How to Build Your Own Claude Skill: A Step-by-Step Guide for Creators

Claude skills run most of my content week now. The first one I wrote never fired once.

I typed the request it was built for. Claude answered in its normal, generic way. The skill sat in its folder, ignored, like a manual nobody opens.

The problem was one line. I had written the description like a label on a jar, and Claude had no idea when to open the jar.

That is what this post fixes. I will show you how to build a Claude skill from a blank folder. The worked example is a real skill I run every week.

Claude skills are small instruction folders that teach Claude how to do one specific job your way. Each skill holds a file called SKILL.md with a name, a description, and the steps Claude should follow.

Claude reads the description first and loads the full instructions only when your request matches. That is why one account can carry dozens of skills without slowing down.

Why I Stopped Re-Explaining My Work to Claude

Skills exist because chat has no memory of how you like things done. Every new chat starts from zero, so you retype the same rules, the same format, and the same banned words. A skill writes those rules down once and applies them every time.

I have been writing online since 2005. I spent 17 years in corporate jobs before that last layoff call.

Alongside all of it, I built this blog, 14 niche sites, Threadeazy, and VoiceLab. Time was always the scarce thing.

When I started using Claude seriously, I noticed a pattern. I was pasting the same 400-word brief into chat three or four times a week.

My voice rules. My no em-dash rule.

My list of words I never want to see. The format for a Threads batch.

That is not a writing problem. It is a storage problem.

Skills solved it. Today my weekly pipeline runs on them.

One writes hooks, and one audits every blog post before it reaches you. Another strips AI patterns from everything I publish.

If you have read my earlier post on Claude Code skills for content creators, you know the story of that pipeline. That post showed what the system does. This one shows how a single skill is actually written, line by line.

What You Need Before You Build Claude Skills

You need three things: Claude Code installed, a plain text editor, and one task you repeat at least weekly. You do not need to know how to code. A skill is a folder and a text file, nothing more.

Here is the short checklist I give people before they start:

• Claude Code installed and signed in. The official Claude Code skills documentation covers the install if you have not done it.
• A text editor. Notepad works. VS Code is nicer because it shows the folder tree.
• One repeated task. Pick something you did at least twice in the last 14 days.
• Two past examples of that task done well. These become your quality bar.

Look, the task choice matters more than anything else in this list.

Beginners pick something huge, like "write my whole newsletter." Pick something narrow instead, like "write five hook options for a Threads post." Narrow skills trigger reliably and give you output you can judge in 30 seconds.

Where Claude skills live on your computer

Claude Code looks for personal skills in one folder: .claude/skills inside your user home directory. On Windows that is usually C:\Users\yourname\.claude\skills. On a Mac it is ~/.claude/skills.

Each skill gets its own subfolder. The subfolder name should match the skill name exactly.

You can also keep skills inside a single project folder, under .claude/skills in that project. I keep my blog-only skills there, so they never fire when I am working on something unrelated.

The Anatomy of a Real Skill: My Hook Writer

A skill has two parts: a short header called frontmatter, and a body of instructions. The header tells Claude when to use the skill. The body tells Claude how to do the job.

Let me break it down with a skill I actually use. My hook-writer skill writes opening lines for LinkedIn, Threads, Instagram, X, and Facebook. Its folder looks like this:

• hook-writer/
• hook-writer/SKILL.md (the main file, 322 lines)
• hook-writer/references/ (supporting files Claude reads only when needed)

The SKILL.md starts with this header between two lines of three dashes:

• name: hook-writer
• description: Write scroll-stopping hooks and re-hooks for LinkedIn, Threads, Instagram, X, and Facebook. Trigger on requests like "write me a hook for", "opening line for", or "hook ideas for". Do NOT use this skill to write the full post body.

That description does three jobs. It says what the skill does, it lists the exact phrases that should trigger it, and it says when not to use it.

The description line is the whole trigger

Here is the thing most tutorials skip. Claude does not read your skill body until it decides the skill is relevant.

At startup, Claude loads only the name and description of every skill you own. When you type a request, it compares your words against those descriptions. If nothing matches, the body never loads, and your 300 lines of careful instructions do nothing.

So the description is not a summary. It is the switch.

What belongs in the body instead

The body holds everything Claude needs after the switch flips. My hook-writer body covers the steps in order: confirm the platform, pick a hook type, write three options, and run a checklist.

Good bodies share a few traits:

• Numbered steps in the order Claude should do them.
• Hard rules written as plain commands, like "Never use em dashes."
• One or two examples of finished output that meets your bar.
• A final checklist Claude runs before it hands the work back.

Keep reference material out of the body. Long word lists, brand guides, and past examples go in a references folder, and the body tells Claude which file to open and when.

How to Create Claude Skills: The 7-Step Build

Building a skill takes about 20 minutes for a narrow task. You create a folder, write the header, write the steps, add examples, and test the trigger. The steps below are the exact order I follow.

Step 1: Write down the job in one sentence

Before you touch a folder, finish this sentence: "When I ask for (this), Claude should (do that)."

My hook writer's sentence read: "When I ask for a hook, give me three options for one platform." If you cannot finish the sentence, the task is still too big.

Step 2: Create the folder

Open your .claude/skills folder and create a new subfolder. Use lowercase letters and hyphens only, like threads-hooks or newsletter-subject-lines.

Inside it, create one empty text file called SKILL.md. The capital letters matter.

Step 3: Write the frontmatter

At the top of SKILL.md, add three dashes on their own line, then the name and description, then three more dashes. The name must match the folder name exactly.

Write the description as three sentences:

1. What it does: "Writes five subject lines for a weekly email newsletter."
2. When to trigger: "Use when I say 'subject lines for', 'email subject', or 'what should I call this email'."
3. When not to: "Do not use for the email body or for blog titles."

If your description contains a colon, wrap the whole description in double quotes. An unquoted colon breaks the header, and the skill fails without any error message.

Step 4: Write the steps

Below the second set of dashes, write the job as numbered steps. Use short commands, one action per step.

Write it like you would brief a sharp assistant on their first day. Say what to check first, what to produce, and what to verify before handing it over.

Step 5: Add your quality bar

Paste one or two past examples you were happy with. Label them clearly as examples of the standard, not templates to copy word for word.

Then add your hard rules. Mine always include no em dashes, no emojis, and a short banned-word list.

Step 6: Move bulky material into references

If your skill needs a long style guide or a list of 50 examples, put that in a file inside a references subfolder. In the body, write one line telling Claude when to read it.

This keeps the main file short. Claude only pays the reading cost when the step actually needs it.

Step 7: Restart and test

Close Claude Code and open it again, so it picks up the new skill. Then run the trigger test in the next section before you trust it with real work.

How to Test Whether a Custom Claude Skill Actually Triggers

A skill that works when you call it by name can still fail when you ask naturally. Test both ways, using at least five different phrasings. If it fires on four out of five natural requests, it is ready.

I learned this the hard way. My first skill worked perfectly when I typed its name with a slash in front. It never fired when I just asked for the thing in plain words.

Here is the trigger test I run on every new skill:

1. Direct call. Type a forward slash and the skill name. This confirms the file loads at all.
2. Exact phrase. Use one of the trigger phrases from your description, word for word.
3. Loose phrase. Ask the way you would on a tired Tuesday, like "give me a few openers for this post."
4. Wrong job. Ask for something the skill should not handle. It should stay silent.
5. Neighbour job. Ask for a task close to it, like a full post instead of a hook. It should step aside.

Score each test as fired or silent. Write the results down, because you will forget them by next week.

What to do when the score is low

If the loose phrase fails, add that exact wording to the description as another trigger. If the wrong job fires the skill, add a clear "Do not use for" sentence.

Then restart and run all five again. Two rounds usually fixes it.

Claude Agent Skills vs Projects vs Saved Prompts

People mix these three up, so here is the plain comparison. A saved prompt is text you paste, and a Project is a workspace with files Claude can see.

A skill is a set of instructions Claude picks up on its own when the job matches.

I use all three. My Claude Projects setup for keeping content in my voice handles tone. Skills handle the repeatable jobs that have steps and checks.

Anthropic calls these Claude agent skills in some of its material, and you will see both names. They mean the same thing. The Anthropic announcement of skills explains the idea from their side.

Common Mistakes That Stop Claude Skills From Working

Most broken skills fail for one of five reasons, and none of them involve code. The fixes take minutes once you know where to look.

Mistake 1: A label instead of a description

"Helps with emails" tells Claude nothing about when to act. This was my first mistake, and it cost me a week of wondering why nothing happened.

Fix it by listing real phrases you type, in quotes, inside the description.

Mistake 2: Folder name and skill name do not match

If the folder is email-subjects and the name line says subject-lines, the skill may not load. Make them identical, character for character.

Mistake 3: Hardcoding rules that live elsewhere

This one cost me more than any other. My content-auditor skill once carried its own copy of my hook checklist.

The real checklist lived in a separate file, and I kept improving it. The copy inside the skill never changed.

For weeks, the auditor graded hooks against an outdated standard, and I did not notice. On 5 September 2026, I fixed it by deleting the copy and telling the skill to read the real file fresh on every run.

The rule: if a rule has a home somewhere else, point to it. Never paste a copy.

Mistake 4: One skill doing five jobs

A skill called "content-helper" that writes posts, emails, and titles will trigger on everything and do each job badly. Split it into three narrow skills.

Mistake 5: Never restarting

Claude Code reads your skill list when it starts. If you edit a description and keep the same session open, you may be testing the old version.

When Building Your Own Skill Is Not Worth It

Building is worth it for jobs you repeat weekly with a clear right answer. It is not worth it for one-off tasks or anything you have not done by hand at least twice.

Here is my honest filter. If I cannot show Claude two examples of the job done well, I am not ready to write the skill.

That is because a skill encodes a standard. With no standard yet, you end up encoding guesses, and Claude follows those guesses with total confidence.

There is also a time cost. My hook writer took four rounds of edits across two weeks before it matched how I actually write hooks.

That is fine for a job I do 10 times a week. It is not fine for a job I do twice a year.

If you would rather start with skills already built

Some readers want the result without the build. That is a fair choice, especially if your evenings are already spoken for.

I packaged the skills I use for blog posts, Threads batches, emails, and YouTube scripts into the Content Creator's Claude Skill Stack. You install them, then use this post to customise the descriptions and rules to fit your own voice. Building one yourself first still helps, because you will understand exactly what you are editing.

How Skills Fit Into a Weekly Content Workflow

One skill saves you a few minutes. Five skills that hand work to each other save you hours. The trick is to build them one at a time, in the order you actually do the work.

My order was simple. Hooks first, because I wrote them daily. Then an audit skill, because I kept missing the same formatting errors.

After that came the anti-AI filter, because every draft needed it. You can read how those pieces connect in my post on the AI content creation workflow that saves 5 hours a week.

Do not build the chain first. Build one link, use it for two weeks, and only then build the next.

Frequently Asked Questions

What are Claude skills and how do they work?

Claude skills are folders containing a SKILL.md file that teaches Claude one specific job. Claude loads only each skill's name and description at startup, then reads the full instructions when your request matches that description. This lets one Claude Code setup carry dozens of skills without filling the context window. My own setup runs more than 18 skills this way.

How do I create a Claude skill without coding?

Create a folder inside .claude/skills in your home directory, then add a text file called SKILL.md. Write a name and a three-sentence description at the top, then numbered steps below. No code is needed, and a narrow skill takes about 20 minutes to build.

Why is my Claude skill not triggering?

The most common cause is a vague description that lacks real trigger phrases. Add the exact words you type, in quotes, to the description, then restart Claude Code. Also check that the folder name matches the name line exactly and that any description containing a colon is wrapped in double quotes.

What is the difference between Claude skills and Claude Projects?

A Claude Project is a workspace with files and instructions that you open deliberately. A skill is a set of steps that Claude picks up on its own when your request matches its description. I use Projects to keep my voice consistent and skills for repeated, step-based jobs like hooks and audits.

Where are Claude skills stored on Windows?

Personal skills live in C:\Users\yourname\.claude\skills, with one subfolder per skill. Project-only skills go in a .claude\skills folder inside that project. Each subfolder needs a file named SKILL.md in capital letters.

How long should a SKILL.md file be?

Keep the main file focused on steps and rules, and move long reference material into a references subfolder. My hook-writer SKILL.md is 322 lines, which is on the long side. Most narrow skills work well at 50 to 150 lines.

Can I use Claude skills on the free Claude plan?

Skill support depends on the Claude product and plan you use, and Anthropic has changed availability more than once. Check the current Claude Code documentation before you build. In my setup, skills run inside Claude Code on a paid plan.

Are Claude agent skills the same as Claude skills?

Yes. Anthropic uses "agent skills" in some of its technical material and "skills" elsewhere. Both refer to the same folder-plus-SKILL.md format described in this post.

How many Claude skills should a beginner build first?

Build one. Use it for two weeks, fix its trigger with the five-phrase test, and only then build a second. I built my hook writer first and my audit skill about a month later.

Can a Claude skill read other files on my computer?

Yes, when it runs inside Claude Code. A skill can tell Claude to open a reference file, a style guide, or a checklist stored elsewhere. My content-auditor reads my live hook checklist on every run instead of keeping a copy.

The One Line That Decides Everything

My first skill failed because of one lazy line. Every skill I have built since works because I now spend more time on that line than on the rest of the file.

Pick one job you repeated this week. Write the one-sentence version, build the folder, and run the five-phrase test before Sunday.

Which task are you going to hand over to your first Claude skill? Tell me in the comments.