Quick answer: Claude Skills are folders of instructions (a SKILL.md file plus optional code or templates) that Claude reads automatically when its description matches your request. Anthropic launched them on October 16, 2025, then released the Agent Skills specification as an open standard in December 2025. Skills require no server, API key, or deployment. They are best for instructions, templates, and workflows. MCP servers are best for live data and external APIs. The two compose well: Skills define the workflow, MCP servers fetch the data.
What Are Claude Skills?
A Claude Skill is a folder containing a single SKILL.md file (with optional supporting scripts, templates, and reference docs). Claude reads the skill *automatically* when its description matches what you're trying to do, then follows the instructions inside.
That's it. No server. No API key. No deployment. Just a folder with markdown.
If you've ever wanted to teach Claude "here's exactly how to write our company memos" or "always use these design tokens when building React components" or "follow this 5-step checklist before deploying" — Skills are the answer. They're the cleanest way to give Claude long-lived, reusable knowledge that doesn't bloat your prompts.
Why Skills Matter — Anthropic's Own Numbers
Skills are not a side feature. According to internal research Anthropic published in December 2025, the company's own engineers used Claude in 60% of their work and reported a 50% productivity boost — a two-to-threefold increase over the prior year. Notably, 27% of that Claude-assisted work was tasks that would not have been done otherwise — internal tools, documentation, and "papercuts" employees called perpetually deprioritized.
That last number is the interesting one. Skills aren't just speeding up existing work; they're making it economical to do work that previously got skipped.
Industry analysts agree this matters. As Arun Chandrasekaran, an analyst at Gartner, told AI Business: "For many of the business use cases that require human input, now you can actually have an AI model with the skill go ahead and do that action on your behalf." That's the practical difference between a chatbot and a workflow tool — and Skills are how Claude crosses that line.
How Skills Actually Work
Every skill has the same structure:
my-skill/
├── SKILL.md # required: instructions + when to trigger
├── scripts/ # optional: code Claude can run
│ └── helper.py
├── templates/ # optional: reference files
│ └── memo.docx
└── examples/ # optional: example outputs
The SKILL.md file has a YAML frontmatter and markdown body:
---
name: company-memo
description: Use this skill whenever the user asks to write a company memo,
internal announcement, or formal communication. Triggers on phrases like
"draft a memo", "write an announcement", "send to the team".
---
# Company Memo Skill
Memos follow this structure:
- One-line subject (under 60 chars)
- TL;DR in the first paragraph
- Three sections max: Context, Decision, Action Items
- Sign with "— [Author], [Title]"
Always use templates/memo.docx as the base format.
When the description's trigger phrases match the user's request, Claude reads the entire SKILL.md, follows the instructions, and produces output that matches the skill's standards.
Skills vs MCP Servers: The Real Difference
Both Skills and MCP servers extend what Claude can do. But they solve different problems:
| Skills | MCP Servers |
|---|
| What it is | A folder with markdown + optional code | A running HTTP server exposing tools |
|---|---|---|
| Lives where | On disk, read at runtime | On a server, called via API |
| Best for | Instructions, templates, workflows | Live data, external APIs, real-time queries |
| Setup time | 5 minutes (write markdown) | 2-3 hours (build + deploy server) |
| Network required | No | Yes |
| Third-party access | No (private to you) | Yes (depends on the API) |
| Example use cases | "How to write our reports", design system rules, code style guide | Read Gmail, query Postgres, post to Slack, fetch from Google Search Console |
The simplest rule:
- Use a Skill if the task is "follow these instructions" or "use this template."
- Use an MCP server if the task is "go fetch this live data" or "make this change in an external system."
The Built-In Skills Anthropic Ships
Out of the box, Claude has access to several skills designed for common workflows:
docx— Create, edit, and read Word documents (.docx). Handles styles, tables, page numbers, letterheads.pptx— Build PowerPoint decks with consistent formatting and speaker notes.xlsx— Generate Excel files, including formulas, charts, and multi-sheet workbooks.pdf— Create, fill, merge, split, watermark, and OCR PDFs.frontend-design— Production-grade UI patterns and design tokens for React/HTML output.pdf-reading— Extract text, images, and tables from PDFs (including scanned ones).file-reading— Decide which tool to use for which file type (PDF, docx, xlsx, archives, ebooks).
These are why Claude can now produce a polished Word doc or PowerPoint deck without you specifying every formatting detail. The Skills do the heavy lifting.
When to Use Skills (Real Examples)
1. Style guides and writing standards
If your team has a specific way of writing — voice, terminology, structure, no-no phrases — that's a Skill. Write it once, Claude follows it forever.
---
name: digidog-blog-style
description: Use when writing blog posts for DigiDog. Triggers on "write a blog post",
"draft an article", or any content creation task for digidog.org.
---
DigiDog blog style:
- Direct, opinionated, no buzzwords
- Open with the problem, not the solution
- Code examples must be runnable
- Always include a comparison table or list
- End with a real CTA, never "let me know if you have questions"
2. Recurring workflows
Anything you do every week that has the same steps. Customer onboarding checklist. Deploy procedure. Code review checklist. Weekly status report format.
3. Project context
Persistent context that doesn't fit in a prompt. We use skills for ongoing client projects — the full stack details, key contacts, repository structure, ongoing issues. Claude reads it automatically when we mention the project name.
4. Templates with logic
Templates that aren't just files but also include "here's how to fill it out" rules. Company memo template + the actual structural rules for memos. A pitch deck template + the narrative arc that must be followed.
When You Still Need MCP
Skills are powerful but they're static. They can't:
- Query live data (yesterday's revenue, current weather, your inbox)
- Take actions in external systems (send an email, create a Jira ticket, deploy code)
- Access third-party APIs (Stripe, GitHub, Salesforce, Google Search Console)
- Maintain state across conversations (search history, user preferences, cached results)
For any of those, you need an MCP server. The good news: Skills and MCP servers compose. You can have a Skill that says "when the user asks for revenue analysis, use the stripe_get_payments MCP tool, format the output as a Markdown table, and add commentary using our standard structure."
That's actually the most powerful pattern. Skills provide the *workflow*, MCP servers provide the *data*.
How to Create Your First Skill (10 Minutes)
If you're using Claude Code or Claude Desktop, you can drop skills into your ~/.claude/skills/ directory (Claude Code) or configure them in your settings. For Claude API access, you upload skills via the project.
Step 1 — Create the folder:
mkdir -p ~/.claude/skills/my-first-skillStep 2 — Write the SKILL.md:
---
name: my-first-skill
description: Use this skill when the user asks to summarize meeting notes
or call transcripts. Triggers on "summarize this meeting", "extract action
items", "what did we decide", or when given a transcript.
---
# Meeting Summary Skill
Always produce a summary with exactly these sections:
Decisions Made
- Bullet list, one decision per line
- Include who owns the decision
Action Items
- Format: [Owner] - [Action] - [Due]
- If no due date mentioned, use "No deadline set"
Open Questions
- Anything left unresolved
- Tag with @[person] if a specific person needs to answer
Next Meeting
- Date if mentioned, otherwise "TBD"
Step 3 — Test it. Open Claude, paste a meeting transcript, and ask "summarize this meeting." The skill triggers automatically because the description matches.
That's the whole loop. Five minutes of work, and Claude now has a reusable workflow that produces consistent output every time.
The Practical Implication for SaaS
This is where it gets interesting from a business perspective.
A lot of SaaS tools today are essentially expensive UIs wrapped around someone else's workflow. Notion templates. Asana project structures. Sales playbooks. Email sequences. Onboarding checklists. People pay $20-$200/month per user for products that mostly enforce a specific way of doing things.
Skills replace that pattern at near-zero cost. A Skill defining your sales playbook costs nothing to "deploy," can be updated by editing one markdown file, and runs inside your AI assistant rather than requiring everyone to log in to yet another tool.
For internal workflows — onboarding, documentation, reporting, standard procedures — Skills will eat a meaningful chunk of the SaaS stack within 12 months. The companies that recognize this early will operate with fewer tools and faster cycles.
What's Next
If you're thinking about which parts of your stack to migrate to Skills + MCP first, the rough order is:
1. Internal documentation tools — Notion templates, Confluence pages → Skills
2. Standard operating procedures — How-to guides, checklists → Skills
3. Light reporting — Weekly status reports, dashboards → Skills + MCP (for data)
4. Communication templates — Outreach emails, follow-ups, proposals → Skills
The integration patterns are still being figured out — Anthropic shipped Skills in October 2025 and turned the Agent Skills spec into an open standard that December — but the direction is clear: more of what we call "SaaS" today will be local instructions + targeted API access, not full-blown apps.
If you want help mapping which parts of your stack are replaceable with Skills + MCP (and which still need real software), see our AI Integration service for what we ship end-to-end, or book a free AI Operations Audit. We'll go through your current tools and show you exactly where the cost can come out.
Frequently Asked Questions
What is a Claude Skill?
A Claude Skill is a folder containing a SKILL.md file with instructions Claude reads automatically when the file's description matches the user's request. Skills can include optional scripts, templates, and reference documents. They require no server, API key, or deployment.
Are Claude Skills the same as MCP servers?
No. Skills are folders of markdown instructions read at runtime. MCP servers are running HTTP services that expose tools the AI can call. Skills are best for workflows and templates. MCP servers are best for live data and external API access. They compose well — Skills define the workflow, MCP servers fetch the data.
When did Anthropic release Claude Skills?
Anthropic launched Claude Skills on October 16, 2025. In December 2025, the company released the Agent Skills specification as an open standard — similar to how it had previously open-sourced the Model Context Protocol — and partnered with Atlassian, Canva, Figma, Notion, Stripe, and other vendors to ship a prebuilt skills directory.
Where do I install Claude Skills?
Personal skills live in your local Claude Code or Claude Desktop directory (typically ~/.claude/skills/). API users upload skills through their project configuration. Each skill is one folder containing a SKILL.md file and any supporting files.
Can I share a Claude Skill with my team?
Yes. Skills are just folders of files, so you can commit them to a Git repository or share via any file-sharing method. Many marketing and engineering teams maintain Skill repositories that all members install locally for consistent AI behavior across the team.