Key Takeaways
- A Claude skill is a folder with a SKILL.md file in it: a name and a description at the top, then the instructions Claude follows.
- The description decides whether the skill gets used, because Claude reads it first and only opens the rest when a task matches.
- You can write a skill yourself, have Claude draft one with Anthropic's skill creator, or record your screen in Cowork and let Claude propose one.
- Test it on three tasks from your own work before you share it, and check that it doesn't fire on tasks it shouldn't touch.
To create a Claude skill, make a folder named for the skill and put a SKILL.md file in it.
Start the file with a name and a description that says what the skill does and when to use it. Then write the steps Claude should follow.
Zip the folder and upload it in Claude under Customize > Skills. Then try a few prompts that should trigger it, and refine the description until Claude picks it up at the right moments.
What You're Building, and When a Task Deserves a Skill
A skill is a folder Claude can load when a task calls for it. At minimum, it holds one file, SKILL.md, with a name and a description at the top and your instructions below.
Claude doesn't read every skill in full on every task. Anthropic's developer docs explain that only the name and description load at the start, and Claude opens SKILL.md once the skill becomes relevant, so you can keep many skills on without slowing anything down.
The Exchange's Claude skills directory covers what skills are, how to install ones other people built, and 87 that are ready to use. Check it before you build, since the skill you need may already exist. Most are free once installed.
A task deserves its own skill when you catch yourself pasting the same context into chat again and again. Anthropic's docs suggest a simple test: do the task once with Claude, notice what you had to explain, and turn that into the skill.
Good first skills share three traits, according to Claude's Help Center. They solve one repeatable task, they follow clear instructions, and they're focused on a single workflow.
The Easiest Way: Describe It in a Claude Chat
I don't write my skills by hand. I open a Claude chat, describe the task I want a skill for, and ask Claude to build the skill and guide me through it.
Claude asks what it needs to know and drafts the files, then tells me what to check before I save it. I don't open another window or look anything up, and it feels like explaining a task to a colleague.
A first message can be as plain as this:
I want a skill that turns my sales call transcripts into CRM recaps. Build it with me and guide me through each step.
Anthropic's developer docs back this up: Claude handles the skill format on its own, with no special prompt needed.
How to Create a Claude Skill, Step by Step
Here are the detailed instructions, with the logic behind each step, for anyone who wants to know how a skill works underneath. They're also what Claude does for you when you ask it in chat.
Step 1: Pick One Task and Write Down Three Examples
Choose a task you do every week, such as turning call notes into a CRM recap. Then write down three requests you'd give Claude for it, taken from last week's work.
Those three become your tests later, since Anthropic's docs recommend building the tests before the instructions.
Step 2: Create the Folder
Make a folder and name it after the skill, in lowercase with hyphens, such as writing-call-recaps. The Help Center asks for the folder name to match the skill's name, and a mismatch is one of the listed reasons an upload fails.
Step 3: Write the Name and Description
Create a file called SKILL.md in the folder. It starts with a short block of settings between two lines of three dashes, called the frontmatter.
Both fields have limits:
| Field | Limit | Write it as |
|---|---|---|
| name | Up to 64 characters: lowercase letters, numbers, and hyphens only, with no "anthropic" or "claude" | Lowercase with hyphens, matching the folder |
| description | Up to 1,024 characters | What the skill does, then when to use it |
The description is the part Claude reads to decide whether to use the skill. Say what it does and when to use it, in the third person, as the developer docs advise.
"Turns a call transcript into a CRM recap" works, while "I can help with calls" doesn't.
Step 4: Write the Instructions
Below the frontmatter, write the steps the way you'd brief a capable new hire. Skip anything Claude already knows, such as what a CRM is, and spend the words on what's specific to your team.
Include one example of a good result. The Help Center lists examples as one of the traits of the best skills, because they show Claude what success looks like.
Keep the file short: the developer docs recommend staying under 500 lines and moving longer material into separate files.
Step 5: Add References and Scripts Only If You Need Them
A template, a style guide, or a list of rules can live in its own file, such as references/recap-template.md. Mention it in SKILL.md so Claude knows when to open it.
Keep every reference one level away from SKILL.md. The docs warn that Claude may only skim a file reached through another file.
Scripts suit steps that must come out the same every time, like checking a format. Leave them out of a first skill, because instructions alone are enough for most tasks.
Step 6: Package and Upload It
Zip the folder so the folder itself sits at the root of the ZIP, with your files inside it. Then, in Claude, go to Customize > Skills, click "+", choose "+ Create skill", then "Upload a skill", and pick the ZIP.
Skills need code execution turned on, and the Help Center lists them on Free, Pro, Max, Ultra, Team, and Enterprise plans.
Using Anthropic's Skill Creator and Other Shortcuts
You don't have to write the files yourself. Claude offers two shortcuts, and both end with the same folder you'd build by hand.
| Route | How it works | Best for |
|---|---|---|
| Write it yourself | You create the folder and SKILL.md, as in the steps above | Learning the format, or full control over every line |
| Skill creator, in chat | Turn on Anthropic's skill-creator skill, describe your process, and Claude asks questions, writes SKILL.md, and packages a ZIP | A process you can explain in words |
| Record a skill, in Cowork | Record your screen and voice while you do the task, and Claude proposes a skill for you to review | A process that's easier to show than to describe |
The skill creator route is the one most beginners should start with. The Help Center guide has you open a new chat with a line like "I want to create a skill for quarterly business reviews."
Then you answer Claude's questions, with enough detail that someone capable but new could follow it.
Anthropic's developer docs add a useful method on top. Use one Claude chat to write the skill, then test it in a fresh chat with only the skill loaded, and bring back whatever the fresh chat got wrong.
Recording has limits to know before you plan around it. According to the Help Center, it runs in Cowork in Claude for Mac on Pro, Max, Ultra, and Team plans, and a recording lasts about 10 minutes.
It isn't available in chat, on Windows, or on Free and Enterprise plans.
Whichever route you take, read the SKILL.md before you save it. A draft can sound right and still miss the one rule your team always follows.
A Worked Example: A Call Recap Skill
Here's a first skill a sales or customer success team could build in an afternoon. It turns a call transcript into a recap for the CRM and a follow-up email draft, in the team's format every time.
The folder holds two files:
writing-call-recaps/
├── SKILL.md
└── references/
└── recap-template.mdThe SKILL.md file:
---
name: writing-call-recaps
description: Turns a sales call transcript into a CRM recap and a follow-up email draft in the team's format. Use when someone shares a call transcript or asks for call notes.
---
# Writing Call Recaps
1. Read the whole transcript before writing anything.
2. Fill in references/recap-template.md, keeping its headings in the same order.
3. Under Next steps, list only what someone agreed to on the call, with an owner and a date for each.
4. Quote the buyer's own words for their main problem. Don't paraphrase it.
5. Draft the follow-up email under 120 words. It confirms the next step and nothing else.
Rules:
- If no date was agreed, write "No date agreed." Never invent one.
- Leave out pricing unless the buyer raised it first.
- Return the recap and the email as two separate blocks, so each can be pasted on its own.The template it points to sets the headings once, so every recap reads the same:
# Call Recap: [Company], [Date]
## Attendees
## Their main problem, in their words
## What we showed
## Objections raised
## Next steps (owner, date)The description is 162 characters, well under the 1,024-character limit, and after saying what the skill does it names the two moments to fire: when someone shares a transcript, or asks for call notes.
The rules carry the team's judgment. "Never invent a date" and "leave out pricing" are the kind of thing a new rep learns in their first month, and the skill makes Claude follow them from the first recap.
Test, Install, and Share Your Skill
Test the skill on the three requests you wrote in Step 1, plus a couple it should ignore. Anthropic's developer docs recommend at least three test cases, and a request that shouldn't trigger the skill tells you as much as one that should.
| Test | What to try | It passes when |
|---|---|---|
| It fires | Each of your three requests from Step 1, such as pasting a call transcript | Claude's thinking shows "Using writing-call-recaps," and the recap follows the template |
| It stays quiet | An unrelated request, such as summarizing a blog post | The skill isn't used |
| It follows the rules | A transcript where no date was agreed | The recap says "No date agreed" |
When Claude doesn't pick the skill up, the description is usually the cause, and the Help Center's troubleshooting list starts there.
Check the skill is switched on in Customize > Skills, check the description says when to use it, and try naming the skill in your request.
If your team uses more than one Claude model, run the tests on each. The developer docs note that a skill written with a powerful model in mind may need more detail for a faster one.
Custom skills you upload are private to your account until you share them. On Team and Enterprise plans, an owner has to allow sharing in Organization settings > Skills first, and can also provision a skill so it appears for everyone.
The same folder works beyond the Claude apps. Skills also run in beta in Claude Code, and through the API with the code execution tool, and the Help Center points to the open Agent Skills standard at agentskills.io for tools that adopt it.
Before you share a skill, check it for secrets. The Help Center's security advice is short: never hardcode API keys or passwords, and review any skill you download before you turn it on.