#644 Neil: 5 AI Agent Skills I Tested For Research, Code And Design

0:00Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called chat-concierge and it's simplifying car shopping using self-reflection and layered reasoning with live API checks. It doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trading value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. With the American Express Platinum card, I can unlock experiences like no other. Since I'm always booking my next trip, I'd love that I can earn points on travel. Plus, I get a resume benefit, so you know I'm hitting the restaurants everyone's talking about. And you can find out your welcome offer after you apply, which could be as high as 175,000 points. For experiences like no other, there's nothing like Platinum. Learn more at americanexpress.com slash explore dash Platinum terms apply.

1:00Think about your daily AI workflow for a second. You sit down with your coffee, you over-to-fresh chat window, and then you paste that massive five paragraph wall of instructions you use every single day. You explain the exact same context to the machine again and again. That is not real automation. You have not built a system, you have just hired a very forgetful assistant. Yeah, exactly. It feels incredibly repetitive. You know, you are essentially starting from zero every single morning. The cognitive load of just managing the AI becomes totally exhausting. Welcome to today's deep dive. We are looking at a fundamental shift in how we work with these models. We are moving away from manual prompting. Today, we were exploring AI agent skills. Right. These are entirely reusable workflows. You plug them directly into your environments like chat GPT work or cloud code. Once they are installed, you never have to start from zero again. And I have to make a vulnerable admission here. I mean, I still wrestle with prompt drift myself.

2:00Oh, it happens to everyone. It is incredibly common. I end up copy-pasting the exact same instructions every day. Prompt drift is when the AI slowly forgets your rules over time. It gets incredibly tedious. I keep telling myself there has to be a more permanent way to set these systems up. They're absolutely is. And today, we are going to prove it. We are going to build a complete software project entirely from scratch. We will use five specific reusable skills to do it. Five skills. Okay. First, we tackle market research to find a real problem. Then we secure our code environment. Next, we engineer the actual application. After that, we generate a promotional video. And finally, we polish the user interface. We are going from a blank screen to a fully launched product. Okay, let's unpack this. We have to start with the research phase. Because before you build any project, you need to know what people actually care about right now. Right. Building a solution, looking for a problem, is the fastest way to fail. You need to identify real

3:01urgent pain points in the market. And finding those pain points brings up what we call the 15 open tabs problem. We all know that feeling intimately. You are trying to research a hot new tech trend. You have read it open on one monitor. You are scanning hacker news on the other. You are digging through chaotic YouTube comments. Yeah. You are looking at GitHub repositories. You read a literal mountain of text. But at the end of the day, you still do not know what real people actually think. It is exhausting. It is kind of like having a digital ear to the ground. But the ground is just a million people shouting at once. You are trying to find authentic human opinions hidden inside an ocean of internet noise. So how do we find those insights without spending an entire week reading forums? We use a skill called last 30 days. It is specifically designed for chat GPT work. It bypasses all of that manual labor. It autonomously scans recent community discussions across all those major platforms. This is just give a generic summary of

4:05the hype. No, it specifically targets common problems, disagreements and actionable takeaways, citing exact sources. So it filters the noise but keeps the receipts. Exactly. It does not just summarize the general sentiment. It utilizes grounded search. It pulls specific source citations. In a recent test, it actively analyzed 14 massive Reddit threads and 13 hacker news stories simultaneously. So it is actively linking its conclusions back to the source material. Precisely. It separates the slick marketing hype from actual user complaints. During the test, the user asked the agent to research AI agent skills themselves. They wanted to know what developers were actually installing. They wanted to see the common security concerns. And most importantly, they asked it to find opinions that actively disagreed with the general hype. That is such a crucial filter. Finding out what people hate is usually much more valuable than finding out what they like. Absolutely. The model came back with specific actionable data.

5:07It flagged workflow friction and system overload as major user complaints. People are annoyed by how hard it is to string these complex tools together. It pulled out the practical takeaways. Which gives us our project direction. We know the market is frustrated with workflow friction. We know what to build. Which brings us to the second phase. Security. Right. Now that we know what to build, we need to download specialized tools from GitHub to help us build it. But blindly downloading AI tools is a massive risk. It is a huge vulnerability. When you install an AI agent skill, you are giving third party code access to your local machine. A clean readme file is not enough. A high number of GitHub stars is absolutely not a security guarantee. You hear a lot about supply chain attacks these days. A supply chain attack is when hackers infect legitimate software you trust. That is a perfect definition. It is like hiring a trusted caterer for a party, but their ingredient supplier poisoned the flower. The caterer does not know and neither do you. Militious code can hide very deep within seemingly popular repositories.

6:12Yeah. You are giving an AI agent permission to act on your behalf. You need to know exactly what it is allowed to touch. So how do we lock the front door before we start building? We use a tool called Skill Spector. It was developed by Nvidia. And it is a plugin designed specifically for Claude Coork. Skill Spector acts as your digital security guard. It reviews a skills GitHub repo before you even hit the install button. How does it actually work under the hood? Does it just read the code and guess if it is safe? It performs what is called static analysis. You just paste the GitHub repository URL into the chat. The agent scans the entire project structure without actually running the code. It meticulously maps out the dependencies. Let's quickly define dependencies for clarity. Dependencies. Extra external code needed to run a program. Spot on. So it looks at those external files. Instead of just checking the basic instructions, it actively monitors things like network access and credential handling. It ensures your private data is not leaking. It checks if the code is trying to read

7:16your local files. I have to calmly challenge the necessity here. I mean, isn't a clean readme file and a lot of GitHub stars enough of a safety check? Traditional virus scanners look for known bad files. But AI tools are dynamic. They execute scripts on the fly. A script that quietly copies your API keys and sends them to an external server might not look like a traditional virus. You need a tool that understands the intent of the code. The testing process is very revealing. The user asks skill specter to review a complex engineering skill called superpowers. Right. And it did not just return a simple yes or no. It returned a specific caution verdict. That is so interesting. It didn't find anything explicitly malicious. But it flags specific behaviors for a manual review. It highlighted the exact files that interact with the local system. It showed which instructions caused the risk. It acts as an intelligent starting point for your own security review. What exactly triggers it to issue a caution? It flags any hidden scripts,

8:17external dependencies, or anything that interacts with your local files or network. Basically, it flags anything that talks to the outside world. Precisely. Okay. Our environment is vetted. It is secure. It is finally time to actually write the software. We are moving to the foundation phase. This is where things get incredibly powerful. With our environment secured, we can start engineering. Coding agents are amazing at writing code quickly. But relying on single line prompts usually leads to chaotic code bases. Oh, absolutely. If you try to dictate every single step, you end up manually managing the entire process. Planning, testing, and debugging usually require separate manual instructions. It feels like dictating every single keystroke to an intern. You spend more time managing them than you would have spent just writing the code yourself. It should be more like stacking legal blocks of data with a lead engineer. You want to collaborate on architecture, not micromanaged to syntax. That is exactly the friction point. But the superpowers

9:22skill collection for cloud code changes that dynamic entirely. It is not for simple one line requests. It manages the full software development lifecycle. How is it different from just asking a standard chatbot to write a web app? It handles the architecture brainstorming. It manages the actual implementation. It runs the automated testing. It even handles bug fixing autonomously. Wow. It is the difference between dictating to a junior intern versus handing a blueprint to a master builder and walking away. Let's look at the test scenario. They ask cloud code to build a simple task management web app to solve that workflow friction we found earlier. It needed titles, owners, due dates, and statuses. It needed filtering capabilities. And it needed a clean interface for desktop and mobile. What do the agent actually do first? Well, first, it did not write a single line of code. It stopped and reviewed the requirements. Then it created a formal step-by-step implementation plan. That methodical thinking step is so crucial. Large language models need to

10:22break complex tasks down before executing them. It shows genuine architectural planning. After the plan was set, it built the application locally. But here is the fascinating part. It didn't just hand the code back and say it was done. It autonomously spun up a testing environment. Really? Yeah. It tested the main user flows on both simulated desktop and mobile screens. Wait, it tested its own work visually. Yes. And during its own testing, it caught a layout issue. It found the bug. It analyzed the root cause and it fixed it autonomously. All of this happened locally on the machine. Nothing was pushed to the cloud. That autonomous workflow is what matters most. Is this full workflow too heavy for just tweaking a small feature? Yes. If you're making a tiny change, the extensive planning and testing steps become unnecessary overhead. Use it for building houses, not for painting a single wall. That is a great way to think about it. You want to match the tools complexity to the tasks actual scope. Sponsor, this deep dive is supported by our partners. If you are looking to streamline your

11:26own digital workflows, consider exploring tools that integrate seamlessly with your daily tasks. Finding the right platform can completely transform how your team manages data and scales operations. All right, let's jump right back into the deep dive. So we have a working task management app. The code is solid, the test have passed. But a great app sitting on your hard drive is useless if nobody knows about it. We need to explain this app to the world. We need video marketing. Video is essential. But jumping into complex video editing software can take weeks. This brings us to the fourth skill. Bringing the project to life with a tool called Re-motion. Re-motion is another skill for cloud code. It approaches video creation from a completely programmatic angle. It creates videos using React code rather than dragging keyframes. React code is just a popular language for building user interfaces. Meaning it writes software code to generate a video. Exactly. You are not manually scrubbing a timeline or dragging layers around.

12:27If you have ever tried to manually align keyframes in an editing bay, you know the pain. It is incredibly tedious, pixel-pushing work. Re-motion allows the AI agent to build motion graphics mathematically. It builds data visuals and captions entirely through code. Whoa. Imagine generating thousands of custom motion graphics just by tweaking a line of code. You could personalize a promo video for every single user. That completely changes the scale of content creation. It allows for infinite scale and perfect brand consistency. For the test, they asked the agent to create a 15-second motion graphic promo for our new app. They wanted an AI assistant icon in the center. They wanted five feature icons appearing around it. The instructions were very specific. They requested smooth animated lines connecting the icons. They wanted clean typography and a minimal background. And the agent planned the animation sequence first. It established the scene timing. It programmed the mathematical curves for the main transitions. Then it rendered the frames locally. And just like the engineering agent, it reviewed its own work.

13:29It actually found a few layout problems during the review phase. Right, the research text label was physically hitting the main title. The bottom labels were sitting too close to the edge of the screen. The agent caught these specific visual issues. It adjusted the code variables. It fixed the spacing before executing the final render. That self-correction loop is just brilliant. It really is. Why use a coding agent for video instead of a standard AI video generator? Because it gives you perfect repeatability, precise timing, and exact layout control for motion graphics. It gives you exact control over timing, unlike generative video. Code-driven video gives you absolute mathematical control. Which brings us to the final phase. The lodging of our app functions perfectly. The promotional video is done. But there is still one lingering problem when you actually open the application. The user interface. It probably looks like a generic AI generated template. We have all seen that exact look. Oh, absolutely. The buttons are weirdly spaced.

14:31The fonts are boring. The colors are perfectly symmetrical, but somehow lifeless. It is functionally fine. But visually, it feels like a high school project. It lacks a human touch. It lacks taste. And that is exactly what the fifth skill addresses. It is a clawed code plugin called Kraft. Some developers also refer to it as impeccable. We are moving from hard logic to pure aesthetics here. Kraft focuses entirely on visual polish. It improves typography. It refines subtle animations. Crucially, it removes those obvious AI design patterns. It is important to understand how this specific tool is used. You do not use Kraft to build an interface from scratch. You only apply it to a front end that already works. It is exactly like staging a house before a sale. The plumbing works. The walls are up. The foundation is solid. But now you need to bring in the right furniture and lighting. So it actually looks beautiful and inviting to a buyer. That is a perfect analogy.

15:31And Kraft has very specific focused commands to act as that interior designer. You can type slash critique for a full user experience review. You can use slash type set to fix awkward font pairings. You can use slash arrange to balance the layout perfectly. My absolute favorite command from this section is slash anti-slop. It is a brilliant feature. It actively scans the code and mathematically removes those generic clunky layout choices that scream an AI made this. It adjusts the visual hierarchy. Let's define visual hierarchy quickly. Visual hierarchy arranging design elements to show their level of importance. Right. Making sure the user's eye naturally goes to the most critical action button first, rather than getting lost in a sea of equally sized text. During the test, the user fed Kraft, the working front end of our task app. They told it to keep the core structure completely intact. They just wanted improvements on that visual hierarchy, the spacing and the responsive behavior. The plug-in started with a short written critique.

16:32Then it applied the visual improvements directly to the React code. It is that simple. Does running these commands risk breaking the core functionality we just built? No, you instruct it to keep all existing features working and purely focus on visual hierarchy and spacing. It changes the paint, but leaves the plumbing alone. Exactly. We have covered a massive amount of ground today. We walked through an entire software build. Let's synthesize all of this. What is the overarching theme here? AI agent skills act as a crucial bridge. They move AI from being a neat novelty, a chap bought you bounce ideas off of, to being a reliable, professional tool that executes complex, multi-step workflows. But the secret is not hoarding plugins. You do not need to install every single skill you can find on GitHub. That just creates more management overhead for yourself. The real power lies in targeted application. You need to pick the one skill that maps directly to a workflow you already repeat constantly. Exactly. If you spend hours reading community trends,

17:33install less 30 days. If you constantly worry about code vulnerabilities, use skill specter. If you manage complex engineering projects, leverage superpowers. If you meet precise marketing animations, use remotion. If your interfaces look generic, apply craft. It is about matching the exact tool to the exact friction point in your day. This entire workflow raises a massive question. We are looking at a future that is arriving very quickly. If AI agents can now autonomously plan software architectures, if they can secure the code, if they can build it, debug it, animate the marketing and polish the design, all based on real-time market data, what happens to the concept of a startup? That is a staggering thought. When the friction of creating high-quality software drops to near zero, does the value shift entirely away from the code itself? Does it shift entirely to the taste of the person guiding the agent? We want you to really think about that. But for now, we want you to try this practically.

18:34Pick just one repetitive task you do this week. Just one. Install a single AI agent skill to handle it. See if a genuinely saves you time. Stop explaining your exact workflow from scratch every single morning. It is time to fire that forgetful assistant. Thanks for joining us on this deep dive. We will catch you on the next one. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called chat-concierge and it's simplifying car shopping, using self-reflection and layered reasoning with live API checks. It doesn't just help buyers find a car they love. It helps schedule a test drive. Get pre-approved for financing and estimate trading value. Advanced, intuitive and deployed. That's how they stack. That's technology at Capital One.