Claude Skills

Claude Skills: How to Set Them Up, What Files They Use, and Real End to End Examples

What Claude Skills Are
Claude Skills are organized folders that bundle instructions, resources, and sometimes executable scripts, so Claude can handle a specialized task in a repeatable way. The goal is to capture a workflow once and reuse it whenever you ask for that kind of work. 

How Claude Decides to Use a Skill
A Skill has a required name and description in YAML frontmatter at the top of SKILL.md. Claude uses that metadata as an initial signal to decide whether the Skill applies. If it does apply, Claude can then load the rest of the Skill content on demand rather than flooding the context with everything upfront. Anthropic describes this as a progressive disclosure approach. 

Kinds of Skills You Might See in Claude
Claude can use Anthropic provided Skills, like enhanced spreadsheet creation, Word documents, PowerPoint generation, and PDF creation and processing, when code execution and file creation are enabled. You can also upload your own custom Skills as ZIP packages, and on Team or Enterprise plans owners can provision Skills for everyone in the organization. Claude also points to a Skills Directory with partner built Skills that often work with MCP connectors for tools like Notion, Figma, or Atlassian workflows. 

When Skills Are Worth Using
Skills are most useful when the task is specific, repeatable, and quality depends on following the same steps every time. Anthropic recommends that good Skills clearly define when they should be used, stay focused on one workflow, and include examples when helpful. 

Concrete Setup Steps in Claude

Enable Skills in Claude
In the Claude app, Skills are enabled from Settings in Capabilities. You also need code execution and file creation enabled, because Skills require that feature preview capability. 

Upload Your Custom Skill
After your Skill folder is ready, compress it as a ZIP and upload it from the Skills section in Settings in Capabilities using Upload skill. Your uploaded Skill then appears in your Skills list and you can toggle it on or off. 

Important privacy note
The Claude Help Center states Skills run in a secure sandboxed environment with no data persistence between sessions. 

Skill Folder and ZIP Structure That Actually Works

Minimum required file
Every Skill must be a directory that includes a SKILL.md file. That file must start with YAML frontmatter containing nameand description. 

Suggested clean layout
Here is a clear structure you can copy. The ZIP should contain the Skill folder, and inside that folder is SKILL.md.

• customer_support_triage_skill.zip
 • customer_support_triage_skill folder
  • SKILL.md
  • resources folder
   • tone_guide.pdf
   • issue_taxonomy.csv
   • reply_templates.md
  • examples folder
   • example_ticket_01.txt
   • example_output_01.md
  • scripts folder
   • sanitize_text.py
   • extract_fields.py

This matches the idea that a Skill can be simple instructions or a multi file package with resources and executable code. 

Full Example Skill: Customer Support Triage

What this Skill does
It takes messy customer messages and produces consistent triage output: classification, urgency, next action, and a draft reply in your team tone.

SKILL.md content example

---
name: Customer Support Triage
description: Classify incoming support tickets, set urgency, recommend next actions, and draft concise replies in our support tone.
dependencies: python>=3.10
---

Purpose
Help Claude triage customer support tickets quickly and consistently.

When to use
Use this Skill whenever the user provides customer tickets, support emails, chat logs, or asks for ticket classification, prioritization, or reply drafts.

Inputs
The user may provide one ticket or a batch of tickets.
Each ticket may include any of: subject, product, customer name, order id, screenshots, error text, or a free form complaint.

Required output format
For each ticket, output a section with these fields:
Ticket ID
Issue type
Product area
Urgency
Confidence
Missing info
Recommended next step
Draft reply

Urgency rules
High: account access blocked, billing charged incorrectly, security concerns, production outage
Medium: feature broken but workaround exists, partial outage, repeated errors
Low: how to questions, cosmetic issues, feature requests

Style rules for draft replies
Use a calm helpful tone.
Confirm the user goal in one sentence.
Ask only for the minimum missing details.
Offer a simple next step.
Keep it under 120 words unless the user requests longer.
Never promise a timeline unless the ticket includes an SLA.

Use resources
tone_guide.pdf defines our voice and phrases to avoid.
issue_taxonomy.csv defines the allowed Issue type values.
reply_templates.md provides approved short templates.

Safety and privacy
Do not request passwords or full payment card numbers.
If the user includes sensitive data, advise them to redact it and proceed with what remains.

Quality checks
If confidence is below 0.6, say why and ask one clarifying question in Missing info.

This reflects the official requirement that SKILL.md starts with YAML frontmatter containing name and description, and that the description is critical because Claude uses it to determine when to invoke the Skill. 

Resource file examples

resources issue_taxonomy.csv example

issue_type,product_area,notes
Login trouble,Authentication,Includes MFA and password reset
Billing dispute,Billing,Unexpected charge or refund request
Bug report,Core app,Repro steps required if possible
Feature request,Product,Collect use case and urgency

resources reply_templates.md example

Template: Missing details
Hi {name}, thanks for reaching out. I want to make sure we solve the right problem.
Could you share {missing_details}? Once I have that, I can guide you to the next step.

Template: Workaround provided
Hi {name}, thanks for the details. While we investigate, here is a workaround you can try: {workaround}.
If it does not help, please tell me {one_question}.

How you use it in chat

Input example
You paste a ticket like this:

Ticket ID: 18422
Subject: Charged twice
Message: I was billed two times for the same monthly plan. One charge is pending and one posted. Please fix ASAP.
Account email: user@example.com

Output example
Claude should respond in the exact structure your Skill asked for:

Ticket ID: 18422
Issue type: Billing dispute
Product area: Billing
Urgency: High
Confidence: 0.86
Missing info: Order or invoice id for the posted charge, last 4 digits of the card used, and the date of the charges
Recommended next step: Verify duplicate charge in billing system, confirm whether pending will drop, initiate refund if duplicate posted
Draft reply: Hi there, thanks for flagging this. It sounds like you are seeing two charges for the same plan and you want the extra charge removed. Could you share the invoice id for the posted charge and the dates of both charges, plus the last 4 digits of the card used? In the meantime, a pending charge can sometimes drop off automatically, but if we confirm a duplicate posted charge we will fix it right away.

Why this works well as a Skill

It is a repeatable workflow, it includes clear instructions, it defines when to use it, and it includes examples, which matches Anthropic’s own best practice guidance for Skills. 

Second Example Skill: Brand Guidelines for Documents and Slides

What this Skill does
It teaches Claude to apply your brand voice and formatting rules whenever it generates content such as emails, documents, or presentations.

Minimal file set that still works
• brand_guidelines_skill folder
 • SKILL.md
 • resources folder
  • brand_voice.pdf
  • do_and_dont_examples.md
  • approved_phrases.txt

How you would use it
Input prompt example:
Create a customer update email about a delayed shipment using our brand voice and the apology rules in the Skill.

Expected output behavior:
Claude drafts an email that matches the voice rules and avoids forbidden phrases, because the Skill tells it exactly what to check before finalizing.

This fits the official idea that Skills extend Claude with specialized knowledge and workflows and can be used to create documents in a consistent way. 

Third Example Skill: Meeting Notes to Action Items

What this Skill does
It takes raw meeting notes and converts them into a consistent summary, decisions, risks, and action items.

Input example
You paste a messy notes blob.

Output example
Claude outputs:

• Attendees
• Decisions
• Risks and blockers
• Action items with owner and due date
• Open questions

This is exactly the kind of repeatable transformation Skills are designed for, and the Help Center explicitly encourages including examples and focusing on one workflow. 

Tips for Making Your Skill Reliable

Make the description extremely specific
Anthropic notes the description is critical because Claude uses it to determine when to invoke the Skill. Write it like a trigger rule, not marketing text. 

Define strict output formats
If you want consistent outputs, specify exact fields and ordering. Your triage Skill above is a good pattern.

Put real examples in the Skill package
Examples help Claude imitate the exact quality and structure you want. 

Treat Skills as portable packages
Agent Skills have been published as an open standard intended for cross platform portability, and the Agent Skills site describes them as a reusable format across different compatible agent products.