Skills are markdown files that teach Claude new capabilities. No code, no infrastructure, no build steps. Write instructions in plain English, and Claude follows them every time.
A skill is a SKILL.md file with YAML frontmatter and markdown instructions. You write what you want Claude to do, and Claude adds it to its toolkit. When a relevant task comes up, Claude reads the skill and follows the instructions automatically. You can also trigger skills directly by typing /skill-name in any Cowork session.
Skills follow the Agent Skills open standard from agentskills.io. This means your skills aren't locked into Claude. They work across multiple AI tools that support the same specification.
Everything is file-based. Markdown and JSON. No programming languages, no APIs to call, no servers to run. If you can write a checklist, you can write a skill. The hard part isn't the format. It's knowing which workflows to encode and how to write instructions that produce consistent results. That's where working with a specialist makes a difference.
Skills fall into two categories based on how Claude uses them.
Domain knowledge Claude applies to whatever it's currently working on. Conventions, patterns, style guides, terminology. These run inline alongside conversation context.
Examples:
Typically set user-invocable: false so they stay hidden from the / menu. Claude draws on them automatically.
Step-by-step instructions for specific actions. Generating reports, reviewing documents, processing data. Usually triggered by typing /skill-name in your session.
Examples:
Often set disable-model-invocation: true so they only run when you ask for them.
Every skill starts with a YAML frontmatter block (between --- markers) followed by markdown instructions. Here's a complete example:
---
name: weekly-report
description: Generates weekly performance reports. Use when asked for weekly summaries.
argument-hint: [date-range]
disable-model-invocation: false
user-invocable: true
allowed-tools: Read, Grep
context: fork
model: sonnet
---
When generating a weekly report:
1. Pull data from connected sources
2. Organize by department
3. Highlight anomalies
4. Format as PDF using company templatename
/command trigger. Lowercase, hyphens allowed, 64 characters max. This is what appears in the autocomplete menu and what you type to invoke the skill. Example: weekly-report becomes /weekly-report.
description
argument-hint
[date-range] shows as /weekly-report [date-range] in the menu. Whatever the user types after the command gets captured as $ARGUMENTS.
disable-model-invocation
true, the skill only runs if the user types /skill-name. Claude won't trigger it on its own. Set this for skills that make significant changes, send emails, or modify files. When false (the default), Claude can auto-invoke when it thinks the skill is relevant.
user-invocable
false, the skill is hidden from the / menu entirely. It becomes background knowledge that Claude draws on silently. Good for reference skills like style guides, conventions, or domain terminology.
allowed-tools
allowed-tools: Read, Grep lets the skill search and read files without pausing for approval.
context
inline runs it in the main conversation, which is great for quick reference lookups. fork runs it in an isolated subagent with its own context window. Use fork for long tasks that produce large outputs, so they don't clutter your main thread.
model
A skill can be a single SKILL.md file. Or it can be a folder with supporting files that Claude references when running the skill:
my-skill/
├── SKILL.md # Main instructions (required)
├── template.md # Template for Claude to fill in
├── examples/
│ └── sample.md # Example output showing expected format
└── scripts/
└── validate.sh # Script Claude can execute
Only SKILL.md is required. The supporting files are optional but powerful. Templates enforce consistent output formatting. Examples show Claude what "good" looks like. Scripts let Claude validate its own work.
Skills can exist in four locations. Higher-priority locations win when skill names conflict.
Highest priority. Set by organization admins. Apply to all users in the org. Read-only for individual team members. This is how companies enforce standard workflows across their team.
~/.claude/skills/Your personal skills, available across all projects. Put your daily workflow skills here: your writing style guide, your preferred report format, your weekly review process.
.claude/skills/Shared with your team via git. Project-specific skills that everyone on the project uses. Good for team conventions, project templates, and shared workflows.
<plugin>/skills/Skills bundled inside plugins. Available wherever that plugin is enabled. Namespaced as plugin:skill-name to avoid conflicts. The 11 official Anthropic plugins each ship with their own skills.
Five skills come pre-installed with Claude Code and Cowork. You don't need to create or configure them.
/batch <instruction>Runs large-scale changes in parallel. Claude spawns one sub-agent per unit of work. If you need to process 50 files the same way, /batch handles them concurrently instead of one at a time.
/claude-apiLoads the Claude API reference for your programming language. Useful when you're building integrations or custom tools that call the Claude API.
/debug [description]Troubleshoots session issues by reading debug logs. When something isn't working right, /debug digs into the logs and tells you what went wrong.
/loop [interval] <prompt>Runs a prompt repeatedly on a schedule. Pair this with /schedule for recurring tasks like daily file organization, weekly report generation, or hourly data checks.
/simplify [focus]Reviews recent changes for quality. Spawns three parallel review agents that check your work from different angles. Think of it as an automated peer review.
The real power of claude cowork skills is building your own. Here are three examples from different industries.
A real estate agent runs this every Monday morning. It pulls new listings, calculates price changes by neighborhood, checks mortgage rate trends, and outputs a client-ready newsletter. What used to take 2 hours now takes 3 minutes.
---
name: weekly-market-report
description: Generates a weekly market analysis report for client distribution
disable-model-invocation: true
---
Generate a weekly market report:
1. Pull new listings from MLS connector (last 7 days)
2. Calculate median price changes by neighborhood
3. Identify notable sales (above/below ask)
4. Check mortgage rate trends
5. Format using the template in ./templates/market-report.md
6. Save to ~/Reports/weekly/ with date in filename
7. Draft email summary for client newsletterZack Shapiro's "Claude-Native Law Firm" approach: encode your analytical frameworks, preferred formats, and professional judgment into skills. Claude opens .docx files at the XML level and writes tracked changes attributed to the reviewing lawyer's name.
---
name: contract-review
description: Reviews contracts for standard issues and red flags
---
When reviewing a contract:
1. Identify contract type (NDA, MSA, employment, lease, etc.)
2. Check for missing standard clauses per ./reference/required-clauses.md
3. Flag unusual liability provisions
4. Note indemnification scope
5. Check termination conditions
6. Compare fee structure against market rates in ./reference/market-rates.md
7. Generate a summary memo with: risk level (low/medium/high),
key terms, recommended changesThis pattern, open-sourced at github.com/rebelytics/one-skill-to-rule-them-all, watches other skills perform and logs corrections you make. Over time, it updates skill files with those learnings. Your skills get better the more you use them. A marketing team used this approach to build banner upload, dev ticket generation, and image alt text skills that improved with each batch.
The meta-skill observes each session, records when you correct Claude's output, identifies patterns in those corrections, and applies them to the source skill files. It's how you go from "good enough" to "exactly right" without rewriting skills from scratch.
Skills aren't static checklists. Four features make them responsive to context.
$ARGUMENTS
Captures whatever the user types after the command. When you type /weekly-report Q1 2026, the skill receives $ARGUMENTS = "Q1 2026". Your instructions can reference this variable to customize the output based on user input.
Reference templates, examples, and scripts from within your SKILL.md. Claude reads them when needed. A market report skill can reference ./templates/market-report.md for formatting, and Claude fills in the template with fresh data each time.
Start with a brief SKILL.md and reference detailed docs in separate files. Claude loads them only when needed. This keeps token usage efficient. A complex compliance skill might have 10 pages of reference material, but Claude only reads the relevant section for each task.
Set context: fork to run a skill in an isolated subagent with its own context window. Long tasks won't clutter your main conversation. The subagent completes the work and returns results. Good for report generation, batch processing, and multi-step analysis.
NYC Claw identifies your repetitive workflows and turns them into custom skills. One-time setup, ongoing value. No consulting retainer required.
We map what you do weekly and monthly. Which tasks take hours? Which follow the same pattern every time? Which require copying data between tools? Those are your skill candidates.
Each workflow becomes a skill with step-by-step instructions, supporting templates, and example outputs. We configure the frontmatter so skills trigger at the right time with the right permissions.
Skills work best with connectors wired up. We configure your CRM, project tracker, email, and document storage via MCP so skills can pull and push data automatically.
We iterate on prompts until output quality matches your expectations. Then we train your team on using the skills, customizing them, and building new ones. See the setup guide for more on the process.
Want to understand pricing before the call? Check the Cowork pricing guide for plan details and what's included at each tier.
Every business has 5-10 tasks that follow the same pattern every week. Let's turn yours into skills that run in minutes instead of hours.
Book a Free 15-Minute Call15 minutes · No commitment · We'll map your top skill candidates
.claude/skills/ directory). Think of plugins as the container and skills as one type of content inside it.
.mcp.json file. When a skill runs, it can use any connector that's been set up: HubSpot, Slack, Google Drive, Salesforce, and 30+ others.