What are Claude Skills, and how do companies build their own?

A Claude Skill is a folder with a SKILL.md file, holding a short name, a description of when to use it, and step-by-step instructions, plus optional scripts, templates and reference files. Claude loads it only when a task matches. Companies build their own by capturing a workflow people repeat, writing a precise description, testing it on real tasks, then publishing it to the organization, often bundled with connectors as a plugin.

Skills are how you teach Claude, powered by Anthropic, the way your company does a task: your report format, your contract checklist, your month-end commentary style. This guide explains how they work, where they run, how to build one properly, and the settings an administrator uses to govern them. Details are as of October 2026, from Anthropic's documentation and help center.

How does a skill actually work?

Skills use what Anthropic calls progressive disclosure. Claude does not read every skill in full at the start of a conversation. It reads only each skill's name and description, decides whether the current task needs one, and only then opens the instructions and any files those instructions point to. Scripts run in the code execution environment, and only their output enters the conversation, not the code itself.

Two consequences follow. You can install many skills without filling the context window, and you can bundle large reference material, such as a full chart of accounts or a style guide, at no cost until it is needed. But the description carries the whole load of deciding when the skill is used. A vague description means the skill never fires, or fires when it should not.

How are skills different from prompts, projects, connectors and plugins?

Where do skills run, and how are they shared?

Skills work in the Claude apps, Claude Code and the Claude API, but custom skills do not sync between the API and the apps. Anthropic's documentation is explicit: a skill uploaded to claude.ai must be uploaded separately to the API, and API skills are not available in claude.ai. Keep the source in Git and treat each surface as a deployment target.

Anthropic also ships pre-built skills for PowerPoint, Excel, Word and PDF that Claude uses automatically when creating documents in the apps, and that API users reference by ID (pptx, xlsx, docx, pdf). They are not included in Claude Code.

What are the rules for writing a SKILL.md file?

• Two required frontmatter fields: name and description.
• name: at most 64 characters, lowercase letters, numbers and hyphens only, no XML tags, and it cannot contain the reserved words anthropic or claude. Anthropic suggests gerund names such as preparing-variance-commentary.
• description: non-empty, at most 1,024 characters, no XML tags, written in third person, and it must say both what the skill does and when to use it, with the terms people actually type.
• Keep the SKILL.md body under 500 lines. Move detail into separate files and link them directly from SKILL.md, one level deep, because Claude may only preview files that are referenced from other referenced files.
• Give any reference file longer than 100 lines a table of contents at the top.
• Use forward slashes in paths, name files by content (revenue-recognition-rules.md, not doc2.md), and avoid dates that will go stale.
• When a skill uses connector tools, name them fully as ServerName:tool_name so Claude finds the right one.
• Prefer a script for anything that must be exact, such as validating a file or computing a reconciliation, and tell Claude whether to run it or read it.
• Never put API keys, tokens or passwords in skill files.

A good description reads like this: Drafts month-end variance commentary for the management pack using the company template. Use when the user asks for variance explanations, flux analysis, budget-versus-actual commentary, or the monthly management report. A bad one reads: Helps with finance.

How do you build a company skill, step by step?

Which skills should a company build first?

What settings govern skills across the organization?

Team and Enterprise Owners manage skills and plugins in Organization settings > Plugins & skills, which has Inventory, Marketplaces, Organization skills, Requests and Policy tabs. On Enterprise, anyone with a custom role that includes managing Libraries can do this too. The Policy tab holds the switches that matter:

• Skills: turns skills and plugins on or off for everyone. Turning it off also stops Claude Code from loading plugins synced from claude.ai accounts.
• User-created skills: whether members can create or upload their own. Off removes the option entirely.
• Skill sharing: whether members can share their skills with specific colleagues. Sharing is view-only, and on Enterprise, Share with groups extends it to user groups.
• Publishing: Requires review (nothing reaches others until an Owner or reviewer approves), Open, or Off. Team defaults to Open. The Enterprise default became Requires review on October 2, 2026. A reviewer cannot approve their own request.
• Skills and plugins scanning (Enterprise): scans skills members upload or edit for hidden code execution, data transfer and safeguard tampering. Failed or unfinished scans block the skill. Scanning does not cover skills uploaded through the API, skills already present when it was turned on, or organizations with CMEK, zero data retention or HIPAA configurations.

For each item in Inventory you set Default access: Not available, Available to install, Installed by default, or Required. On Enterprise you can override that per user group, and the most permissive group setting wins when someone belongs to several. Organization-uploaded items keep a version history you can compare and revert. Note that skill data is not covered by zero data retention arrangements; it follows Anthropic's standard retention policy.

When should you package skills as a plugin?

Use a plugin when a team needs several things installed together: for example a sales plugin with the CRM connector, the account-brief skill and a command to draft follow-ups. Owners can sync the company's own plugins from a GitHub or GitLab repository, automatically on each push to the default branch if a webhook is set up, so the repository stays the source of truth. A plugin marked Required is always on and members cannot remove it. Adding a plugin does not sign anyone in: bundled connectors still have to be added by an Owner and connected by each member.

What mistakes make skills fail?

• Descriptions that are vague or written as I can help you, so the skill never triggers or triggers on the wrong requests.
• Too many overlapping skills active at once. Each description competes for attention; consolidate narrow skills only when evaluations show no loss.
• The author approving their own skill. Separate authoring from review.
• Uploading a skill to the Claude apps and expecting the API or a Bedrock-based Claude Code to have it.
• Leaving production API agents on the latest skill version. Pin versions, because a new upload changes what every request uses.
• Installing skills from unknown sources. A malicious skill can direct Claude to run code or send data out. Treat installation like installing software.

Abstrakt Solutions designs and builds skill catalogs and plugins with the review process above; see our Claude skills and plugins services. If you are still choosing where to start, our free Claude buyer's guide walks through the options.