What are Claude skills? A plain guide for teams
What a Claude skill is, what sits inside the folder, how it differs from a project or a connector, and what makes one worth having on your team.
King & Company
In short
A Claude skill is a folder containing a SKILL.md file of instructions, plus optional scripts, templates, and examples, that Claude loads by itself when a request matches the skill's description. The skills worth having hold your firm's own procedure for one job, such as a proposal or a weekly report, and they are plain files you own and can read. Anthropic's guidance is to test a skill on real cases, keep it short, and audit any skill that comes from someone you do not know.
Claude skills are, in Anthropic's words, "folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks." For a team that runs on documents, a skill is where the firm's procedure for one job gets written down once, with its template and a few good examples beside it, so that nobody has to retype it into a chat window.
If your team already uses Claude, you have seen the pattern a skill replaces. Someone writes a long paragraph explaining how the firm formats a client memo, what the weekly report has to include, and which mistakes to avoid. The result is good, and next week they write the paragraph again, slightly differently, and a colleague writes their own version. A skill holds that paragraph in one place and Claude picks it up when the job comes around.
What is inside a skill folder?
Every skill is a folder with a file called SKILL.md. The file opens with two required fields, a name of up to 64 characters and a description of up to 1,024 characters, and the rest is instructions written in plain language. Anthropic compares the folder to "an onboarding guide you'd create for a new team member."
Beyond that one file, a skill can bundle three kinds of material:
- More instructions. Additional files for the parts of the job that only come up sometimes, such as a separate guide for one unusual type of engagement.
- Resources. Templates, examples, and reference material, such as your proposal template and two finished proposals you are proud of.
- Scripts. Small programs Claude runs for the steps that must come out the same way every time.
A skill for a weekly report might hold the instructions, the report template, last month's four reports as examples, and one script that fills the firm's Excel file.
How does Claude decide when to use a skill?
Claude loads a skill in three stages, which Anthropic calls progressive disclosure. The name and description of every installed skill are always loaded, at about 100 tokens per skill. The body of SKILL.md is read only when a request matches the description. The bundled files are opened only when the instructions call for them. An idle skill takes up about as much room as its own description.
This is why a team can keep many skills installed. The Help Center pages we read give no maximum, and Anthropic's authoring guide talks about Claude choosing "the right Skill from potentially 100+ available Skills."
It also means the description is the trigger. Anthropic says the description is what Claude matches a request against, so it has to say both what the skill does and when to use it. The guide gives "Helps with documents" as an example of a description to avoid. When a skill never seems to fire, the description is the first thing to check. A description that works reads more like "Drafts the firm's weekly client status report from the project tracker export. Use when someone asks for a weekly report, status update, or client summary."
Skills compared with projects, custom instructions, and connectors
These four features get confused because all of them shape what Claude does. Anthropic's Help Center separates them this way:
| Feature | What it gives Claude | When it applies |
|---|---|---|
| Skill | The procedure for one kind of task | Loaded only when a request matches it, anywhere in Claude |
| Project | Background knowledge for one body of work | Always loaded inside that project |
| Custom instructions | General preferences | Every conversation |
| Connector (MCP) | Access to an external service or its data | When Claude needs to reach that system |
A connector and a skill work well together. The connector lets Claude reach the folder where client files live, and the skill tells it how your firm turns those files into a memo. If you are weighing a skill against a larger build, we cover that choice in AI workflow vs agent vs skill vs custom software.
What does a good skill look like, and why do downloaded ones disappoint?
Search for skills and you will find long lists of them to download. Our view, from building them, is that a generic skill gives generic output, because it cannot contain the things that make the work yours: your template, your fee language, the section a partner always rewrites. Claude can already write a reasonable proposal without help. A skill earns its place when it carries what only your firm knows.
Anthropic's authoring guidance points the same direction. It says to start from the assumption that Claude "is already very smart" and to add only the context it does not already have, and to keep the body of SKILL.md under 500 lines, moving detail into separate files. Three other points from that guide matter more than the folder format:
- Test before you write. Anthropic says to "create evaluations BEFORE writing extensive documentation": run Claude on representative tasks without the skill, note where it falls short, and write only enough to close those gaps. The guide suggests three test scenarios, and three real past jobs where you already know the right answer serve that purpose well.
- Use a script where the output must not vary. The guide says to prefer scripts for deterministic operations and describes bundled scripts as more reliable than code Claude writes on the spot. Filling the firm's Excel template belongs in a script. Deciding what the cover note should say belongs in the written instructions.
- Show examples. The guide notes that examples convey style and level of detail more clearly than descriptions alone.
For a worked set of skills in one industry, see Claude skills for commercial real estate.
How do you write your first skill from work you already did?
The best source material for a skill is the set of corrections a senior person makes to a draft. Anthropic describes the method in its guide: complete a task with Claude using normal prompting, notice the context you keep supplying, and then ask Claude to create a skill that captures it. The same page says Claude understands the skill format and will generate a properly structured SKILL.md when asked, which is why a person who does not write code can produce an instruction-only skill.
A working session looks like this. Pick one recurring job. Have the person who owns it do the job with Claude on a real case and correct the draft out loud: this section is too long, we never quote that figure, the client name goes in the header. Ask Claude to turn those corrections into a skill. Then open a fresh conversation, give the skill a different real case, and see what it still gets wrong. Anthropic's guide recommends the same loop of using the skill on actual tasks, observing where it struggles, and refining.
Writing your own works well for instruction-only skills. It is worth getting help when a skill needs scripts, has to feed other systems, or sits inside a longer process with a review step, which we describe in how to build an AI workflow. That is the work we do in weekly sessions with a client's team, and you can get in touch if you would like to talk it through.
How does a team share skills?
Anthropic's Help Center says skills are available on Free, Pro, Max, Team, and Enterprise plans and require code execution to be enabled. On Team and Enterprise plans, an owner turns the feature on in organization settings, and from there the Help Center describes three routes: an owner can provision a skill for all users, a member can share a skill with specific people, and a member can publish a skill to the organization's library. A custom skill is uploaded as a ZIP file of the folder.
Three cautions apply. The developer documentation lists custom skill uploads in claude.ai on Pro, Max, Team, and Enterprise plans only, so someone on the Free plan should check what their own account offers. Anthropic's announcement of skills was updated on December 18, 2025 to add organization-wide management, while its developer documentation still says claude.ai custom skills are individual to each user. The Help Center reflects the newer state, and you should check your own admin settings before planning a rollout. The developer documentation also says custom skills do not sync between claude.ai, the API, and Claude Code, so a firm using more than one of those manages its skills in each place.
Are Claude skills safe to install?
It depends on where the skill came from. Anthropic's documentation says to use skills only from trusted sources, meaning those you created yourself or obtained from Anthropic, and to treat a skill like installing software. It warns that a malicious skill can direct Claude to run code or use tools in ways that do not match its stated purpose, and that this could lead to data leaving your environment, depending on what Claude has access to.
If you must use a skill from elsewhere, the same page says to audit every file in it, including scripts and other resources, and to be wary of skills that fetch content from external URLs. It also notes that Enterprise organizations can turn on content scanning for custom skills uploaded in claude.ai and Claude Cowork. Have your IT or security lead decide who may install skills and from where, and fold that into the rules you set for what an AI agent can read, write, and send.
Who owns a skill, and can it move to another AI tool?
A skill is a set of files. You can open SKILL.md in any text editor, read every instruction, and keep the folder wherever you keep other firm documents. Anything we build for a client is the client's to keep, and a skill makes that easy to verify because every file in it can be opened and read.
The format is also portable. The Agent Skills format was developed by Anthropic and released as an open standard, and the products listed as supporting it include ChatGPT and Codex, Gemini CLI, Cursor, GitHub Copilot, and VS Code. A skill that relies on a specific connector or on scripts written for one environment may need adjusting when it moves, so test it in the new tool before you depend on it. The written procedure, which is the valuable part, travels as it is.
Common questions
What is a Claude skill in simple terms?
It is a folder with a file of written instructions, and sometimes scripts and templates, that tells Claude how to do one specific job. Claude keeps only the skill's name and description in view and opens the rest when a request matches that description.
Do I need to know how to code to create a Claude skill?
Not for a skill made of written instructions. Anthropic's authoring guide says you can ask Claude to create a skill and it will produce a properly structured SKILL.md file. Code comes in when a step has to produce the same result every time, such as filling a fixed spreadsheet template.
What is the difference between a skill and an MCP connector?
Anthropic's Help Center says a connector built on MCP links Claude to external services and data, while a skill supplies the procedure for completing a task. A connector gives Claude access to your document system, and a skill tells it how your firm writes the memo once it has the documents.
Can an admin roll out skills to the whole firm?
Anthropic's Help Center says owners of Team and Enterprise organizations can provision skills for all users, and that members can share a skill with specific people or publish it to the organization's library. Anthropic's developer documentation still describes claude.ai custom skills as individual to each user, so confirm the behavior in your own admin settings before you plan around it.
Do Claude skills work in other AI tools?
The format is published as an open standard at agentskills.io, and that site lists ChatGPT and Codex, Gemini CLI, Cursor, GitHub Copilot, and VS Code among the products that support it. A skill that depends on a particular tool's scripts or connectors may still need adjusting when it moves.