible. Sid is gonna open up Claude, show you some of the possibilities of what happens when you have Claude skills. And then, truly, I'm gonna take a lot of time today for us all to build a skill together in Claude. Hopefully, y'all can follow along at home. And just to also set the stage here, I know everyone's kind of on a different, place in their AI journey since we've got, comments going on in the chat. Out of curiosity, I just wanted to get a sense of familiarity here. If you've used Claude's skills, could you drop a thumbs up in the chat if you've used a Claude skill before? You could use a thumbs down too if you haven't. Okay. Lots of thumbs up. Awesome. Very cool. Okay. People are using quad skills. And if you have it, no need to respond. You can even give a thumbs down if you want. But I'm curious too. Has anyone made a cloud skill? If you have, could you drop a smiley face? Maybe something a little different than the thumbs up. Give me a awesome pause. Got it. Kaylee. Cool. Okay. So we've got some folks who are familiar with skills today. Everyone's sort of on a different, you know, a different page in their, like, AI adoption journey. So, hopefully, no matter where you are on that journey, there's something you can learn today or take home with you. We will have a little bit of q and a at the end. We've saved some time for that. But like I mentioned, please feel free to ask, throughout the session. Just go ahead and and pop your question in the chat, and we'll get to it live. So a couple of core concepts here, like I mentioned, that I want to, establish, and these may be new to you or just super obvious, but I think, two things that unlock everything you wanna do in in cloud. And one is connectors. I think this is known to most folks. Connectors are what, allow access to your data and systems, also called an MCP, an MTP server. It's just how cloud reaches your data and your tools. Siddhant will show off a little bit of the assembled MCP. We recently released in his, demo in Claude. MCP is just a technology underneath that makes connectors possible. And then today, what we're really talking about, all about Claude skills. So quite a few of you are familiar with using cloud skills. We're gonna talk about best practices for building them today. It's really just this reusable set of instructions. You'll see these both in action shortly, and we'll go deeper. But I like these analogies for both of these, bug concepts. Connector, it's like an integration. You can securely work with all your systems, whether that's assembled or is it Slack or Salesforce, Notion, whatever without all the custom integration rigmarole. And connectors go both ways. So you can read, meaning you can access all that data, look at it. We can also interact with it. So you can also take action in those tools with connectors, which is pretty sweet. And then a skill. I like this analogy for skills. Just an SOP, a standard operating procedure, something you document for, like, a new hire. But for Claude, you make it once, and then you can run it constantly. It's really great, for something that needs a consistent process and then also a consistent output. And today, what's great is right next to the chat window, there's a little tab called docs. That's where you can get your free pack of cloud skills that we've made for workforce management planning, workforce vendors, planners, and support leaders. You can customize and distribute those skills however you like because in practice, the skill is really just a zip file, which is why they're so popular today. I mean, there's all these, gosh, like websites with thousands of user generated skills, but these we've created specifically for support operations. Everyone can access those in that Notion doc. We're gonna use a few of them today and show you how you can customize them for your operation. And If you don't share a cloud account, by the way, you know, everyone can see your shared skills if you have a cloud account that you share with folks, can share the same skills. If you don't have a shared cloud account, you can just send them in a file like we did with the zip file in Notion. And really the power with connectors and skills, I'd say most folks know about MCP these days. I think skills are still a little bit underrated even though they're kind of, they've been around for some time. I think most folks who are using, agents or Claude today, they're not using skills a ton. Even, us at assembled, we are, just sort of beginning to see how powerful skills could be. Exactly, Paul. Like, they're super powerful once you've created your own and especially once skills are connected to all your applications with your connectors, then I mean, you're really just TEDxing the way that you use Claude. So just a little background and slide where, but now I'm gonna pass it over to Siddhant. We can actually get into Claude and, talk about what's possible. So, Siddhant, I'll pass it over to you. Oh, and you're on mute there, Siddhant. Classic webinar mistake. Isn't that great way to start? Alright. So here we get into Claude. And I think everything that Cassandra was talking about from a MCP perspective is important from the point of view of skills are great from an instructional perspective. But how can you supercharge the value you get out of Claude? And that's kind of where the MCPM, the connectors come in. So, Paul, to your point, skills are amazing. On top of skills, everything that works is based on data. And I think one of the key factors in all of the things that we're talking about is getting the data right. assembled spends a lot of time in ensuring that we get the data correct, and it's audited, verifiable, traceable. The same can be said about other systems as well. So for example, over here, I have my connector section opened up, and I have my cloud instance connected into a bunch of tools. As I was going through this and I was thinking about, hey. What example do we talk about? A couple of weeks ago, we had a conversation with a customer, and a very interesting use case came about. They were looking at how to ensure and I hope all the planners on the webinar have already done this, but they wanted to ensure that they retain coverage for the fourth of July long weekend because their teams in The US are unavailable. And what they wanted to make sure is that would it be possible to enable AI voice automation across dominant queues so that they can still get good SLA coverage across those couple of days. So now in this example, what I have is Snowflake, which many of you might have as your data source, where I can cross connect tools such as Assembled, Snowflake, Salesforce. And this really opens up the power of cloud where you can do cross system analysis. Typically, you would go to a Power BI or a Tableau or any kind of a BI tool to run examples like this. I'm sure many of you planners or leaders have had situations where somebody is breathing behind your back for a particular report or a particular analysis. That's the kind of stuff that in cloud can be done in just a few minutes. So what I wanna show right now is one of two ways in which you can create a skill. The first way is where you ask questions in cloud, try to get insights. Once you get those insights, you figure out a way to act on them and action on them. And once you action on them, the final step is I've done this once. How do I ensure I never do this again? And that's where the automation comes in. That's where the skill creation comes in. So for example, now let's go to a new chat and look through my snowflake data and figure out what percentage of phone conversations were resolved via AI. This is a starting point for me in this example to understand with the AI operations I have running. Many of you might have some level of customer agents that are accessible across chat, voice, or email, but the primary goal over there for your tier one or tier two tickets would be to deflect or automate them. So what I'm trying to do right now is connecting into Snowflake, as you can see right now that cloud is doing, and getting a baseline on what would be the rough resolution rates that I should be expecting. Because what I want to do after this is figure out based on the resolution rate that AI can offer me, how can I offset this across different queues within my assembled instance? So that's something that we're gonna work through in this example. As this cooks through, you can see that the number we're getting is about 9.86%. So about, like, 9.9 or 10% of our volume currently is automated across voice and is not handled by a human. Now the next thing I wanna check would be let's get a sense of what's the data split across my queues in assembled. So I just pulled some data from Snowflake. And now I wanna check what are my dominant queues within assembled. The reason for that is, in this example, my voice AI setup works in a way where I have to define a SOP. And that SOP is something that I have to define at a queue level. So I want to try to figure out a way where I can get, like, maximum bang for the buck. So in this case, for example, it's checking assembled, and it's pulling out the different queues and trying to give me a sense of what would be the overall spread of data. While this cooks, I'm gonna just switch over into the webinar to see if anybody has any questions. So far, so good. Instead, what might be interesting is could you show me. just, like, if you want the plus sign, could I just see your list of connectors? I'm always. curious what people have. Absolutely. So in this case, for example, you're seeing it's connecting Snowflake and assembled over here. In assembled, it's trying to look at the actual versus forecasted data to get a sense of what is my overall queue split. And here's what I get, the different splits across the queues that are coming in. It takes a second. And, again, I put in a slightly bigger query. But the idea and the goal over here is this is how generally people go about their day within cloud, but they're trying to find answers to questions that they don't have the answer to. And it's a probe and figure out what might be the best way to look at the dataset that's available to you. So in this case, for example, right now, it's falling and adding up the values that we're seeing over here. Again, this is a referencing exercise in terms of getting data. Typically, when we create skills, there are slightly different ways in which we have the LLMs do math. Many of you in the past might have seen that sometimes LLMs can be a little funky when it comes to math, adding, updating numbers. So that's something to just watch out in a column that they're providing over here. As this cooks in, I can see that some of the volume across my queues is what Assembled is able to pull out, and it's adding across the date range over last week. Let me just now figure out the next set of a prompt just so that we are on time over here. So what I want to do next is for the July 4, long weekend, which is, like, fourth to sixth. All my US agents will be off. Walk me through the phone staffing gap this dates, then assume a 9.9% AI resolution rate. And let's take a look at a particular queue. I think tier one has a good amount of volume, so let's pick tier two. And write back a forecast adjustment into Assembled for phone and tier one for the long weekend. Once you're done with this, send a Slack message to support team leaders. Okay. So cloud is done cooking over here. And as you can see, it's giving up the volume across different queues that I have. And you can see inbound and tier one account for 15% to 17% of the traffic over here. So going back to the example of what our customers did, they looked at the overall data that they got from Snowflake. They connected that into Assembled. Then they realized, you know what? It might be best to attack either tier one or inbound as a queue, set up a automation. And based on that, the overall volume that their humans will have to work on would be offset by the amount of automation that they can see, which in this example, I'm assuming it to be 9.9%. And what I want to do is I want to write back this as a forecast adjustment. So many of you who are Assembled customers or many of you who are not Assembled customers, just to talk about what a forecast adjustment is, within assembled, you have forecast on what the demand might look like. And you have the ability to make tweaks to this dataset in the form of adjustments or carving out outliers. Now what I'm doing in this example is in a headless way, I'm trying to impact this directly from cloud into assembled. So up until now, we were looking at this from a perspective of let's ask a question, get an insight. And now we're looking at this from the other way, which is I need to now act on it. So this is my first step into acting. And as you can see now, it's, again, pulling into Assembled. It's looking at the buckets, and it will write back the forecast adjustment. So this is the action I'm taking on top of it. To additionally compound this, it's not just an action happening in vacuum. The action is also getting communicated over Slack to everybody within the team. So this is one way to ensure there is some level of audit within the automations you set up is that you add a layer of communication or a layer of confirmation. And let's see as this cooks. Let's check if there are any questions from the team so far. So far, so good. So if you see over here, what's happening is that, a, we're able to look at the agents within the environment I have who are tagged as agents who work in The US, and it's figuring out what the forecasted volume will be, what the overall gap is, and it tells me that I am in a deficit right now. So that's kind of the analysis part that I was talking about. Now look at the second part, which is before it writes back the adjustment, it's asking me if I should proceed with it or not. And I kept the write back action within my Assembled MCP to a permission level that first requires me to explicitly allow it. And this is something we recommend as best practice where within your MCP connectors, you will always notice that the rights and privileges that the connector has are typically split between read and write. Read actions, generally, we allow across the board, and we recommend that you always set up write actions to have a approval layer on top. It's just from a best practice perspective, you don't want something written in while this is running, just to show you guys again what I was talking about, is if you look at the Assembled MCP as an example over here, and you will see this across other tools as well, you will see what are the read only scopes available, and you will see the write scopes available. So if you see from a permission perspective, there are three levels of permissions over here. One is blocked, one is allow across the board, and then the option in between is where it needs an approval. So if you see all the right tools have been set to, it needs approval at the moment. I love seeing people's, MCB connections. I see what you're connected to and what you're not here. So it's interesting. Mhmm. And, again, your setup might vary. The kind of connections you have might be completely different. But, also, talking about this as a starting use case right now. There are more advanced scenarios where you might actually set this up within Cloud Code. You might have other APIs connecting in. So there's a lot that you can do with this. Right now, what we're doing is a starting level exercise on how you can convert your natural conversations from just getting insights to actions and then finally getting those into automations. So now if you come into Assembled, quickly change the data ranges over here, and then change the queues that we have to label. You can see the deflection now. So the adjustment's now showing in over here. So this automatically tweaks my forecast based on what I just sent in through Claude. And in a second, I can show this to you as a Slack message that's come in as well, and that's available over here. See if this opens up. Ah, there you go. As you can see right now, we just sent a Slack message sent via Cloud, and everybody now knows about can anyone see my Slack, or am I opening up a pop up? That's a good question to share. It looks like a pop up there. Yeah. Oh, just a pop up. Yeah. Okay. One second. Let me And we've got a question here from Paul. Are there any. limitations on what actions and data are available in the Assembled MCP? That's a very good question. So in terms of right scopes at the moment, we have started off with adjusting forecast as a starting point. We're expanding this one by one into different areas. And in terms of the road map of what it will be, let me open the Assembled APIs, and you will notice a lot of almost all objects within the Assembled API will eventually be part of the MCP roadmap. But today we've started with what we call the forecast adjustments. Like I said, for slightly more advanced use cases, you could always always go into cloud code and connect the Assembled API as well. Alright. So now we've done our actions. We've sent the Slack message. We've sent the adjustment over. Let's do the final part, which is this flow is something that I really, like, don't wanna do again. It took me a lot of time. So in this case, let's turn this into a skill called US holiday automation. And the steps I'm giving over here is to, a, make sure for any upcoming US holiday, a week in advance, You pull the agents who are available for unavailable for that period, check the overall automation rates you're seeing across Snowflake, and then create the forecast adjustment into Assembled as well as inform people in Slack. So now what cloud is doing is everything that we've just worked on, it's gonna convert this into an abstracted way that allows you to set this up as a skill. And while this runs let me go into the questions and see if there were any questions that are coming. So we covered limitations, disassembled mask, PII. Yes. It does. In the MCP scopes right now, we've set it up in a way that we do not get your PII data because it's for workforce management. And at the integration layer within your instance itself, we would have set up gates to ensure that we scrub or mask PII before consuming itself. So the MCP on top of that is only consuming data that's been prescrubbed and is devoid of any PII. Any other questions from the team? Absolutely, Paul. As as yeah. As as weeks and months go by, I think we're gonna have more and more objects opening up. I recently made a cloud code automation, which use our API for slightly more advanced scenarios, which, again, pertain to the work that I do with the assembled. But, for users who are non tech and don't want to tinker around in some of those territories, the MCP will expand. As this rolls up, I think a couple of things to just give as best practice advice. This example had a lot of calculations. The starting part was very much like inside generation. In terms of creating a skill, typically, when you are running a lot of heavy math and you want to be deterministic, always, at the instructions, add things such as do not estimate, compute all totals, Maybe use a Python function just so that the math is being done on a calculator and you don't have any LLM induced errors. That's one, recommendation that we do. Make sure that the LMS are always giving you verifiable numbers so you can check and audit everything that comes through. So those are, like, best practices that we typically recommend. And I will say, I think the good thing about the Assembled MCP and most MCPs connected to other data is, we're doing all the the calculations on the assembled side. You're not getting, like, your forecast from math that's happening within Claude. But I think we are showing this as creating a skill. I think one way we can also maybe even while that loads, I can talk a little bit about the other ways to create a skill as we go back to sort of first principles here, and maybe I'll even share my screen here if you don't mind, Siddhant. Of course. Let me give it over to you. Amazing. Okay. So I think Siddhant was sort of showing here, really the the steps of, like, interacting with, your data with assembled with your connectors and then creating a skill from the outputs you have there. Now let's say like, let's go back to first principles, and I wanna just, sort of level set on how you should think about these in general. And whether this is a Gemini gem, a Claude skill, a custom GPT, I think this is a really often overlooked step when it comes to, creating any type of custom agent. Planning is super important, especially in these days where we are wasting a bunch of tokens and that has suddenly become important. So if you don't wanna spend your time prompting back and forth, taking a second to plan your inputs is is important. So first, really simply just identifying your use case, I think, is key. Like we mentioned, a more repetitive task with, consistent process and outputs is a great candidate for a cloud skill. And then, the exercise that we do here even at assembled is defining the starting and end states. It's often like a whiteboard session will do. And, really, it's just saying what's your starting point, which may be zero. Maybe you have nothing. And what's the ideal end state for this skill? So is it a PDF, a Notion doc, or is it a whole workflow like Siddhant just showed where you have, actually a, some data that runs an adjustment to assembled and then a Slack message? Is it that whole workflow? Once you start from your starting point, you have point a, here's where we're at. Maybe it's like, I do have a report I like, but I like to turn it into an automation, and then I understand my end state. Here's all the things I'd like it to do. Then you can fill in the input. So it's like, do I need to find some reference files to give my agent so it knows what good looks like? What tools do I need to connect to. Do I need to be connected to my data warehouse, to Snowflake, to assembled, to Salesforce, whatever it might be? And what should this data look like? What should we be pulling out? Also to determine all these things upfront so you're not looking around for these docs or setting up your connections while you're building. And so as we level set there, we'll just get back into cloud. No more slides. I do wanna note, I wanna make it feel like we're building together. Let's get in there, follow along if you can. But in the interest of time, we have preloaded a few chats. You could even see, like, in this example, we try to keep it live, but we preload things. Skills really take a lot of processing power. Often when I'm creating a skill, it'll be something I do in the background. I have my, like, cloud window open. I'm doing something else. I'll kinda come back to it, do some tests, adjust it. Yeah. It's not something I think you're you're gonna do in just one chat window. So with that, why don't we just get into our chat here? So from Assembled, I just want to give you all a little bit of the anatomy of a skill. If you go into Cloud here and you hit customize, this is where you'll find your skills and connectors. Here are all of mine. You can see some of my very marketing focused skills here. I'm going to open one up, a simple skill here just so you can see the makeup of it. Right here, I'm gonna show you one called internal comms. And this is a skill in the anthropic library. Anthropic has a ton of premade skills that you can add quite easily. And this just shows a little bit of the makeup. A skill is really just a prompt and then some reference files. So this one's quite simple in that we've got the prompt, and then we've got some examples. Like I mentioned, you can throw some examples into your skills so the agent knows what good looks like. Here, you can see we've got a company newsletter, some FAQ answers, general comms. These are great examples of all of, the examples you might give to Claude. And a simple skill too may not have may not have any examples. It could just be the prompt itself, which is totally fine. That's also effective. But they can also get as complex as this skill I'll show you that I also highly recommend you add if you are going to be building skills. So this is the skill builder skill creator skill, quite meta. And here you can see the makeup is quite complex. We've got sub skills. These are the agents here. So Skills used to be called agent skills. There's a ton of sub skills within this, skill creator skill. There's also several assets. It can even get as complex as having custom Python scripts. All of these things, it's, it can get as complex or as simple as it needs to be. But these ones in the anthropic marketplace, are very simple to add and really, really useful. I highly suggest, especially for today, adding the skill creator skill. That's gonna make you every skill you create follow, Anthropic's exact instructions without you having to do too much. All you do when you add a skill this way is just hit the plus sign. It will show up with all your other skills. And then there is another way to install skills and that is just by uploading them. So you're gonna go here and you are going to hit create skill and you're gonna upload. And what we're gonna do is we're gonna take that Claude skill from, this Notion doc that I mentioned. This is in the docs tab in the webinar. It says, free cloud skill pack. This is the cloud skills that we've created, specifically for support ops. We've got a few in here that you can take and make your own. Right now we're going to check out the performance report skill. This is definitely a good candidate for, core skill, something you do repeatedly. It's usually the same metrics, something you run on a weekly, monthly, quarterly basis. And what you do, as I mentioned, just a zip file. You can just download that. And right here in upload, you're going to add your performance report skill after you have saved it locally. And here it is. Super simple. Like I mentioned, there are the anthropic skills that you can find on the marketplace, but there are also a ton online, which you can use as a great starting point. Upload them just like this. Awesome use cases. Definitely recommend checking them all out. This one in particular, you could see the description here. The intention is that this produces a super tight word doc with a headline, a scorecard. It will create, some key KPIs, root cause analysis, decisions needed. So you can go ahead download that. And what I like to do when I first download a skill is to just test the output. So I'm gonna go into a new chat right here, and we can talk about how we trigger this skill. So most folks I think a few folks here, you sent me a a thumbs up that you've used skills before. So this is something you're used to doing. But the cool thing I like about skills is they can run automatically with the right context. So, for example, you might have seen in, my bank of skills, I have one called stop flop. And, what that one does is when I'm writing something with Claude, it will run automatically to make, Claude's output just sound more human. So no em dashes or other classic AI tells when it comes to writing. And I don't actually have to call the agent. I don't have to call the skill. It has enough context. Claude will have enough context to know that if I'm doing a writing task, it will automatically use that skill. But if you'd like to trigger the skill intentionally, all you do, super simple slash command, and then you can see here's all my skills. I've got them all. They have the description, which is great since there could be so many. Then I just hit Run performance report. Now, this one, what's great is this one has two paths. If you are connected to assembled via MCP, it will pull that data. However, if you're not connected to the assembled MCP, it will just ask you for the information it needs to, to create the report. And so I've got clogged cooking to use a sense term here like that, clogged cooking. But in the meantime, I've preloaded this same query here, and we can go through this. So you can see I just called performance report, and and it's gonna ask me since this, skill was set up to either run on weekly, monthly, quarterly basis. Let's get this set up. It's gonna ask on cadence. I said, let's go weekly. It's pulling some data automatically from assembled. If I didn't have assembled, this agent would just, ask for that data, whether it's being pulled from another platform and you're just uploading it or it's from another connector. Then it also asked, did anything happen this week that's not in the data? Product launches, outages, campaigns, I think that's a good ask, but I made sure it was not. Nope. Okay. Great. Now it gives me my weekly performance report, and here it is all loaded up. You can do it okay. Here you can see got a TLDR at the top, a scorecard, like I mentioned, the KPI trends. And what I like a lot about creating, this type of reporting in Claude is rather than just dashboards, you really can you can actually get some analysis that's automatic. I mean, that's what AI is great for. Right? It's like looking at the numbers and doing the analysis that you typically spend a lot of time doing. And while, you know, we're never going to replace human judgment, we should always confirm and and look through all these. It's surprisingly great at this level analysis, especially across tools. So it can save you a lot of time. So let's take this and make this our own. I think, you know, this is great, but, obviously, this is not gonna be helpful to my business if it's not customized to my needs. So it's important that we customize this with our own information. So whether that's our own connectors, our own context, refining the prompt to to be specific to what we want, Generic skills like this are a great starting point. And, I'd sid if you have one takeaway from today, go explore all the templates that exist out there, the ones we've provided in this Notion doc, but go and make them your own. That's really, the major the the major unlock here is making this useful for you. So here's how we do that. There are really two ways to create a skill at this point. So you can go in we're gonna go back to customize. And here, I'm gonna go back into skills. And to customize this performance report skill that, I've just downloaded, I can actually just hit edit. Oh, not found. You've got it. Let me give it a refresh. We're going to go into edit. Oh, and I believe, sid, I hope I'm pronouncing your name there correctly, but a 100% you will get a recording of this session. It will come into your email, very shortly right after the session. Yay. So what you're gonna do here in your skill that you've just downloaded, super simple. You're gonna go edit with Claude. And what's great is since we've already downloaded the skill creator skill, we've got a prompt that it's set up for us. Help me edit the performance report skill using Skill Creator. Perfect. And this is where you can just make this your own directly in this conversation. So you can sid, you know, call different connectors, get this information from Workday, from Salesforce, from your BI tool. You can update it to pull different metrics from those tools, whatever is most important to you. You can customize the format. So here, we've got a little memo that creates a a little Google Doc, but it could be a slide deck. It could be an email, whatever you'd like. And then, also, you can set it up in a workflow. So you could connect it to your email or your Slack, set it up on a trigger so that it runs every week. This is where you make all those customizations and really easy just to go back and forth in the conversation with Claude and do it conversationally. So now we sort of customize skills and Siddhant also showed us an example where we did sort of one workflow and then we turned it into a skill. How do we make a skill totally from scratch? And I'll show you how to do that here. Again, we're gonna head into customize into skills. And what we're gonna do is we are going to add skill again. We're gonna create a skill, and you can write skill instructions. I highly doubt people are doing that, unless they are, you know, creating, like, copy and pasting the just the skill instructions from somewhere. I've done that from, like, a repository of skills like a GitHub or something. But highly recommend when you're creating something from scratch, create with Claude, particularly because we have that skill creator skill. So you're gonna hit create with Claude, and it's gonna set you up for success already. Let's create a skill together using your skill creator skill. How many times can we say skill today? The skill creator skill. So what I would like to do though, is I would like to make a skill that, is a brand guideline skill that I can use to make anything I do in assembled branded. Because I'm thinking about that performance report I got. You know, if I'm downloading that from online somewhere, obviously, that's not my brand colors. It's not it doesn't have my logo. And there's probably also a lot of repeat use in having just a brand guideline skill. Obviously, any output, that I create in Claude, if it can have my brand guidelines in it, then amazing. So let's just tell it what to do here. And if we think through, our planning diagram from earlier too, what we need, specifically for my needs at assembled, I know I'm gonna need my brand guidelines, obviously. Now what I could do, old school style, is download them from Figma. I know I have a brand system and brand guidelines in Figma, but I know since I'm connected to Figma here, my connectors, I've got also got click, g two, gamma, all sorts of stuff in here, but I'm connected to Figma, so I don't need to upload or download anything. I do know where the link is. I'm gonna make it a little bit easier for the agent. I'm just gonna say I'm gonna skip this step because I know what to do here. I'm gonna say this is where my brand system is. I'm also gonna say, help me make a skill that makes all your outputs branded like assembled. And I've got that loaded up here. I'm just gonna copy and paste. Help me make a brand guidelines skill. Check out these guidelines in Figma. And now I could just say go find it in Figma, and that would have, like, some level of success, but not a ton. You know, let's just make it easier, especially these days. I think there was a time up until so very recently that we'd say, oh, just let it run. You know? Go let it find things in the connector. If you can bring the link in, that might help you save some some credits, make sure you don't run out of, usage. Also, if you are, have any questions about the model selection here, SONNET is a 100% great for everyday use. I like to have SONNET on with higher effort. So it uses less tokens than a, a bigger model like Opus, or even Haiku. But, it's max effort, so it will return a good result here and not burn through all my tokens. So as that's running, it's getting through all sit on the muse's opus. We have a big credit limit at Assembled. Let's say that. So as we run through the, as Claude's cooking here, I'm gonna pull up where I've, preloaded that. You can see here, I did the same thing. I asked, let's create this skill together. Help me create a brand guidelines skill. Great. You can see where we've had this conversation. It pulled up the typography, the colors from Assembled. It also asked me, would you like the colored typography for now, or would you also like the logo? I want the logo. Awesome. So then it told me some stuff, and I stopped paying attention because I was like, great. I have this skill. Very cool. Amazing. And what you do here is you just save the skill. You don't need to download it and add it. If you wanna download it to share it with other folks, you can. But in your own Cloud account, you can just hit Save. And I didn't do that yet because I know that I don't like reading all these things. So what I did instead of saving it immediately is I took that performance report from earlier and I just said, test it on this report. What that's going to do is it's going to just give me the output that I want immediately. You can see here, I'll pull up my files. These are all the things that Claude pulled up. It gave me a report that looks like this. If you're familiar with Assembled, you will know that this is not our logo, unfortunately. So Cloud is not perfect. You have to test these things. I'd say that's another best practice with with Cloud is really like when you're creating a skill, test it, don't set it and forget it. Test it with new things, see what it's doing before you publish. So I could see that that was not accurate. Right? So what I did always use the assembled logo and not that random sparkle. Right? So it did. It adjusted. And then we see that finally it has correct brand guidelines. And what I can do from here, I'm pretty happy with the output. I'll save the skill, different name. Well, you can't have double name skills. But, either way, you save it from here. And the best part is even if you've saved it, if you're in a chat and you see the skill is giving you an incorrect output, in that same chat window, you don't have to go into customize, whatever. You can say just like update the skill to not have this output or to use this instead. And so you can constantly update your skills, based on the outputs that you're getting. Great. And then what you'd want to do, I'd say the next save or the next sort of unlock here is chaining these skills together. What you can do is you can actually go, since I've got brand guidelines together. This is still cooking. I can actually just do a new chat and say, I'm gonna run performance report and my brand guidelines skill. So you can chain skills together like this, and I have pulled up this output already here. The brand guidelines and the performance report. It might do some work here to to try to, understand both those skills and work on them together. But what it did here was that performance report I showed earlier, it applied my brand guidelines skill. Oh, and this one actually is my, an older one, but we see let's go like this and see what it does. Here we are. Let's go like this. Branded weekly report. There we are. Show this guy. There is my nice assembled logo. Can you see this window? I'm not sure it here. I think you can't. It's like this. There we go. Candid. I need that nice Assembled logo. Come on. There we go. Nice. Okay. This is in my brand guidelines now. And now, obviously, there are use cases in which you'd want to actually call those two skills separately. But there are also instances where you wanna just add obviously, every time I run my performance report skill, I just want to say add the brand guidelines skill to the performance report skill. And when you do that every time, just like we showed that that skill creator skill at the top, it had a bunch of sub skills, it will add the brand guidelines skill as a sub skill, so it will run every time. Cool. So, I mean, I think this is just really scratching the surface of what skills can do in Claude. And, really, with this just this type of stuff, what I highly recommend is just getting into the tool and turning some of these knobs yourself. What is great is there are so many templates, existing skill guides, and the chat itself makes it really easy. If you're ever running into any issues, you're definitely going to just ask Claude itself. What better expert? So and I wanna just back in time, I wanna get back into our sort of best practices here as that ends our sort of demo portion. I want to add let's go stop. I just wanna hit on some best practices here, some common gotchas to look out for in general. And feel free to add any q and a in the chat since we're coming up on time. Let's go share. I think one big best practice advice from a security standpoint often overlooked is as you look at skills and look at skill marketplaces, like, get skills from places that you trust. There are a lot of areas where prompt injection can be used to get company and sensitive information and data from your AI tools. So just be mindful about what sources you're getting cloud skills from. That's course, a really good trusted. We verify our skills, but across the board, just make sure you're vetting the sources. that is a really good point, Siddhant. And now, Amit, just convey a couple of key points here. So the thing I wanna really convey about skills, some common gotchas, is, these three things we sort of covered. What can happen, a couple issues that we've seen that you'll run into. One, we've mentioned this quite a bit. If you're not customizing the skill, with enough of your context, it's just not gonna be very helpful. If you're not adding enough of your connectors, your own business data, it's not gonna be useful to you. So just always add the right connectors and context. The other one that is, a common gotcha that folks get into particularly because when you are updating your skill in your chat window, you're often not going back to that customized place, right, where you're actually looking at the whole prompt. So what can happen is your prompt will get really, really long, and it will get very, unwieldy. And if you add, too many instructions in your skill prompt, it it's gonna be, like, contradictions for Claude to choose from, and it's just gonna create, exactly, Paul, too long hallucinations. You wanna be as concise as possible. Not only if you give a big long, instruction prompt, you're you're gonna have more errors. It's also gonna blow up your token usage because Claude just has to think harder. So keep your prompt concise, and there's plenty of guides out there on how to do that. And then finally, another gotcha, you said it and forget it. We build agents whether they're in Claude, whether they're in, ChatGPT, Gemini, Dust. We are tool agnostic for our own internal tooling at Assembled. One thing across all agents as a best practice is these skills wear out fast just as your business changes quickly so to do your AI agents. So just just make sure to always update it, test it initially, but you're gonna wanna update it as you find new and creative ways that people or even you are using your, using your skills. So now just with a little less than ten minutes, ten minutes left, we've got a couple questions. Feel free to add some more questions in the chat too. I've also would really love to hear I'm gonna launch a poll here. This is the start of a session that we are oh, I'm just interrupted by your, your good point, Paul. It's true. 200 to two fifty lines is a good rule of thumb for your prompt. To get back to, what we're talking about here, I'd really like to ask all of you. I'm very curious. What would you like to see next when it comes to AI skills for support operations? I've just launched a survey here. You should be able to see taken over your engagement panel. We just asked open ended, like, what would you like to see in terms of, like, what we can offer for AI skills, any type of skills regardless of Claude for, support operations? We're gonna be doing this type of webinar that's just educational once a month. I think as we all, use AI more in our daily workflows, we're all learning and growing more, and it's something that we're just trying to apply to what we do here at Assembled and how we can can make everyone's lives here a little bit easier. Yes. And don't forget, I see I think we've answered all the q and a. We've got that PII question, schedule adjustments. Great. Oh, and, Kaylee, yes. I can definitely share with you the stop slop skill. I will add that to our, to our Notion doc, which should be updated with, new skills as they come. Another one in the chat. Oh, of course, Kaylee. And, a couple final pieces of housekeeping. I mentioned, definitely, you will get the recording after this. Check your inbox. Don't forget to grab that free pack of clogged skills in the docs tab. We will add more to them. It's a living Notion doc, so we'll be updated. And, again, answer the poll if you can. Let us know what you'd like to hear about. We are super excited to be doing this session once a month at a minimum. And as always, you can learn more or just get in touch with us at Assembled dot com. You can always just email us, hit us up on LinkedIn. That's where Melissa is a, an influencer if you've seen her LinkedIn posts. But, yeah, in general, really happy to have you here, and we do hope you have a great long weekend if you are celebrating Independence Day here in The US. And if not, we just hope you have a great rest of your day. Thanks so much for joining us. Thanks, everyone. Have a great one.