How to Create a Claude Skill (No Code Required)

Creating a Claude skill is how you teach Claude to do a repetitive task your way, once, so it runs from a single command forever after. The good news for non-technical people: you don't write any code, and the reliable method is the opposite of what most people try. This guide walks through how to create a Claude skill from scratch, using the approach we teach after building dozens of our own.

A skill is a reusable set of instructions Claude follows to complete a specific task. Building one well is less about clever prompting and more about capturing a process that already worked.

Watch: the full skill-building walkthrough is in our free Claude Cowork course.

The one rule: do the task first, then save it

The single biggest mistake people make is opening the skill builder and saying "create a skill that does X." Don't. You're asking Claude to imagine a process instead of capturing one that works.

The reliable method is to do the task with Claude first, on real inputs, and only turn it into a skill once the output is right. The skill then becomes a recording of what actually worked, not a guess. Everything below follows from that.

Step 1: Turn on the Skill Creator

Before you start, switch on Anthropic's Skill Creator, the skill used to build every other skill. You'll find it under Customize, then Skills, then browse the available skills, and enable it. It should be one of the first things you turn on.

Step 2: Do the task with Claude, on real inputs

Walk Claude through the job step by step, because you're the expert and you're teaching it your method. Use a real draft, a real spreadsheet, a real folder, not a hypothetical. Prepare anything it will need first, the way a designer gathers assets: your template, your logo, your fonts, an example of the finished result.

Then do the task as a normal conversation. If the first attempt is close but not right, treat it like working with a colleague: ask for the change, iterate, and keep going until the output is what you want.

Step 3: Save it as a skill

Once the result is right, turn it into a skill, using the Skill Creator or the "turn this into a skill" option. Give it a clear name (this is hard to change later, so pick carefully), and bake in any rules you discovered along the way, for example that an image should auto-resize to fit.

The Skill Creator even tests what it builds before saving. For a skill that resizes an image, say, it checks the script against both a short and a long input, so you know it works before you rely on it.

Step 4: Check what got packaged

You don't need to read code, but it helps to open the finished skill and confirm the pieces are there. A skill is a folder with three kinds of content: the instructions (a SKILL.md file in plain markdown), any resources it needs (templates, fonts, logos, saved scripts), and, for more advanced skills, helper agents. We once built a skill that quietly left out a key image, which only surfaced when someone else ran it; a quick look at the contents catches that.

Step 5: Test it in a fresh task

Always test a new skill in a clean task, not the one you built it in, because the building conversation already has all the context loaded. Hand it a realistic input and check the result end to end. When it works, the job that took a careful back-and-forth the first time now runs from one command, and you can feed it a whole batch at once.

Updating and sharing your skill

Skills aren't fixed. To change one, use the same method: do the new version of the task with Claude first, confirm it, then update the skill and save over the old one. To add an option, get Claude to produce it once, verify it, then bake the choice in so the skill asks which you want next time.

Sharing depends on your plan. On Team or Enterprise you can share a skill with your whole organisation, and your future updates reach everyone automatically. As an individual you can download a skill as a file and send it on, though a downloaded copy won't get your later updates. Either way, this is how one person's expertise becomes something the whole team runs. If you want the day-to-day mechanics of running skills, our guide to using Claude skills covers it.

A worked example: a tweet-image skill

Here's the loop on a real build. The goal was a skill that takes a line of text and returns a branded tweet-style image. First we prepared the pieces a designer would: a reference image, the font, the brand colours, and a sample of the finished look. Then we did the task once in conversation: "adapt this image, swap in our logo, use this text, make it 1080x1080, and shrink the font if the text is too long." The first result was close; the font weight was a little thin, so we asked for a heavier one and iterated until it was right.

Only then did we save it, telling the Skill Creator to bake in the auto-resize rule. It ran its own evaluation, smoke-testing the script on a short tweet and a long one to confirm the text fits both, then packaged the skill. We tested it on a fresh batch of fifteen quotes and got fifteen correct images. That is the whole method: do it once properly, save what worked, let the Skill Creator harden it.

Common mistakes when creating skills

A few errors come up again and again, so watch for them.

Asking the builder to invent the skill. The most common one: opening the Skill Creator cold and describing what you want. Do the task first; you're capturing a process, not commissioning one.

Skipping the asset check. We once built a skill that quietly left a key image out of its resources, which only showed up when a teammate tried to run it. Open the finished skill and confirm the pieces are there.

Making it too generic. A skill called "write content" will trigger on everything and do none of it well. Keep skills narrow and specific; that's also what makes their description reliable.

Building in a cluttered conversation. Save the skill in the same context where you did the task, so it knows exactly what worked, then test it in a fresh task with the context cleared.

Questions people ask when building a skill

Do I need to know how to code? No. You do the task in plain language and the Skill Creator writes the files, including any scripts.

Can I edit a skill by hand later? Yes, the SKILL.md is plain markdown, but it's usually easier to ask Claude to update it after you've redone the task.

Will my skill work for a teammate? On Team or Enterprise, a shared skill stays owned by you and your updates reach everyone. A downloaded copy works for them but won't get your later changes.

Build your first skill this week

The whole method is one loop: do the task with Claude, get the output right, save it. Pick a job you already explain to Claude more than once, run it together, and save it. That first skill teaches you the pattern, and after that every repeated task becomes a skill you never explain again.

If you want help turning your team's repetitive work into a real skill library, our hands-on Claude Cowork training takes you from your first task to skills your whole team can run.