AI Agent Skills and MCP: Better Together | apidays India 2026

Thank you everyone for being here. Hello everyone. My name is Rishab Gupta and welcome to API Days 2026 and I'll be presenting on the topic AI skills and MCP better together. So my name is Rishab Gupta. I work as a growth engineer two at RT Camp where my responsibilities are to help marketing and sales team with my AI engineering and automation efforts. I like to go as tech and product speaker at multiple conferences. So in past I've given talks at Cloud, Get Together, City JS, React JS and Notion like meetups. I also like to tinker around AI automations. This is kind of my hobby wherein I like to you know, talk in form of video content or you know, write blogs about AI automations so that people can understand how they can you know, avoid the manual work they are doing on the day-to-day basis. Okay. So let's get started. So the course of the talk will be that first I'll talk about what is MCP, then I'll talk about what are skills, then we'll talk about how we can use these two together to get the best results for ourselves and then finally I'll talk talk about a fun little demo which I built using Swiggy MCP and on top of that I've built a skill that will help me do quick orders. So MCP stands for model context protocol and it was introduced by Anthropic in 2024 and it is a standard protocol by which AI clients like Cloud, ChatGPT connect to external tools, database and systems. Since MCP usually needs a runtime, it runs via node via NPX or you're typically running an MCP server yourself. Now this reminds me of a quick How many of you have used connectors with Claude? Lot of you. Anyone would like to share that experience on what connector did they use or how was your experience with connectors in general? Anyone? Okay. So okay, so you might have used connectors. So for those of you who don't know about connectors, these are some ready-made integrations that let AI access tools without you managing any setup or back-end. So you might have heard about Gmail and Claude, right? So you can connect your Gmail directly to Claude and Claude can basically answer questions from you when you ask about Gmail, right? So let's say you ask it to summarize emails, right? You ask it to like things like you want to know the top 10 business emails that came to you last week. So it will be easily able to answer that and connectors are basically running on an MCV protocol. So behind the scenes, what it does is that it gives a tool call to, you know, the MCV protocol and it gets the result. Like get email tool call will be made and it gets the result to Claude and Claude will basically summarize the emails or, you know, get the top 10 emails for you. Uh 1 second. Now what are agent skills? Agent skills were also introduced by Anthropic in 2025 and these are some reusable capabilities for AI agents. So you might have you know, on your day-to-day basis, you might have experienced this thing that you want to send a message to your colleagues or write a mail to your boss, right? So what you do for that is that plan a rough sketch for the message and then you put it to Claude. You ask Claude that, "Hey, can you fix this message? Make this message better, you know, polish it up." All those things, right? But when you get the result, you feel very hesitant to share that direct message from, you know, Claude to uh your colleagues or boss, right? And the reason being that AI uses uh uses its own skill that it has built, and it writes sentences with EM dashes. It writes sentences that have, you know, text that are not uh like very familiar to us, like words like additionally, specifically, uh effectively, right? These words we don't generally use in our sentences, but AI obviously uses it. And when you share it with people uh you know, they automatically catch that it's written with AI, right? In order to solve that problem, what we can do is that we can have a skill built on top of that, right? We can write our own skill with a tool name uh with like a name, description, and we can write certain instructions. And in those instructions, you can personalize it saying that, "Hey, don't use EM dashes in the sentences." Or you can say that make the sentences polite, uh make it concise, right? When you write all these things, and the next time when you come to your uh agent and use that skill file, it will write it it will write a message that is much more personalized. You will feel that it's coming from you rather than it's coming from AI. Correct? So, let me show you an example in Claude. So, this is the Claude interface, and I'm in uh settings and inside skills. So, there's a skill by the name of humanizer. So, what it simply does is that you can remove AI-generated traces from that text, right? So, you will see that it's a humanize skill. It is first of all completely written in markdown format, right? So, skills are basically written in markdown format, and when you have to invoke a skill, right? You first in the description tell it that this is the name of the skill. Then you write a small description so that AI, when it's trying to understand multiple skills, it is able to identify which skill do you want, right? It will see that the humanizer skill needs to be invoked when you are saying sentences like fix this message or polish this message. Right? So, inside the skill, what you will see is that there are multiple things. The task, core rules at a glance, personality and soul, how to add tone, right? So, it's a very long written skill and this I've downloaded from skills.sh, which is a website by Vercel in which there are a lot of skills available, not only for humanizing sentences, there are also skills available for, you know, writing front-end design skills, having back-end, you know, design skills and everything. Right? So, this is a skill part and you might ask that how do you add a skill in Claude? So, there are four different ways in which you can add skills in Claude. I think in this version there are only three, but there is a new new update as well in which you can record a skill, which I'll talk about. Right? So, first thing is you can create with Claude. This is generally not recommended in my opinion because when you ask AI to only write the skill, then it's very difficult for you to personalize it according to you. Right? This can be used after you write some skill instructions. So, the second part is important wherein you can write your own skill instructions, you can write personalized messages or you can write out things that you actually need in a skill. Right? What are the instructions it needs to follow, right? You might have some company guidelines as well. You can mention it in a skill file and then after that you can give it to Claude to polish it up. Right? After that you will get the best results for a skill. Then the third option is you can upload a skill. So, uploading a skill is through skills.sh, right? The website that I told you. It has a lot of verified skills available that expert people have written, which you can directly use in Claude. Right? And the other way is you can copy your markdown from GitHub or somewhere and you can use it directly. Uh the reason I would not recommend highly recommend the copying and pasting part is because there are some prompt injection uh you know things associated to skills as well, right? If you blindly copy and paste a skill from somewhere, there could be a possibility that there are some prompt injection patterns that are listed into it and I would not recommend that. So, always download or upload a skill into your Claude using uh you know, a recognized platform. Right? And the fourth thing that I was talking about, right? So, in this version it's not available, but now you can record your screen and skill will be built automatically by Claude. So, you can just simply record what you're doing. Just uh make sure that you don't have any sensitive data around. Right? You can record your screen. Claude will automatically see what all steps you have taken to complete a task and then it will build a skill according to that. Right? And after that you can use that skill whenever you need to repeat that task again uh when you are working on the same task. Right? So, that is there. You can browse skills as well that are available in Claude. Okay. So, as I said, there are like humanizer is not the only skill that is available. So, in case you feel that you are a like since you are a front-end developer, you can also use front-end design skills. You don't want that AI should be generating some you know, pink and blue gradient on your website. It is using that same font again on your website. So, you can definitely use front-end design skills. It will give you a much more mature front-end design that you can use and share it with your clients. Right? Now, we talked about agent skills. We talked about MCP, but it's important for us to understand what are the pitfalls of MCP. Where MCP falls apart. Right? So, MCP access often led to hallucinations or struggled due to lack of clear tool chaining instructions. Without defined step-by-step SOPs, agents couldn't sequence action effectively. So, to give you a simple example for this as well, let's say you are at the first day of your job. You are asked to generate a report and publish it on Google Docs. This is a task that you know, your boss has assigned you and you are you have to do it 5 days a week. Correct? So, and you have to do it efficiently as well, right? Because it's time uh where everyone can use AI. Uh so, you decide that you can use Claude for it. Right? You go to Claude. You tell Claude that, "Hey, can you generate me a report on a specific topic?" And it will generate the report for you. And what you can do next is that polish it up, right? You often ask that you need to alter these things. It will alter it for you. And once you have the, you know, proper documentation present and the report present, you can simply connect Google Docs MCP with your Claude and ask it to publish it. And it will easily be able to do it because your MCP tool like MCP already has the tools tool description to publish the report and it will be easily able to do it. But the second day, what uh the problem is that you are given a lot of tasks. Right? For that, you need to install multiple MCP tools. Like you installed Notion MCP, uh GitHub MCP, Jira, Asana, and lot more MCP. At this point, you have almost 10 MCP tools connected to your AI client. Third day, when you will go to generating a report, right? Because you had this task of every day generating a report, what happens is that uh you generate a report. The report uh comes to you, right? Then when you ask it to publish it, you push in the same prompt, like publish it. But instead of publishing it to Google Docs, what it does is it publishes it to Notion pages. And why this problem happened was because we were, you know, prompting it very vaguely because at the initial day we felt that just saying that publish it would directly publish it to Google Docs because at that time only one MCP tool was connected. But when you have 10 MCP tools connected, you need to be you know, talking to your AI in a very structured way. You need to define step-by-step SOPs so that agents can work effectively. Now, what if we could combine MCP and agent skills together? What if we could, you know, write these uh step-by-step instructions to our agents and then agents can take the action effectively? So, MCP and skills can work together. So, skills instantly trigger with preloaded context reducing the latency in decision-making. The same context is reused across multiple skills. It ensures that uniform output and consistent behavior from an AI agent. Right? So, if I go back to the same example, the other way of doing solving the same problem could be that I could have created a skill like a report generator skill. So, in that skill I could have given it a name, description, and after that I could have written step-by-step instructions. Right? The first step would be that you need to write a report. So, of course, as a prompt I would say that this is a report and I would invoke it via skill. Right? I would give it a topic, and in the instruction I would write that make it within 200 to 250 words, you know, write in a proper formal way, don't use em dashes, and you know, all these instructions. So, it saves me time as well. So, the next time when I prompt my AI, it would automatically know that these are the instructions I have to follow, and after that it would give me a much more polished and richer you know, report so that I don't have to repeat myself again and again to AI to, you know, properly fix those reports. Right? The second instruction that I would give in my skill is that you need to publish the report via the Google Docs. And I would mention the tool call as well. Because if I'm very clear with which tool I want to use and which MCP I want to use, it will easily be able to do the task efficiently, right? And the second major advantage of this pattern is using MCP and skills together is that it is cost-effective as well. Because if you think deeply that why do we use, you know, smarter models? Because they are good at decision-making, right? But in case in cases where you already have predefined instructions, like it's already written in a markdown file. So agents like Hyku 4.5 with probably medium or high efforts can easily do the report generation task. They can simply read through the instructions. They can follow, you know, step-by-step instructions of, you know, generating the report. And it can easily identify tool calls as well, right? It can see that it needs to write the report to Google Docs. It will easily follow both the steps and get the task done. So at this point you're saving cost as well when you don't have to, you know, use smarter models and decision-making models to get the task done. Now, time for a quick demo. Now, how many of you use Swiggy Instamart? Lot of you. Yeah, almost all of you use Swiggy Instamart. And how many of you know that Swiggy has its own MCP? Okay, very few. So, okay, I I'll show you a demo wherein I used Swiggy MCP with my skills built on top of it so that I can do quick orders, right? So something like I when I have repeated repeated things, like when I have to order something on, say, weekly basis or monthly basis, I can create a quick orders list. Okay. So I have this Swiggy quick orders, uh you know, skill that I've built. And if I click on this, what you will see is that it has some predefined grocery list on Swiggy Instamart, right? Each list is curated for the specific occasion. I just say the name and the list gets built. So, what you will see is that I have predefined list. What you would do with Claude is that I would tell Claude that, "Hey, can you order all these items for me on a weekly basis or monthly basis?" But, do I actually need to prompt Claude with, uh you know, these things every day? Since this task is repeating, I can just simply tell my Claude that, "Hey, I need to order daily need items." Right? It will automatically understand that the user is asking for, uh let's say, tender coconut, paneer, eggs, and coconut cookies. Right? It will automatically be able to understand that part. Right? Let's say I'm sick someday, right? I know all these things are required on a sick day, so it will automatically understand that these are the things that are there. Right? Guest in 20 minutes, movie night, right? Hydration list. Now, the second most obvious question is that Claude may ask that, "Where do you want to get it delivered?" Like, right? Since most of the times I'm uh there are multiple addresses saved, so I'm specifically looking at a uh home to address, right? So, it can default to that address, and it will also give me a budget. Right? I've defined a budget like 500 is the max that you can use so that it never exceeds and goes beyond, you know, 500 uh rupees on ordering anything. Right? So, after that, what you will see is there is a workflow, right? I saw the Swiggy MCP configuration. I saw what all tools are available, and in each step I'm defining which tool call needs to be made. So, first step is to identify the list. It will easily be able to identify the list uh from the uh predefined uh list, and then it gets the address. It calls get address via the Swiggy MCP. From the returned list, find the Find the address whose label matches home too. So, uh it's like behind the scenes it will have the, you know, label matching for home too, like whatever address is there. And then it stores the address ID for subsequent calls. So, the address ID is also there. So, whenever you are building a skill on top of MCP, always make sure first you do the manual efforts of always doing going through it, right? You go through the tool tool calls whenever the task is getting done. You go through the instructions that are written, right? After that, you build a skill so that it's very effective. Correct? Then the next step is that it checks the existing cart. So, I don't want that, you know, let's say I uh added something previously and it gets cleared uh the next time, right? If If it is adding something, it should always ask me that if uh if your cart is empty, then should I add on top of it? Or if it's clear, then uh I should create a new one, right? Then uh there are multiple tool calls like clear cart, update cart. And then it searches and matches products, right? So, there are, you know, products available multiple places like sometimes the order is also not available, right? So, product is missing or something like that, right? Then after that, it creates an order summary. So, you'll see that the whole order summary is present. And finally, the most interesting part what I find is that you can find discounted alternatives as well. Let's say you were ordering some item A. It was present at price X, but there is an item B present for it at price Y, which is of course lesser than X, right? So, it often makes sense that we can order the second item instead of the first one, right? So, this is what uh is there in the discounted alternative. So, you don't have to go manually search for these uh you know, secondary item. It will do it automatically for you, right? It will ask you, "Would you like to replace these items?" And then it does a budget check. Then it adds to cart and then it does a checkout confirmation. So, for payments it is mostly cash on delivery. If you use Swiggy Dine Out or Swiggy normal MCP, you'll get the option for you know, card payments. In in case of Swiggy Instamart it was not available. So, yeah. And finally the trigger phrases, I would say so whenever it sees words like order my daily needs, daily stuff, daily orders, it will automatically know which skill I need to invoke to get the task done, right? So, that is there and these are some trigger phrases that I've written for invoking the skin. All right. That's all from my side. Thank you everyone for being patient listeners and I appreciate API Days for having me here as a speaker. For those of you who are curious to understand more about AI tech and automations in general, feel free to scan the QR codes and you'll be able to connect with me. Thank you everyone.