Claude Skills: What They Are and How to Create Your First One

1. Check your plan and turn on code execution

Skills are on the Free, Pro, Max, Team and Enterprise plans, and they need code execution switched on, per Anthropic's Using skills in Claude article (checked 15 September 2026).

On Free, Pro or Max: open Settings, then Capabilities, and turn on Code execution and file creation. Then open Customize in the left sidebar and choose Skills.

On Team or Enterprise: an owner turns on code execution and skills in Organization settings first. Then each member uses Customize, then Skills.

Menu names move between app versions. Some older Anthropic tutorials still say Settings, then Capabilities, then Skills. If a label doesn't match, search the help center for "using skills".

Why: skills run inside Claude's code environment. With code execution off, there's nothing for a skill to run in.

2. Set your safety rule before you add any skill

Anthropic's security guidance says to use skills only from trusted sources: ones you wrote or got from Anthropic. A malicious skill can direct Claude to use tools or run code in ways that don't match what it claims to do, and could leak data. Anthropic says to treat it like installing software.

Write your rule down now: I only enable skills I wrote, Anthropic's, or ones I've read line by line. For anything from GitHub, a marketplace or a friend, run the check below first. It's free, and Claude does the reading. Keep in mind that a skill with access to your connectors can reach what you can reach. Set agent guardrails before any skill touches client data.

Why: the setup mistakes you can undo. A skill that quietly sends your client list somewhere, you can't.

I downloaded this skill and haven't enabled it. Read every file in it, including scripts.
Tell me in plain English:
1. What it says it does, and what its steps tell you to do.
2. Any web address it contacts, any file it reads or writes outside its own folder, and any command it runs.
3. Anything that doesn't match its stated purpose.
4. A verdict: safe to enable, enable after I remove lines X to Y, or don't enable.
Don't run anything while you check.

3. Watch a built-in skill work first

Before building your own, see what a skill feels like. Anthropic's ready-made skills handle Excel, Word, PowerPoint and PDF files, and Claude uses them automatically when relevant. Paste the prompt below with your own numbers.

Why: you'll notice you didn't name any skill. Claude spotted a spreadsheet job and opened the right instructions. Your own skills should work the same way.

Make me a spreadsheet of these [4] weeks of leads by source, with a total row and a simple chart:
[paste your numbers here]

4. Do one job by hand in a normal chat

Pick a job you repeat weekly and do it with Claude in a regular chat. Use the template below. Correct it until the output is what you'd send. Every correction is a rule the skill needs.

The most common mistake is writing the skill first. You end up guessing at rules you've never said out loud. Do one clean run, then save it. If your prompts need work, better AI prompts covers the basics.

Why: Claude already knows general skills like writing and math. What it can't know is your format, your numbers and your taste, and those only come out when you correct a real draft.

Job: [one sentence: what gets made, like "the weekly client update email"]
Input: [what I'll give you each time: notes, a spreadsheet, a transcript]
Format: [sections, length, where the numbers go]
Rules: [what to never do, words to avoid, what needs my approval]
Good example: [paste one past version you liked]

Do the job with this week's input: [paste it]. Then ask me what you got wrong.

5. Save the job as a skill by asking Claude

In the same chat, once the output is right, paste the prompt below. Claude uses Anthropic's skill-creator skill to write the SKILL.md file, organize any files you shared and package it, per the help center's create a skill through conversation guide. If Claude says it can't create skills, check that skill-creator is switched on under Customize, then Skills.

Save the skill Claude makes, then find it under Customize, then Skills, and turn it on. To change it later, use the Edit with Claude option described in How to create custom skills.

On a Mac with Cowork? On Pro, Max and Team, Customize, then Skills, also offers Record your screen. You do the task for about 10 minutes and Claude builds the skill from what it saw, per the same article.

Why: saving from the chat where you did the job keeps every correction you made. Starting a skill from a blank page loses them.

That output is right. Turn what we did in this chat into a skill.
- Name: lowercase words joined by hyphens, like weekly-client-update.
- Description: under 200 characters. Say what it does, then when to use it, using the words I typed in this chat.
- Steps: numbered, including every correction I gave you.
- Put the good example and any template in separate files inside the skill, and say when to open them.
- Where I made a judgment call without explaining it, ask me instead of guessing.

6. Write the description as a trigger

The description is the only part Claude sees before deciding to open a skill. Anthropic's skill authoring best practices give 3 rules. Write it in third person. Say what the skill does, then when to use it. Use the words you'd type.

Anthropic's own skill-creator skill adds a fourth: Claude tends to undertrigger skills, so make the description a little pushy.

Watch the length. Uploading to claude.ai caps the description at 200 characters and the name at 64, per How to create custom skills (checked 15 September 2026). Claude Code allows much more, as step 10 shows. So if you want a skill everywhere, write it to 200.

Names: Anthropic's field rules allow lowercase letters, numbers and hyphens, and reserve the words claude and anthropic. Skip names like helper or tools.

Why: a skill with a good body and a weak description is a skill that never runs.

# Won't fire: vague, no situations
description: Helps with client stuff

# Fires: what it does, then when, in words you type (161 characters)
description: Writes the weekly client update email from rough notes: wins, numbers,
  next steps and what is needed from the client. Use when the user asks for a client update.

7. Copy one of these 3 business skills

Here are 3 skills built for the jobs owners repeat most. Each description fits the 200-character claude.ai limit. Fill in the brackets with your own numbers and rules, since I can't know your targets.

Weekly ad report. It reads numbers you paste or a connector pulls, and it never invents a missing number. The judgment calls behind keep, fix or pause are the ones in a Meta ads audit.

Client update. It turns your rough notes into the same clear email every week.

Content in your voice. It works from real examples of your writing, which beats describing your tone in adjectives.

To use one in claude.ai, paste it into a chat and say "make this into a skill I can save", then turn it on. In Claude Code, save each as SKILL.md in its own folder, as step 10 shows.

Why: starting from a working structure is faster than a blank page. Still run each once by hand before you trust it.

---
name: weekly-ad-report
description: Turns last week's Meta ads numbers into a one-page report with a keep, fix or pause call per campaign. Use when the user asks for the weekly ad report or pastes ad numbers.
---
1. Use only the numbers provided or pulled from the ad account. If one is missing, write "missing" and never estimate it.
2. For each campaign: spend, leads, cost per lead, and the change from the week before.
3. Keep if cost per lead is under [your target]. Fix if it's over target for 1 week. Pause if over target for [2] weeks with under [X] leads.
4. Never change the ad account. Recommendations only.
5. Output: one line on top saying whether I need to act, then a table, then 3 suggested next steps.

---
name: weekly-client-update
description: Writes the weekly client update email from rough notes: wins, numbers, next steps and what is needed from the client. Use when the user asks for a client update.
---
1. Sections in this order: Wins, Numbers, What we're doing next, What we need from you.
2. Under 200 words. Plain language, no jargon.
3. Only numbers from my notes. Flag anything that looks wrong instead of fixing it.
4. Match the tone of example-update.md in this skill folder.
5. End with one clear question for the client. Draft only; I send it.

---
name: content-in-my-voice
description: Writes posts, captions and emails in the owner's voice using saved examples and a banned-word list. Use when the user asks for a post, caption, hook or newsletter.
---
1. Before writing, read voice-examples.md in this skill folder. Copy the sentence length and openings, not the topics.
2. Never use these words: [your banned list].
3. Short sentences. One idea per paragraph. No dashes.
4. Never invent a story, result or number about me or a client. Ask for one.
5. Give 2 versions and say which one you'd post.

8. Test it cold, then fix the file

Open a new chat. Describe the situation without naming the skill, like "here are last week's ad numbers, what should I do?" If the right skill opens, the description works.

If it doesn't open, name the skill in your request. If it works when named, the file is fine and the description is the problem, so go back to step 6. If it still gets the job wrong, the steps are the problem.

When the output is wrong later: ask Claude why, fix it, then say "update the skill with that fix". Correct the file, not the chat.

Why: you'll use skills in fresh chats, so test them there. A correction you only give in chat is gone next week.

9. Pick your next skills, one a week

List the 5 jobs you paste instructions for most often. Turn one into a skill each week, in order of time saved. Anthropic's tutorials suggest starting points like a weekly team update, sales call prep and customer feedback analysis.

Then turn off skills you stopped using under Customize, then Skills. On Team and Enterprise plans you can share a skill with people or your whole organization, and recipients can use it but not edit it.

A skill that runs on a schedule and reports somewhere you already read starts to look like a business operating system. Get the single skill reliable first.

Why: 5 skills that fire every week are worth more than 50 that never do.

10. Power users: skills in Claude Code (terminal)

This step and the next two are for Claude Code users. If you only use claude.ai or Cowork, skip to the troubleshooting table.

In Claude Code a skill is a folder on your computer, not an upload. Per the Claude Code skills docs (checked 15 September 2026), a personal skill lives at ~/.claude/skills/<name>/SKILL.md and works in all your projects. A project skill at .claude/skills/<name>/SKILL.md inside a repo reaches everyone who uses that repo. You can run any skill by typing / and its folder name.

You make a skill by creating the folder and writing the file by hand, as the docs show. The description plus the optional when_to_use field can run to 1,536 characters combined before Claude Code cuts it off. Keep the body under 500 lines and move detail into files you link to.

The surface trap: skills in ~/.claude/skills/ don't load in Cowork, cloud sessions or routines. Those load the skills enabled on your claude.ai account instead.

My own folder shows what goes wrong when nobody checks. My median description is 526 characters, fine for Claude Code and far over the 200-character upload limit. My median SKILL.md is 299 lines, and 69 of my 244 skills run over 500. 54 carry a triggers: field that isn't in the supported field list, so it does nothing. Running claude plugin validate on the folder found one skill with no frontmatter at all.

Why: Claude Code gives you more room and more control, and more ways to break a skill without noticing.

# 1. make the folder
mkdir -p ~/.claude/skills/weekly-client-update

# 2. write SKILL.md inside it (name, description, steps)

# 3. check the whole folder for broken frontmatter
claude plugin validate ~/.claude/skills

# 4. test it
/weekly-client-update

11. Power users: what 244 skills cost in context

Only names and descriptions load at startup. The body loads when a skill fires, reference files load when opened, and a script's code never enters context, only its output, per Anthropic's Agent Skills overview.

Measured on my machine on 8 September 2026, one 33-skill plugin added about 3,327 tokens to every session, around 50 to 220 per description, while each skill's body cost about 650 to 11,700 tokens only when used. The same 116-line procedure cost about 944 tokens on every turn as CLAUDE.md, and 53 tokens as a skill until needed.

The real ceiling: the Claude Code docs say the skill listing gets a budget of 1% of the model's context window. Past it, Claude Code shortens descriptions, starting with the skills you use least. A huge folder doesn't slow you down. It makes your rarest skills stop matching, which looks like a bad description.

Run /skill-doctor (Claude Code 2.1.252 or later) to see what each skill costs and which never get used, then turn those off.

Why: once you pass a few dozen skills, deleting the dead ones does more than writing new ones.

/skill-doctor                              # cost and use per skill, flags never-used ones
/context                                   # the Skills row shows what the listing costs now
claude plugin details <plugin-name>        # token cost of one plugin's skills
claude plugin list                         # what's installed and enabled

12. Power users: downloaded skills and marketplaces

Search for "claude skills github" or "claude skills marketplace" and you'll find thousands. Anthropic publishes its own in the anthropics/skills repository, and the claude.ai directory under Customize lists skills, connectors and plugins in one place.

Everything from step 2 still applies, and more so in Claude Code, where a skill has the same network access as any program on your computer. Read before enabling. Skills can also run shell commands before Claude sees the text. The disableSkillShellExecution setting blocks that for skills from your user, project and plugin folders, per the skills docs. To compare behavior with nothing custom loaded, the docs point to --safe-mode.

Why: a popular repo isn't a security review. Stars measure attention, not what the code does.

// settings.json: stop skills from running shell commands on load
{ "disableSkillShellExecution": true }