What Are Claude Skills? What They Are, How They Load and When to Use One

Claude skills in one sentence

A Claude skill tells Claude how to perform a repeatable job. Its description helps Claude decide when the skill applies, its main instructions explain the method, and supporting files are available when the job needs them.

In our view, a skill works well when the job has a clear trigger and a set method. Put the review standard, report format or research checklist in the skill. If the method should last beyond one chat, a skill may fit better than a long prompt copied by hand.

Prompt, skill or MCP: choose by the job

Start with the smallest tool that solves the problem. In practice, a prompt fits a one-time instruction. A skill carries a reusable method. MCP gives Claude access to an external system.

Anthropic's rule is concise: explaining how to do something belongs in a skill, while access belongs in MCP. For example, use a skill to define the checks for a content review. Use an MCP server to supply the records.

How a Claude skill loads

Claude does not need to load every file from every skill into the conversation at startup. Anthropic documents three loading levels. Claude reads the name and description as metadata at startup. It loads the main instructions when the skill is triggered. It opens reference files or runs bundled scripts only as needed.

Claude reads the name and description, loads the instructions when the skill is triggered, and opens resources only as needed. In our view, this sequence is the main design constraint. The description must make the trigger clear. The main file should stay focused and avoid becoming a library of every detail.

<style>
.xtr-sd.sd-flow{--sd-accent:#0d0d0d;margin:32px 0;padding:24px;background:#f8f9fa;border:1px solid #c9d3d2;border-radius:10px;color:#171717;line-height:1.45} .xtr-sd.sd-flow .sd-steps{list-style:none;counter-reset:sd-step;display:grid;grid-template-columns:repeat(3,minmax(0,1fr));gap:36px;margin:0;padding:0} .xtr-sd.sd-flow .sd-steps li{position:relative;counter-increment:sd-step;margin:0;padding:14px;background:#f3f7f7;border:1px solid #c9d3d2;border-radius:8px;font-size:15px;color:#171717} .xtr-sd.sd-flow .sd-steps li::before{content:counter(sd-step);display:block;width:28px;height:28px;margin:0 0 8px;border-radius:50%;background:var(--sd-accent);color:#fff;font-size:14px;font-weight:700;line-height:28px;text-align:center} .xtr-sd.sd-flow .sd-steps li:not(:last-child)::after{content:"\2192";content:"\2192" / "";position:absolute;top:50%;right:-34px;width:32px;transform:translateY(-50%);font-size:30px;line-height:1;font-weight:700;text-align:center;color:var(--sd-accent)} @media(max-width:900px){ .xtr-sd.sd-flow{padding:16px} .xtr-sd.sd-flow .sd-steps{grid-template-columns:1fr;gap:44px} .xtr-sd.sd-flow .sd-steps li:not(:last-child)::after{content:"\2193";content:"\2193" / "";top:auto;bottom:-42px;right:auto;left:50%;width:32px;transform:translateX(-50%)} }
</style>

Level one: discovery metadata

The name and description do more than label the folder. Anthropic says Claude loads this metadata at startup and matches the description against the user's request. A vague description can make a good skill difficult to discover. A useful description says what the skill does and when it should be used.

Level two: the working instructions

When the request matches, Claude loads the body of the skill's main instruction file. This is where the method belongs: the sequence, decisions, quality checks and boundaries that should remain consistent. It should tell Claude what a successful result looks like without forcing every supporting detail into the initial context.

Level three: resources on demand

Reference documents, templates and scripts can sit beside the main instructions. Anthropic says references enter context when read, while scripts run through the command environment and only their output enters context. That lets a skill keep specialist material available without loading all of it for every request.

What belongs in the skill folder

A small skill needs a clear main file and only the resources the job uses. Its metadata should make the trigger clear. The main file should cover the set method. Other files should hold details that are not needed on every run.

• Metadata: a name and a description that identify both the capability and its trigger.
• Main instructions: the repeatable decisions, steps and success checks.
• References: standards, examples or templates that are read only when relevant.
• Scripts: deterministic operations that are safer or more reliable as code than as repeated prose.

In our view, this division prevents two common mistakes. Putting the whole knowledge base in the main file makes every triggered use heavier. Hiding an important decision rule in a reference file means Claude may never open it. Put the method in the instructions and the supporting depth in resources.

Where custom skills live

In Claude Code, Anthropic says custom skills are filesystem based. A personal skill can live in ~/.claude/skills/, while a project skill can live in .claude/skills/. A project location is useful when the method belongs with a particular codebase; a personal location is useful when the same method follows one user across projects.

Custom skills do not automatically sync across Claude surfaces. A skill available in Claude Code is not automatically available in claude.ai or through the API. Treat each surface as a separate installation point. Ask where a skill is installed before troubleshooting it.

Trust checks before you use a third-party skill

A skill can include instructions and executable code, so it deserves the same review discipline as any other automation. Anthropic advises using skills from trusted sources and warns that a malicious skill can direct Claude to call tools or run code in ways that do not match the advertised purpose.

Before using a third-party skill, read the main instructions and inspect every bundled script. Check what tools it asks Claude to use, what files it can reach, what network calls it makes and whether the visible description matches the actual behavior. If the package is too large to review, that is a reason to pause rather than a reason to trust it.

Review every instruction and script before use. Pay special attention to any file or service access.

A practical way to decide

Ask four questions in order. Is the instruction needed more than once? Does the job have a stable method? Does it need outside access? Can the instructions and any code be trusted? A repeatable method points toward a skill. Outside access points toward MCP. A one-time request with neither points back to a prompt.

Consider this split for an AI search team. Put the review method, proof and report format in the skill. Supply data through files or an approved link. Xtrusio helps teams examine how brands and sources appear in AI answers. The skill is the reusable method around that work, not a substitute for proof.

Official sources

These primary pages cover skill structure, use across Claude surfaces, API use, and MCP.

• Anthropic Agent Skills overview
• Skill authoring best practices
• Extend Claude with skills
• What are skills?
• Using Agent Skills with the API
• Anthropic: extending Claude with skills and MCP
• Public repository for Agent Skills
• What is the Model Context Protocol?

Last checked: October 27, 2026.