Summary
With how many features Anthropic (and pretty much every other AI lab) drops every week, it’s nearly impossible to keep up with everything. And I say that as someone whose job quite literally revolves around testing these features, so trust me, I get it! While an element of monotony has now started to creep in with these launches, one Claude feature that I’ve always thought didn’t get the spotlight it deserves is Skills. This feature basically lets you teach Claude how to handle specific tasks using a reusable set of instructions, resources, and workflows. I’ve been using Skills since the day the feature dropped, and here are a few Claude Skills I think you’re seriously missing out on. Superpowers Claude finally learned to slow down If you’ve been using AI tools for coding for a bit now, you might’ve noticed that they’re often a little too eager to get into the actual coding bit. You throw the tool a task, and more often than not, it instantly starts writing code before taking the time to properly understand what you’re asking for. That can work perfectly fine for smaller tasks, but once you’re building anything even slightly complicated, all those little assumptions the model makes along the way can come back to bite you. This is why I’ve always been a fan of Claude Code’s Plan mode, but Superpowers takes that whole idea much further. And rather than being just one Skill that helps you plan better, Superpowers is an entire collection of Skills that work together, with different ones kicking in at different stages of a project. For instance, its brainstorming Skill activates before Claude writes any code at all. Claude starts by asking questions about your idea, exploring different ways it could approach it, and presenting the design back to you for approval. Once you’ve settled on that, another Skill turns the design into an implementation plan, breaking the project down into tiny tasks with the exact files that need to change and steps for verifying each one. Things get even more interesting once Claude actually starts building. Superpowers can spin up a fresh subagent for each task and have its work reviewed before moving on to the next one, or follow a cheaper workflow where Claude handles the implementation itself and brings in a fresh reviewer at the end. When something inevitably breaks, Superpowers has a systematic debugging Skill that’s designed to make Claude figure out the root cause before it starts throwing random fixes at the problem. There’s even a separate verification Skill whose entire job is basically to stop Claude from declaring victory until it has evidence that what it changed actually works! To see it in action, I used the exact same prompt in two fresh Claude Code sessions: “Build me a simple habit tracker where I can add habits, mark them complete for the day, and see a 7-day streak.” Without Superpowers, Claude immediately decided the project was “simple” and “well-defined,” picked the structure itself, and started building. With Superpowers enabled, though, the first thing Claude did was load the brainstorming Skill. Instead of jumping straight into the code, it stopped to ask where I wanted the habit tracker to live, gave me a few options, and then laid out a quick design before touching anything. So, if you use Claude Code for any sort of coding, I can’t recommend Superpowers enough. frontend-slides I let Claude handle the boring part of presentations The art of putting together a presentation brick by brick is something I don’t really enjoy anymore. While it’s fun to do when you’re not on a deadline, picking a layout, figuring out where everything should go, and making sure every slide actually looks like it belongs in the same deck all take far longer than I ever want them to. This is why I’ve spent a good chunk of time experimenting with different AI tools that promise to take some of that work off my plate. The problem is that most AI-generated presentations tend to look exactly like you’d expect an AI-generated presentation to look. You get the same predictable layouts, a bunch of text stuffed into boxes, and very little of the personality you’d probably add yourself. frontend-slides takes a slightly different approach. Rather than generating a traditional PowerPoint deck, the Skill gets Claude to build your presentation using frontend code. That gives it far more freedom with layouts, typography, animations, and the overall visual direction of the deck, instead of forcing everything into the same handful of slide templates! Codebase to Course Claude builds it, then teaches me what it built I’m not a software developer by profession, which means I primarily use AI tools to build things I probably wouldn’t know how to put together entirely on my own. The problem is that this also makes it incredibly easy to end up with a working project without fully understanding how all the pieces actually fit together. Codebase to Course is one of my favorite Skills for fixing exactly that problem. You point Claude at a repository or a project you’re working on, and instead of simply explaining the codebase to you in one enormous response, the Skill turns the entire thing into an interactive course. The courses it generates are built around the actual project, so you’re learning from the code that’s already there rather than a bunch of generic examples. You get separate modules explaining different parts of the codebase, snippets translated into plain English, visualizations showing how things connect, and even quizzes that check whether you actually understood what you just read! Skill Seekers I basically give Claude its own documentation brain I was once confused about a feature Anthropic had added to Claude Code, so I decided just to see what Claude Code would say when I asked it about it. Interestingly, rather than simply relying on whatever it already knew about the feature, Claude Code handed my question off to something called the claude-code-guide. This is actually one of the built-in subagents Claude Code ships with, and its entire job is answering questions about Claude Code itself. When you ask about a feature, setting, hook, or something else related to Claude Code, the main agent can delegate the question to this smaller, specialized agent. It then digs through Anthropic’s current documentation, finds the information it needs, and passes the answer back to the main Claude session. The reason why I went down the rabbit hole above is because the concept seemed super useful when you translate it to how we use Claude for everything else. Whenever I’m working on a new project that involves tinkering around with new documentation, I usually end up doing the same thing over and over again. I send Claude links, paste snippets from the documentation, and keep correcting it whenever it relies on something outdated. Skill Seekers skips that whole cycle, and basically does what the claude-code-guide subagent does for Claude Code, but for whatever you want Claude to understand. You point it at the source material, and it turns that information into a structured Skill Claude can refer back to later. Skill Seekers supports 18 different source types, including GitHub repositories, local codebases, PDFs, Word documents, Jupyter notebooks, PowerPoint presentations, videos, Notion pages, and even Slack or Discord exports. You don’t even have to stick to one source. You can combine documentation, a GitHub repository, PDFs, videos, and more into a single Skill, and Skill Seekers will attempt to detect conflicts and synthesize the information across them. Things get especially interesting when you’re working with code. Give Skill Seekers a GitHub repository or local codebase, and it can analyze the actual source code, tests, documentation, design patterns, and even things like issues and pull requests. From there, it can pull examples out of tests, generate how-to guides, extract configuration patterns, and build an overview of how the project is actually structured. grill-me Claude, please interrogate me before you build anything Piggybacking off the bit above where I said Claude and AI tools generally have a habit of instantly jumping into work mode, these models also have another slightly annoying habit: they’re often way too willing to fill in the blanks themselves. For instance, if you give Claude a half-baked idea of a tool you’d like to build, instead of stopping to ask what you actually have in mind, it’ll typically make a bunch of assumptions and start building around them. While some of those assumptions can certainly be reasonable, there can also be times when Claude fills in a gap very differently from how you would’ve, and you only realize it once it’s already gone quite far down the wrong path. This is why you’ll always find me advocating the fact that you should spend a little more time in the planning phase before letting Claude loose on a project. While one way to do that is by asking Claude Code to get all the questions it needs answered out of the way before it starts building, grill-me turns that into an actual process. Once invoked, the Skill essentially gets Claude to interrogate your idea one question at a time, digging into vague requirements, decisions you haven’t made yet, edge cases you probably haven’t considered, and any assumptions it would otherwise have to make on your behalf. Ultimately, these are just a few Skills I now can’t picture using Claude Code without. They all do wildly different things, but they’ve become absolutely crucial to my workflow now.