Key takeaways:
- Every Claude skill lives in a folder around one SKILL.md file. Claude keeps only each skill's name and description in view (about 100 tokens) and loads the instructions when your request matches.
- On Free, Pro, Max and Ultra you switch on code execution in Settings > Capabilities and manage skills in Customize > Skills. On Team and Enterprise, an owner controls both.
- The proof that a skill fired is a line such as "Reading [skill name]" in Claude's thinking. No line means the request missed the description.
- Most "broken" skills turn out to be a toggle, an admin setting or a phrasing mismatch, and the troubleshooting table below covers each one.
Most advice about Claude skills assumes you live in a terminal: clone the repository, copy the folder, run the command. None of that is required. Everything in this Claude skills tutorial happens in Claude's own interface, with menus you can click and requests you type in plain English.
That matters, because the people who gain most from skills are rarely developers. The content system I run at busyless, where I work as a fractional Head of Content, has shipped over a hundred articles on seven skills, and each of those skills is a folder of plain-language instructions.
Claude skills for non-developers are a reading and writing job, not a coding one. This guide covers what a skill is, how to turn one on, how to check it fired and what to do when it doesn't.
What is a Claude skill, in plain terms
A Claude skill is a folder that teaches Claude one repeatable job. Inside it sits a file called SKILL.md, which opens with a short header holding the skill's name and a description of when to use it. Below the header are the instructions, written in ordinary Markdown. Some skills also carry reference files (a style guide, a template, an approved example) and small scripts Claude can run, which you never have to open as a user.
What makes the format practical is how little of it Claude reads up front. Anthropic's Agent Skills overview puts the startup cost at about 100 tokens per skill, because only the name and description load. The instructions, usually under 5,000 tokens, enter the conversation only when a request matches, and bundled files load only if the task needs them.
If you came here asking what is a Claude skill, the short answer is a folder of instructions with a one-line trigger. Two consequences follow for anyone who isn't technical. The description is the trigger, so one sentence you can read and edit decides when the skill wakes up. And because a skill is text, you can read every word of it before you trust it.
Writing a team's standards down as files a machine can follow is the core idea of content engineering, and a skill is the smallest working unit of it.
Skills are easy to confuse with two older Claude features, and Anthropic's help center draws the line between all three. The table sums up when each one loads.
Feature | When it loads | What it's good for |
|---|---|---|
Custom instructions | Every conversation | General preferences, like "keep explanations concise" |
Projects | Every chat inside that project | Context that builds up over weeks: a launch, a research effort |
Skills | Only when a request matches the description, in any chat | Repeatable procedures: a brief format, editing rules, a report template |
The practical rule is to keep anything that applies to everything in custom instructions, and to turn anything you explain again and again for one kind of task into a skill. Skills also follow the open Agent Skills standard published at agentskills.io, so a skill written for Claude isn't locked to Claude.
Check your plan and one setting first
Anthropic's help center lists skills on every plan: Free, Pro, Max, Ultra, Team and Enterprise. There's one condition that catches most first-timers, because skills run inside Claude's sandboxed computing environment. Code execution has to be switched on, or the skills section either won't appear or shows every skill grayed out.
Where you flip that switch depends on the plan. On Free, Pro, Max and Ultra, open Settings > Capabilities and turn on "Code execution and file creation", which the updated Settings layout calls "Create and edit files".
On Team plans, skills are on by default at the organization level. On Enterprise, an owner has to enable both "Code execution and file creation" and "Skills" in Organization settings > Plugins & skills, on the Policy tab, before anyone can use them.
One discrepancy is worth knowing before you spend an hour on it. The help center says Free users can upload their own skills, while Anthropic's developer overview lists custom-skill uploads on Pro and above. If you're on Free and the upload option is missing, that's the likely reason, and Anthropic's built-in skills still work either way.
How to use Claude skills: turn on your first one
The fastest first session uses a skill somebody else already tested, so you learn the mechanics before you write anything. These are the steps from Anthropic's guide to using skills, in the order you'll click through them:
- Confirm code execution is on in Settings > Capabilities (or ask your owner on Team and Enterprise).
- Open Customize > Skills from the left sidebar.
- Switch on one of Anthropic's built-in skills, which cover Excel, Word, PowerPoint and PDF files.
- For more choice, click the "+" button, then "Browse skills", and click "Add" on a skill in the directory.
- Open the skill and read its description before relying on it, because that sentence decides when it fires.
- Start a new chat and ask for a real job from your week, phrased the way you'd brief a colleague.
The built-in document skills need no further setup. Ask Claude to "create a PowerPoint presentation about Q3 results" and it uses the PowerPoint skill on its own. The directory also carries partner skills from companies such as Notion, Figma and Atlassian, built to work with their connectors.
A skill you add from the directory is enabled by default and is view-only, so you can use it but not change it. If you want a different version, download a copy, edit it and upload it as your own. If you are still choosing what to add, start with the best Claude skills for non-developers.
Turn on skills for the work your team repeats.
The audit finds the procedures your team explains over and over, the ones a skill should carry. If the built-in skills are enough, I will say so.
The audit is $2,500, and the fee comes off whatever you buy next.
Install Claude skills that someone else built
Plenty of useful skills arrive as a ZIP file rather than a directory listing: from a colleague, a vendor, or Anthropic's example repository on GitHub. To install Claude skills from a file, open Customize > Skills, click "+", choose "+ Create skill", then "Upload a skill", and pick the ZIP. The skill joins your list and can be switched on or off like any other.
The upload is the easy part. Anthropic's help center names the two main risks of skills as prompt injection, where text tricks Claude into actions you didn't intend, and data exfiltration through malicious code or leaked data. Its advice is to review any skill from a less-trusted source before enabling it, and it counts a colleague's skill in that group. In practice, that review is five checks:
- Read SKILL.md from top to bottom and confirm it does only what the description says.
- Open every file in the scripts folder, even if you can't read code fluently, and ask Claude to explain each one.
- Search the files for web addresses, and question any instruction to fetch content from a site you don't know.
- Check the license before you share the skill outside your team.
- Prefer sources you can name, and skip any skill whose author you can't identify.
None of this takes long, because most skills are a page or two of text. One rule from my agentic AI training carries over directly: a workflow that reads the open web holds no write access in the same run, so a skill that fetches pages shouldn't also be the one that publishes.
Ask for the skill by its job, or by its name
Most of the time you don't call a skill at all. Claude compares your request with every enabled skill's description and loads the ones that match, and Anthropic says it can combine several skills on one task without you naming each. That's why phrasing matters more than menus: a request that uses the description's own words is far more likely to fire.
When you want certainty, name the skill. The help center's example is "Use my brand guidelines skill to create a presentation", and the same pattern works for any recurring job. If your team runs a monthly content audit from a spreadsheet export, "Use the audit-summary skill on this export" removes the guesswork about whether Claude will match it.
The Microsoft 365 add-ins give you a third route. In Claude for Excel, PowerPoint, Word and Outlook, the skills you enabled in Claude settings come with you, and typing "/" in the sidebar lists them so you can pick one directly.
How to tell whether a skill fired
The difference between a skill being installed and a skill being used is the heart of how to use Claude skills well. Anthropic's skills walkthrough gives the visible check: open Claude's thinking and look for a line like "Reading [your skill's name]". If it isn't there, your request didn't match the skill's name or description.
What you see | What it means | What to do |
|---|---|---|
"Reading [skill name]" in Claude's thinking | The skill loaded | Judge the output against the skill's rules |
No skill line, generic output | The request didn't match | Rephrase with the description's words, or name the skill |
It fires once you reuse the description's vocabulary | Your phrasing was the mismatch | Keep using that vocabulary for this job |
It fires on requests it shouldn't | The description is too broad | Ask the skill's owner to narrow it |
Knowing how to use Claude skills includes knowing when one isn't pulling its weight, so the second check compares outputs. Run the same job twice in fresh chats, once with the skill switched off and once with it on, and read the results side by side. Anthropic's Claude Code docs explain why the chats must be fresh: leftover context from the conversation where you discussed the skill can hide gaps in its instructions.
If the two outputs are indistinguishable, the skill is either not firing or not adding anything. Both are useful findings, and the first is usually fixed in a minute.
Claude skill not triggering? Work down this table
When a skill misbehaves, the cause is almost always one of a short list, and the help center's troubleshooting section covers most of it. The table below collects those causes with the fix for each, in the order worth checking.
Symptom | Likely cause | Fix |
|---|---|---|
No skills section anywhere | Code execution is off | Turn it on in Settings > Capabilities, or ask an owner on Team and Enterprise |
Every skill is grayed out | Code execution disabled for you or your organization | Same setting; owners use Organization settings > Plugins & skills |
An enabled skill never loads | Toggle off, vague description or mismatched wording | Check the toggle, then name the skill in your request |
Upload fails | ZIP too large, folder name not matching the skill name, missing SKILL.md, or invalid characters | Fix the package and upload again |
No create or upload option on Team or Enterprise | An owner turned off user-created skills, or your role lacks the Create skills capability | Ask your owner; provisioned skills still work |
No Share button | Sharing toggles are off by default for organization and group sharing | An owner enables them in Organization settings |
Work from the top down, because the first two rows hide everything below them. A skill can't fire in an environment where skills can't run, and no amount of rephrasing fixes a switched-off setting.
Roll skills out to a team without the guesswork.
The 90-day plan from the audit covers which skills to enable, who owns each one and how you tell whether they fired. It is yours to keep.
One fixed $2,500 fee, credited to whatever you buy next.
Using skills on a Team or Enterprise plan
On a team plan, your skills list in Customize > Skills splits into three sections. Personal skills are the ones you created or uploaded. Shared with you holds skills colleagues sent you directly, and these stay off until you switch them on. Organization skills are the ones published to your organization or provisioned by an owner, and you install those from the directory rather than toggling them in the list.
Shared skills behave differently from your own. They're view-only, so you can enable and use them but not edit them, and when the owner updates one you get the new version automatically at your next use. Provisioned skills appear with a team indicator and can still be switched off if they get in your way. In Cowork, you reach the organization directory through Customize in the left sidebar, then "+", then the Skills tab.
This is the setup that lets a team scale content marketing without scaling the re-explaining. One editor writes the brief format into a skill, shares it, and every writer's next brief follows it, with updates arriving without anyone redistributing a file.
Your claude.ai skills in Claude Code
If you later try Claude Code, Anthropic's tool that runs in a terminal, your skills follow you. Skills you've enabled in your Claude settings load in Claude Code when you sign in with the same account, on Claude Code v2.1.273 or later. Run /skills and they appear under "claude.ai sync". The sync runs when a session starts and checks for changes about every 10 minutes. If you want a guided start there, compare the Claude Code training options first.
The sync is one-way, so Claude Code reads your account and never changes it. It also doesn't run when Claude Code signs in with an API key or through a cloud provider such as Amazon Bedrock. To switch it off, set syncClaudeAiSkills to false in Claude Code's settings.
Claude Code also reads skills straight from your disk: ~/.claude/skills/ for personal ones and .claude/skills/ inside a project. You can type /skill-name to run one directly. The busyless content system runs on exactly these folders, and the "Skills are just folders" lesson in my free Claude Code course shows a real case: a writing procedure of about 400 lines that stays out of Claude's context until someone asks for a guide.
Mistakes that make a good skill look broken
Most complaints about skills trace back to how they're used rather than how they're written, which is good news, because that half of how to use Claude skills is a habit and needs no rewriting.
The commonest mistake is turning on a pile of overlapping skills. Each enabled skill adds its description to Claude's context, and two skills whose descriptions both claim "blog posts" leave Claude guessing which one you meant. A small set of skills with one clear job each is easier to trigger and far easier to debug.
The second mistake is testing in the wrong chat. If you spent twenty minutes discussing a skill and then try it in that same conversation, Claude already knows what you want, and the skill looks better than it is. A fresh chat shows you what a teammate will get on Monday.
The third is uploading a fixed version of a skill without switching off the old one. Two near-identical skills with the same description compete, and you can't tell which one Claude read. Delete the old copy, or at least toggle it off, whenever you upload a replacement.
The last mistake is asking a skill to carry judgment it can't hold. Skills handle rules-based, repeatable work well, and they don't decide what your team should publish. That's the same line I draw in advanced SEO automation: automate the data-heavy, repetitive steps, and keep people on strategy and the final edit.
What a skill won't do for you
A skill teaches Claude a procedure. It doesn't connect Claude to your CMS, analytics or project tracker, which is the job of connectors built on the Model Context Protocol (MCP). Anthropic's help center puts the two side by side: connectors give Claude access to tools, and skills teach Claude how to use them well. A brief-builder skill paired with a search connector is stronger than either alone.
A skill also doesn't move data between apps on a trigger or run a ten-step workflow across five services while nobody is watching. That's what an automation platform does, and our Gumloop review looks at one built for exactly that kind of chained work. Knowing the boundary saves you from writing a skill to do a job that belongs in a different tool.
Where skills compound
One skill fixes one task. A handful that each own one step, such as research, brief, draft and edit, fix the handoffs too, because the standards travel from step to step without anyone re-explaining them. The content creation build in my automations directory is that idea at full size: a Claude Code system that researches, drafts, proofreads and fact-checks long-form articles in a client's voice before an editor signs off.
This Claude skills tutorial stops at the interface. The free course mentioned above picks up from there, with 27 lessons across six modules, no signup, all built from the system I run. It uses Claude Code, so it does involve a terminal, and it's the right next step once the interface version starts to feel small.
If you'd rather have a team build that habit with someone accountable for it, busyless offers AI enablement: six months, one workflow rebuilt with AI each month, and then your team runs it on its own. And if two skills and an afternoon are all you need, I'll tell you that on the call. Book a 30-minute call.
FAQ
Frequently asked
How do I use Claude skills without writing code?
Switch on code execution in Settings > Capabilities, open Customize > Skills, and turn on a built-in skill or add one from the directory. Then ask for the job in ordinary language, or name the skill in your request. That's the appeal of Claude skills for non-developers, because nothing in the loop needs code: Claude matches your request to the description, loads reference files on its own and runs any bundled scripts without you opening them.Do Claude skills work on the Free plan?
Anthropic's help center lists skills on Free as well as Pro, Max, Ultra, Team and Enterprise, as long as code execution is switched on. The developer overview lists custom-skill uploads from Pro upward, so if you're on Free and can't upload your own skill, that's the probable cause. Anthropic's built-in skills for Word, Excel, PowerPoint and PDF files work on every plan with code execution enabled.Why is my Claude skill not triggering?
Usually because your request didn't match the description. Check that the skill is switched on in Customize > Skills, then check Claude's thinking for a "Reading [skill name]" line. If it's missing, rephrase the request with the description's own words, or name the skill directly ("Use my brief-builder skill on these notes"). If it still doesn't load, the description itself needs narrowing or rewriting.How many skills should I turn on?
Fewer than you're tempted to. Each enabled skill adds about 100 tokens of name and description to Claude's context, which is cheap, but overlapping descriptions make it unclear which skill should fire. A good starting set is one skill per recurring task, with descriptions that don't share trigger words. Switch off anything you haven't used in a month, since the toggle keeps it ready for later.Can I edit a skill someone else made?
Not in place. Skills from the directory and skills shared with you are view-only. To change one, download a copy, edit its SKILL.md in plain language, and upload it under your own name, then switch the original off so the two don't compete. Writing and testing your own version is covered step by step in creating a Claude skill.
Written by
