What Claude skills are, and what you would use one for

A Claude skill is a folder holding a short instruction file named SKILL.md that tells Claude how to do one specific task, along with any reference files, templates, or scripts that task needs. Claude reads only the name and description of each installed skill at startup and loads the full instructions when a request matches, so an unused skill costs nothing and a triggered one behaves the same way every time (Anthropic Agent Skills documentation, retrieved 2026-08-22). The format is an open standard published at agentskills.io rather than a feature that belongs to Claude, which is why other assistants and coding agents have adopted the same folder format.

What is a Claude skill in plain terms?

It is a written procedure that your AI can actually follow, kept in a folder you control. The instruction file says what the task is, when it applies, what steps to take, and what to leave alone. Everything else in the folder is material the procedure refers to: a tone guide, a spreadsheet template, a list of account names, a script that does one deterministic thing.

The mechanism is worth knowing because it explains the behavior you will see. Anthropic calls it progressive disclosure and documents it in three stages: the name and description are always loaded, the body of SKILL.md loads when a request matches the description, and bundled files cost nothing until something reads them (Anthropic Agent Skills documentation, retrieved 2026-08-22). Two consequences follow. You can install many skills without slowing anything down, and a skill whose description is vague will simply never fire, which is the most common reason one appears to do nothing.

Is this the same thing as the SOP I already keep in a shared drive?

Yes, and that is the most useful way to think about it. A skill is the AI-first version of a standard operating procedure: the same steps, written so that nothing load-bearing lives only in the head of the person who normally runs them. Anthropic’s own help center describes the point of custom skills in those terms, as packaging company workflows and institutional knowledge so they get applied consistently (support.claude.com, what are skills, retrieved 2026-08-22).

What changes in the translation is not the writing style. A human-first procedure can say “review it and send it to the client,” because a person fills in which client record is authoritative, which mailbox to send from, and whether sending is their call at all. An agent cannot fill any of that in. So the AI-first version names the specific account, the specific inputs, what a good result looks like specifically enough that someone else could check it, and the conditions under which the work stops and comes back to you.

That last part is where most of the value sits, and it is worth writing even if no agent ever runs the file. “Check with me on the big stuff” cannot be executed. “Never message anyone outside the company, never change a permission, never touch more than twelve records in one run, report the count before doing anything” can be. Writing those conditions out is usually the moment a procedure stops being vague.

Do I need to be technical to make one?

No. A skill can be one Markdown file with two lines of metadata at the top and plain instructions underneath, and Anthropic states directly that simple skills need no code (support.claude.com, what are skills, retrieved 2026-08-22). The required fields are name and description; the name is capped at 64 characters and limited to lowercase letters, numbers, and hyphens, and the description is capped at 1,024 characters and has to say both what the skill does and when to use it (Anthropic Agent Skills documentation, retrieved 2026-08-22).

The belief that stops most people is that AI capability sits on top of tool expertise, so you would have to be an Excel expert before you could specify AI work in Excel. The specification is about your work, not about the software’s internals. What it asks of you is four answers, and they are the same four used on every guide on this site.

1. What is this task’s one job?
2. What does a good result look like, specifically enough that someone else could tell?
3. What must never happen without my approval?
4. What supervision does this task need?

An assistant that has read a half-specified procedure will ask about the gaps. Answering those questions is the work, and the answers are the skill.

Scripts are the one part that involves code, and they are optional and last. Add one only after repeated runs show that a fixed piece of code is more reliable than having the assistant produce the equivalent each time.

What skills already exist that I could just turn on?

Three kinds, and the first kind is on by default. Anthropic publishes and maintains its own skills for document work, described as enhanced document creation for Excel, Word, PowerPoint and PDF files, available to all users and invoked automatically when relevant (support.claude.com, what are skills, retrieved 2026-08-22). So three you already have: the Excel skill, which builds and edits real workbooks rather than describing them; the PowerPoint skill, which produces a deck as a file; and the PDF skill, which reads and produces PDFs. Nothing to install.

The second kind is the partner skills in the Skills Directory, described as professionally built skills from partners including Notion, Figma and Atlassian, designed to work alongside those vendors’ MCP connectors (same source, retrieved 2026-08-22). The pairing is the point: the connector reaches the account, the skill carries the method.

The third kind is the ones you write. Anthropic’s own examples of what people use custom skills for are worth reading as a starting list: applying brand style guidelines to documents and presentations, generating communications that follow a company email template, structuring meeting notes to a company format, creating tasks in company tools such as Jira, Asana or Linear following team conventions, and running a company-specific data analysis (same source, retrieved 2026-08-22).

To see what is available in your own account: click Customize, then Skills, then + and Browse skills (same source, retrieved 2026-08-22). This page does not publish a ranked list of the best skills, and will not. Which one is right depends on the task in front of you, and nothing on a ranking can know that.

Can I even use skills on my plan?

Skills are available on the Free, Pro, Max, Team, and Enterprise plans, and the feature requires code execution to be turned on (support.claude.com, use skills in Claude, retrieved 2026-08-22). They are also documented as available in beta for Claude Code users and for API users working through the code execution tool, from the same source.

The setting has an intimidating name and a mundane meaning. Code execution is the same capability that lets Claude create and edit files for you, and on claude.ai the code runs inside Claude’s own container rather than on your computer (support.claude.com, use skills in Claude, retrieved 2026-08-22). In Claude Code the situation is genuinely different and worth knowing before you install anything: there, a skill runs on your machine with the same network access as any other program you run (Anthropic Agent Skills documentation, retrieved 2026-08-22).

The path, as documented on 2026-08-22:

• Free, Pro, and Max. Settings > Capabilities, turn on Code execution and file creation. Then Customize > Skills, and toggle on the ones you want.
• Team. Enabled at the organization level by default. Members go straight to Customize > Skills.
• Enterprise. An owner has to turn on both Code execution and file creation and Skills in Organization settings > Skills before anyone sees the section.
• Uploading your own. Package the skill folder as a ZIP, then in Customize > Skills use the ”+” button, Create skill, Upload a skill. The developer documentation gives the path on claude.ai as Settings > Features rather than Customize > Skills; the help-center path above is the one used here, and your own screen wins over either.

Two limits belong in the same breath as the instructions. Custom skills you upload to claude.ai are private to your own account, and sharing them with colleagues exists only on Team and Enterprise plans, where the org-wide and group sharing toggles start off and an owner has to enable at least one before a Share button appears (support.claude.com, use skills in Claude, retrieved 2026-08-22). And a skill does not follow you between surfaces: one uploaded to claude.ai is not available through the API, API skills are not available on claude.ai, and Claude Code reads its own folders separately from both (Anthropic Agent Skills documentation, retrieved 2026-08-22).

If your organization has turned skills off, that is a decision someone made rather than a missing feature, and the specific question to ask is whether Code execution and file creation and Skills are enabled in organization settings.

This page stops at what one person can write, own and ask for. Anything that promises a result across a whole team, including a shared library of provisioned skills that people actually follow, is a different kind of project with a different kind of budget, and it is not what a page like this can deliver.

How is a skill different from an MCP connector or a plugin?

A skill carries the procedure; a connector carries the reach. Anthropic states the division plainly: the Model Context Protocol connects Claude to outside services and data, skills supply the knowledge of how to complete a task, and the two are meant to be used together, with the connector providing access and the skill providing the method (support.claude.com, what are skills, retrieved 2026-08-22).

That combination is already shipping as a product rather than a theory. The same page describes the skills directory as carrying skills from partners including Notion, Figma, and Atlassian, built to work alongside those vendors’ MCP connectors. On the connector side, the public MCP registry at registry.modelcontextprotocol.io/v0/servers is an open paginated API whose limit parameter caps at 100 results; walking all 243 pages with version=latest on 2026-08-22 returned 24,221 distinct server names. Listing does not make a server official. Two checks settle that, a reverse-DNS namespace on the vendor’s own domain or an io.github.<org> namespace whose organization owns the product’s canonical repository, and the applications that passed are on the best MCP servers page.

A plugin is the packaging rather than a third kind of thing. It is a package that installs a set of pieces at once, which is how a skill gets shared in Claude Code, and it is why skills, connectors, and plugins now sit in one directory together (Anthropic Agent Skills documentation and support.claude.com, what are skills, both retrieved 2026-08-22).

Where a skill and a project pull apart is timing. A project holds background knowledge that is loaded for every conversation inside it, while a skill activates only when a request matches its description and is available wherever you are working (support.claude.com, what are skills, retrieved 2026-08-22). Custom instructions sit at the other extreme, applying to everything you do.

Are skills only a Claude thing?

No. Anthropic developed the format, released it as an open standard, and publishes the specification at agentskills.io, where the front page states that the standard is open to outside contributions and lists the products that have adopted it. The name is what misleads: every piece of vocabulary around this feature is Anthropic’s, so the format reads as one company’s product when it is not. A count of that showcase on 2026-08-22 found 46 named clients, among them ChatGPT and Codex, Gemini CLI, GitHub Copilot, VS Code, Cursor, Goose, Mistral’s Vibe, Databricks, and Snowflake Cortex Code.

So “claude skills” is a vendor’s name for a portable folder format. The practical consequence is that the effort you spend writing down a procedure is not a bet on one assistant company. The reverse is also true and less comfortable: a skill you download is a folder of instructions and code that will run with whatever access the assistant already has, which is why Anthropic’s own documentation says to use skills only from sources you trust, to read every bundled file before enabling one, and to treat the decision the way you would treat installing software (Anthropic Agent Skills documentation, retrieved 2026-08-22). The vendor names prompt injection and data exfiltration as the primary risks and, on the Enterprise plan, offers scanning of uploaded skills that usually finishes in a minute or two (support.claude.com, use skills in Claude, retrieved 2026-08-22).

Can a skill work inside the apps I already use?

Yes, in two different ways, and neither one waited for the app’s own vendor to ship an AI feature.

The first is inside the application window. Skills you have enabled in your Claude settings are also available in the Claude add-ins for Excel, PowerPoint, Word, and Outlook, where you can type / to pick one or just describe the task and let Claude recognize which skill applies (support.claude.com, use skills in Claude, retrieved 2026-08-22). The same page notes that Claude adapts a skill to the surface it is in, and that a skill built around a specific Excel or PowerPoint template can load that template into the open file. A procedure you wrote in Markdown is running inside Microsoft Office because an assistant reached in from outside.

The second is through a connector, which is how a skill touches a system that has no add-in at all. The skill says which records to read, what to do with them, and when to stop; the connector holds the authorization to the account. For scale on what is reachable that way: this site publishes 589 applications from Composio’s connector catalog, which holds 1,000 in total, with a median of 19 actions retrieved each, queried 2026-08-22 from an inventory built 2026-08-14. The 1,000 is the ceiling of the tool that produced the snapshot, which records the truncation in its own metadata, so it is a floor rather than a count. The median matters more than the total, because most applications expose enough actions for a bounded procedure and not enough for an open-ended one. Per-application detail is in the tools index, including Notion, Slack and Asana.

How do I own and maintain one?

Keep it in a Git repository, because the properties you want from a written procedure are the ones Git already has: you can see who changed a step, when, and why, read the difference, and put it back. In Claude Code the folders are read from ~/.claude/skills/ for personal use or .claude/skills/ inside a project, and the project location is the one that ends up in version control and can be shared through a plugin (Anthropic Agent Skills documentation, retrieved 2026-08-22). Nothing about that requires you to be a programmer. It requires a folder and the habit of committing a change with a sentence about why.

Credentials are the one place where a hobby setup and a maintained one diverge sharply. The quick route is an API key pasted into a configuration file, which then sits in the repository next to the procedure, is as broad as whatever it was issued for, and expires one day without telling you. A managed connector keeps the authorization outside the repository instead: Composio stores the auth config and the connected account, refreshes OAuth tokens ahead of expiry, and marks a connection expired only after refresh attempts fail, so a break shows up as a connection status rather than as a procedure that quietly stopped working (Composio authentication documentation, retrieved 2026-08-22). Either way, the rule for the folder is the same: no key, token, or session cookie goes in it.

The rest of maintenance is dating things and keeping the dates honest. Prices, plan gates, settings paths, and thresholds copied into a procedure decay faster than the steps do. Everything in this page about plan availability was checked on 2026-08-22 and will need checking again. When you find something that contradicts what the file says, treat it as new evidence to decide rather than as an automatic replacement, because newer is not the same as more authoritative.

When is a skill the wrong tool?

When the rules of the work are still changing. Specifying a procedure that will be different next month spends the effort in the wrong place, and the specification will be wrong in ways nobody notices until it runs and produces something plausible.

There are four ways of working with an AI on your own stack, and each of them is the right answer for some work and the wrong answer for other work.

None of these is a step toward the next one. Read-only is the permanent right answer for plenty of work, and a written skill is the wrong answer wherever the definition of a good result is still moving. The full version of the table, with what each shape is wrong for, is on the manifesto.

Two other cases to leave alone. A procedure only one person can describe is not yours to write down without their agreement, because writing it converts their judgment into something anyone can run. And a procedure that turns out to need simplifying rather than automating is finished when you simplify it; reshaping the work is a complete outcome, not a delay.

The honest test of whether you have written a skill down properly is to trigger it in a fresh session with no prior conversation to fill the gaps. What breaks is what was still in your head. Check that separately from whether the output was any good, because a skill that never fired and a skill that fired and did the wrong thing need different fixes.

What could not be established here?

• Whether the Free plan can upload custom skills. The two vendor pages disagree as of 2026-08-22. The help center lists Free among the plans where you can enable example skills and upload your own (use skills in Claude), while the developer documentation says custom skills on claude.ai are available on Pro, Max, Team, and Enterprise with code execution enabled and does not list Free (Agent Skills overview). Both were retrieved the same day. Check your own Customize > Skills screen before planning around either.
• Whether a skill runs unchanged across those 46 clients. The count is of products listed on the standard’s own showcase, taken 2026-08-22, and Anthropic’s own help center states that skills are not locked to Claude and that the same format works across platforms that adopt the standard. No cross-client execution test was run here, adopting a specification is not the same as implementing every part of it identically, and bundle-level packaging is explicitly out of scope for the Skills Over MCP working group.
• Whether the partner directory skills need a paid plan on the partner’s side. Notion, Figma, and Atlassian skills are described as working with those vendors’ connectors. What each vendor charges for the connector was not checked.
• Whether any of this saves time, or how much. No measurement was run. Any hours-saved figure attached to skills, including one this site produces later, is an estimate until it is measured against a specific procedure.
• How many procedures in a business are specifiable. No count exists here. A claim that most of them are would be a guess.
• Whether registry-listed MCP servers are the vendor’s own. No namespace check was performed for this page, so it makes no claim about any specific vendor’s server being first-party. Forty-one applications have been checked on the best MCP servers page.
• How many of the 589 a person can turn on unaided. The 589 published applications were not checked for how many are reachable without an administrator enabling something first. That count does not exist yet.
• What the demand for this topic actually means, and which instrument to judge this page on. claude skills returns a reported 40,500 monthly Google searches at competition index 5, measured 2026-08-22, United States, English. That reported figure is a trailing average on a falling series: the term peaked at 90,500 in March 2026 and the latest measured month is 27,100, which is the number recorded in this page’s frontmatter. A decision to keep or retire this page on the reported 40,500 would be made on a number this page’s own evidence says is stale. In the AI channel the same term went from 2 to 321 over twelve months, close to monotonically, and kept rising through the period when Google fell. Absence or decline of reported search volume is not absence of demand, and the AI-channel figure is modeled from People Also Ask data rather than counted from assistant traffic, so read the two as directions rather than as measurements.
• Whether any of the pages currently ranking for this term will still describe it this way next quarter. The plan gates and settings paths above are the fastest-moving claims on this page, which is why each one carries its date.

Sources