--- title: Skills: how I teach Claude to work my way · Domin url: https://www.thedomin.com/post-skills.html --- [← All posts](https://www.thedomin.com/catch-up.html) Post 03 · 19 Oct 2026 · 5 min # Skills: how I teach Claude to work my way by Vojtěch BrandejsPost ## TL;DR - A skill is a folder with a text file in it. Instructions for Claude, sometimes with a script or two. - The best way to make one is to do the task with Claude first, and then turn what worked into a skill. - Downloaded skills are a good start, but over time you should build your own. --- So, skills are the thing I’ve spent the most time on in my Claude setup. I use them to teach Claude how to work: to get a better result out of it, to fix the way it goes about a task, to give it some rules, or just so I don’t have to repeat myself every single time. I keep all of them in one plugin on GitHub, so when I change one I push it there, and every place I use Claude gets the new version with one update. What is a skill? It’s a folder with one main file in it, called SKILL.md. MD is markdown, which is just a text file with a bit of formatting. Inside is written text, rules or instructions for how Claude should do something. The folder can also hold other files the skill points to, or scripts Claude can run, but most of the time it’s just text. ![A drawing of an open folder with a sheet marked SKILL.md sticking out, joined by dashed lines to a scroll marked just text and a small file with a gear on it.](https://www.thedomin.com/assets/doodles/skills-folder.webp) _A folder, a SKILL.md, sometimes a script_ > often just a file on a schedule. That’s worth knowing, because a lot of the “I have agents working for me 24/7” talk is exactly this. Often it’s one markdown file that says what Claude should do, plus a routine that runs it three times a day. Claude Code has this built in: routines run a prompt on a schedule, even when your laptop is closed. Some agent setups are more than that, but a lot of them aren’t. People also mix skills up with CLAUDE.md, which is a markdown file too. CLAUDE.md is what Claude should know in every session, like who you are, how you want answers and the rules of your project. It loads at the start of every session. A skill loads only when it’s needed. At the start Claude usually sees just the name and a short description of each skill, and it reads the whole thing when your request fits. Anthropic’s docs draw the line the same way, saying that a multi-step procedure belongs in a skill and not in CLAUDE.md. So CLAUDE.md is more about you, and skills are more about how the work gets done. ## How I build them The best way I’ve found is to do the work with Claude first and turn it into a skill after. Don’t just tell Claude “create a skill for designing a website.” Design a website the way you want it, go through the whole process together, correct it, explain how it should think about each step, and once you like the result, ask it to turn that into a skill. Anthropic recommends the same thing: finish a task without a skill, notice what you kept explaining, and have Claude write that down. The other thing I do a lot is reverse engineer research. When people study what makes AI writing sound like AI, or what makes an AI voice sound like AI, they end up with a list of the patterns that give it away. You can take that list and turn it around into a skill. My writing skill came from that, and when I was working on voice agents I did the same thing with 33 papers on human speech. > explain the why, skip the caps. But a skill doesn’t have to be a list of rules, and often it shouldn’t be. When I give Claude hard rules, it sticks to them blindly and stops thinking about the task. Teaching it how to think about the work goes much better. Anthropic’s own skill for writing skills says this too: if you catch yourself writing ALWAYS or NEVER in capital letters, take it as a warning sign and explain why the thing matters instead. Strict rules still make sense for steps that break easily, but for most things I’d rather explain. ## What a skill can be The simplest kind is a routine, the thing people like to call an agent. I run one for my email: it goes through the inbox, triages it with my labels and gives me a table of what it did. The same works for your calendar, or any other simple thing you do every day. Then there are workflows. Say you noticed Claude gets design much better when it looks at screenshots of the website, maybe through your Chrome, and you found a process that gives you the result you want. You can turn that into a skill, so the next time you ask it to design something, it takes a screenshot after every change and looks at it on its own. You don’t have to send them anymore. You found something that works, and now it happens every time. And then there are small helpers for things you do all the time. I can send Claude a link to a video and it watches it, transcribes it and writes notes. If I send a recording of a school lecture, it transcribes it, cleans out the noise and puts the notes into my Notion under the right subject. I explained all of that once, and now when I drop in a recording, it reads the skill and does it. It’s just written instructions, no codebase. Skills can also tell Claude to use tools, like a web scraper, and to run other skills. So you can chain small ones into one big workflow: research a topic, scrape what people say about it, write a script, help edit the video, write the description and post it, with each step being its own skill. (I wouldn’t let AI edit your videos yet, but that’s another thing you could reverse engineer, what gives away an AI-edited video.) One of the most installed skills out there, Matt Pocock’s grill-me, is basically one line telling Claude to run another skill. ![A drawing of five boxes joined by dashed arrows: research, scrape, script, edit and post, each with a small icon.](https://www.thedomin.com/assets/doodles/skills-chain.webp) _Small skills chained into one workflow_ You can even have a skill that builds skills. Anthropic has one called skill-creator. I also have one I call Hermes for Claude. It goes through my chats and looks for things I keep repeating, since those are usually the next skill. ## Downloaded skills and your own If your Claude doesn’t feel as smart as people online say it is, this is a big part of why. The people saying it have been using Claude for a while and taught it how they work. It does get better the more you use it, as long as you save what you learn along the way. There are big directories of skills now. On skills.sh the most installed one of all time is find-skills, a skill for finding other skills, with about 3.9 million installs. Near the top you’ll also find grill-me, which makes Claude question you about a plan before it builds anything, and Anthropic’s frontend-design. What I find interesting is how many of the top ones teach Claude a way of working instead of handing it a list of rules. So yes, start with other people’s skills. They help, and some rules really are worth having, like design best practices so you avoid slop. Most of mine started that way too. Of the roughly 60 skills in my plugin, I’d guess 60 to 70% are curated from existing ones. ![A drawing of a shelf: on the left four identical books marked downloaded, on the right mixed books marked your own, with a hand putting one more on the shelf.](https://www.thedomin.com/assets/doodles/skills-library.webp) _Downloaded skills, and the ones you build_ > read it before you install it. Just read a skill before you install it, because it’s instructions your Claude will follow. When Snyk scanned about 4,000 public skills earlier this year, mostly from ClawHub, 13% had a critical security issue. A downloaded skill also helps everyone the same way, so it won’t make Claude work the way you like. That part you build yourself, over time. claude · post 03 explain this like I’m busy A skill is a text file that teaches Claude how you work. Do the task together first, then save what worked. Thanks for reading,![Vojta](https://www.thedomin.com/assets/signature.webp) [← Older · Post 02 The AI bubble: you can’t print a power plant](https://www.thedomin.com/post-bubble.html) [Newer · Post 04 → Making voice agents sound human](https://www.thedomin.com/post-voice.html) Take-home · from this piece2 sheets [**The notes** · a skill is a folder with SKILL.md, mostly just text · CLAUDE.md loads every session, a skill only when it’s needed · do the task with Claude first, then turn it into a skill · explain the why instead of writing ALWAYS or NEVER · small skills chain into one bigger workflow · start with other people’s skills, then build your own Post 03 · 5 min, in 1 page The notes Sheet · 1 page Download PDF ](https://www.thedomin.com/assets/pdf/post-06-the-notes.pdf) [**Before you write a skill** 1. What did you explain to Claude more than once? 2. What did it get wrong until you corrected it? 3. What did you still check by hand at the end? 4. Can you say why each rule matters? Post 03 · ask them yourself Before you write a skill Sheet · ask them yourself Download PDF ](https://www.thedomin.com/assets/pdf/post-06-before-you-write-a-skill.pdf)